We consider language modelling (LM) as a multi-label structured prediction task by re-framing training from solely predicting a single ground-truth word to ranking a set of words which could continue a given context. To avoid annotating top-$k$ ranks, we generate them using pre-trained LMs: GPT-2, BERT, and Born-Again models. This leads to a rank-based form of knowledge distillation (KD). We also develop a method using $N$-grams to create a non-probabilistic teacher which generates the ranks without the need of a pre-trained LM.

We confirm the hypotheses that we can treat LMing as a ranking task and that we can do so without the use of a pre-trained LM. We show that rank-based KD generally improves perplexity (PPL), often with statistical significance, when compared to Kullback-Leibler-based KD. Surprisingly, given the simplicity of the method, $N$-grams act as competitive teachers and achieve similar performance as using either BERT or a Born-Again model teachers. GPT-2 always acts as the best teacher, though, and using it and a Transformer-XL student on Wiki-02, rank-based KD reduces a cross-entropy baseline from 65.27 to 55.94 and against a KL-based KD of 56.70.

Since late 2019, COVID-19 has quickly emerged as the newest biomedical domain, resulting in a surge of new information. As with other emergent domains, the discussion surrounding the topic has been rapidly changing, leading to the spread of misinformation. This has created the need for a public space for users to ask questions and receive credible, scientific answers. To fulfill this need, we turn to the task of open-domain question-answering, which we can use to efficiently find answers to free-text questions from a large set of documents. In this work, we present such a system for the emergent domain of COVID-19. Despite the small data size available, we are able to successfully train the system to retrieve answers from a large-scale corpus of published COVID-19 scientific papers. Furthermore, we incorporate effective re-ranking and question-answering techniques, such as document diversity and multiple answer spans. Our open-domain question-answering system can further act as a model for the quick development of similar systems that can be adapted and modified for other developing emergent domains.

Terms are linguistic signifiers of domain-specific concepts. Automated recognition of multi-word terms (MWT) in free text is a sequence labelling problem, which is commonly addressed using supervised machine learning methods. Their need for manual annotation of training data makes it difficult to port such methods across domains. FlexiTerm, on the other hand, is a fully unsupervised method for MWT recognition from domain-specific corpora. Originally implemented in Java as a proof of concept, it did not scale well, thus offering little practical value in the context of big data. In this paper, we describe its re-implementation in Python and compare the performance of these two implementations. The results demonstrated major improvements in terms of efficiency, which allow FlexiTerm to transition from the proof of concept to the production-grade application.

Training data for machine translation (MT) is often sourced from a multitude of large corpora that are multi-faceted in nature, e.g. containing contents from multiple domains or different levels of quality or complexity. Naturally, these facets do not occur with equal frequency, nor are they equally important for the test scenario at hand. In this work, we propose to optimize this balance jointly with MT model parameters to relieve system developers from manual schedule design. A multi-armed bandit is trained to dynamically choose between facets in a way that is most beneficial for the MT system. We evaluate it on three different multi-facet applications: balancing translationese and natural training data, or data from multiple domains or multiple language pairs. We find that bandit learning leads to competitive MT systems across tasks, and our analysis provides insights into its learned strategies and the underlying data sets.

Text autoencoders are often used for unsupervised conditional text generation by applying mappings in the latent space to change attributes to the desired values. Recently, Mai et al. (2020) proposed Emb2Emb, a method to learn these mappings in the embedding space of an autoencoder. However, their method is restricted to autoencoders with a single-vector embedding, which limits how much information can be retained. We address this issue by extending their method to Bag-of-Vectors Autoencoders (BoV-AEs), which encode the text into a variable-size bag of vectors that grows with the size of the text, as in attention-based models. This allows to encode and reconstruct much longer texts than standard autoencoders. Analogous to conventional autoencoders, we propose regularization techniques that facilitate learning meaningful operations in the latent space. Finally, we adapt for a training scheme that learns to map an input bag to an output bag, including a novel loss function and neural architecture. Our experimental evaluations on unsupervised sentiment transfer and sentence summarization show that our method performs substantially better than a standard autoencoder.

This work concentrates on reducing the RTF and word error rate of a hybrid HMM-DNN. Our baseline system uses an architecture with TDNN and LSTM layers. We find this architecture particularly useful for lightly reverberated environments. However, these models tend to demand more computation than is desirable. In this work, we explore alternate architectures employing singular value decomposition (SVD) is applied to the TDNN layers to reduce the RTF, as well as to the affine transforms of every LSTM cell. We compare this approach with specifying bottleneck layers similar to those introduced by SVD before training. Additionally, we reduced the search space of the decoding graph to make it a better fit to operate in real-time applications. We report -61.57% relative reduction in RTF and almost 1% relative decrease in WER for our architecture trained on Fisher data along with reverberated versions of this dataset in order to match one of our target test distributions.

An interactive instruction following task has been proposed as a benchmark for learning to map natural language instructions and first-person vision into sequences of actions to interact with objects in a 3D simulated environment. We find that an existing end-to-end neural model for this task is not robust to variations of objects and language instructions. We assume that this problem is due to the high sensitiveness of neural feature extraction to small changes in vision and language inputs. To mitigate this problem, we propose a neuro-symbolic approach that performs reasoning over high-level symbolic representations that are robust to small changes in raw inputs. Our experiments on the ALFRED dataset show that our approach significantly outperforms the existing model by 18, 52, and 73 points in the success rate on the ToggleObject, PickupObject, and SliceObject subtasks in unseen environments respectively.

Continual learning (CL), or domain expansion, recently became a popular topic for automatic speech recognition (ASR) acoustic modeling because practical systems have to be updated frequently in order to work robustly on types of speech not observed during initial training. While sequential adaptation allows tuning a system to a new domain, it may result in performance degradation on the old domains due to catastrophic forgetting. In this work we explore regularization-based CL for neural network acoustic models trained with the lattice-free maximum mutual information (LF-MMI) criterion. We simulate domain expansion by incrementally adapting the acoustic model on different public datasets that include several accents and speaking styles. We investigate two well-known CL techniques, elastic weight consolidation (EWC) and learning without forgetting (LWF), which aim to reduce forgetting by preserving model weights or network outputs. We additionally introduce a sequence-level LWF regularization, which exploits posteriors from the denominator graph of LF-MMI to further reduce forgetting. Empirical results show that the proposed sequence-level LWF can improve the best average word error rate across all domains by up to 9.4% relative compared with using regular LWF.

Understanding of causal narratives communicated in clinical notes can help make strides towards personalized healthcare. In this work, MIMICause, we propose annotation guidelines, develop an annotated corpus and provide baseline scores to identify types and direction of causal relations between a pair of biomedical concepts in clinical notes; communicated implicitly or explicitly, identified either in a single sentence or across multiple sentences.

We annotate a total of 2714 de-identified examples sampled from the 2018 n2c2 shared task dataset and train four different language model based architectures. Annotation based on our guidelines achieved a high inter-annotator agreement i.e. Fleiss' kappa score of 0.72 and our model for identification of causal relation achieved a macro F1 score of 0.56 on test data. The high inter-annotator agreement for clinical text shows the quality of our annotation guidelines while the provided baseline F1 score sets the direction for future research towards understanding narratives in clinical texts.

As the volume of long-form spoken-word content such as podcasts explodes, many platforms desire to present short, meaningful, and logically coherent segments extracted from the full content. Such segments can be consumed by users to sample content before diving in, as well as used by the platform to promote and recommend content. However, little published work is focused on the segmentation of spoken-word content, where the errors (noise) in transcripts generated by automatic speech recognition (ASR) services poses many challenges. Here we build a novel dataset of complete transcriptions of over 400 podcast episodes, in which we label the position of introductions in each episode. These introductions contain information about the episodes' topics, hosts, and guests, providing a valuable summary of the episode content, as it is created by the authors. We further augment our dataset with word substitutions to increase the amount of available training data. We train three Transformer models based on the pre-trained BERT and different augmentation strategies, which achieve significantly better performance compared with a static embedding model, showing that it is possible to capture generalized, larger-scale structural information from noisy, loosely-organized speech data. This is further demonstrated through an analysis of the models' inner architecture. Our methods and dataset can be used to facilitate future work on the structure-based segmentation of spoken-word content.

This technical report summarizes our method for the Video-And-Language Understanding Evaluation (VALUE) challenge (https://value-benchmark.github.io/challenge\_2021.html). We propose a CLIP-Enhanced method to incorporate the image-text pretrained knowledge into downstream video-text tasks. Combined with several other improved designs, our method outperforms the state-of-the-art by $2.4\%$ ($57.58$ to $60.00$) Meta-Ave score on VALUE benchmark.

Adversarial attacks and backdoor attacks are two common security threats that hang over deep learning. Both of them harness task-irrelevant features of data in their implementation. Text style is a feature that is naturally irrelevant to most NLP tasks, and thus suitable for adversarial and backdoor attacks. In this paper, we make the first attempt to conduct adversarial and backdoor attacks based on text style transfer, which is aimed at altering the style of a sentence while preserving its meaning. We design an adversarial attack method and a backdoor attack method, and conduct extensive experiments to evaluate them. Experimental results show that popular NLP models are vulnerable to both adversarial and backdoor attacks based on text style transfer -- the attack success rates can exceed 90% without much effort. It reflects the limited ability of NLP models to handle the feature of text style that has not been widely realized. In addition, the style transfer-based adversarial and backdoor attack methods show superiority to baselines in many aspects. All the code and data of this paper can be obtained at https://github.com/thunlp/StyleAttack.

In recent years, researchers tend to pre-train ever-larger language models to explore the upper limit of deep models. However, large language model pre-training costs intensive computational resources and most of the models are trained from scratch without reusing the existing pre-trained models, which is wasteful. In this paper, we propose bert2BERT, which can effectively transfer the knowledge of an existing smaller pre-trained model (e.g., BERT_BASE) to a large model (e.g., BERT_LARGE) through parameter initialization and significantly improve the pre-training efficiency of the large model. Specifically, we extend the previous function-preserving on Transformer-based language model, and further improve it by proposing advanced knowledge for large model's initialization. In addition, a two-stage pre-training method is proposed to further accelerate the training process. We did extensive experiments on representative PLMs (e.g., BERT and GPT) and demonstrate that (1) our method can save a significant amount of training cost compared with baselines including learning from scratch, StackBERT and MSLT; (2) our method is generic and applicable to different types of pre-trained models. In particular, bert2BERT saves about 45% and 47% computational cost of pre-training BERT_BASE and GPT_BASE by reusing the models of almost their half sizes. The source code will be publicly available upon publication.

Open-Retrieval Generative Question Answering (GenQA) is proven to deliver high-quality, natural-sounding answers in English. In this paper, we present the first generalization of the GenQA approach for the multilingual environment. To this end, we present the GenTyDiQA dataset, which extends the TyDiQA evaluation data (Clark et al., 2020) with natural-sounding, well-formed answers in Arabic, Bengali, English, Japanese, and Russian. For all these languages, we show that a GenQA sequence-to-sequence-based model outperforms a state-of-the-art Answer Sentence Selection model. We also show that a multilingually-trained model competes with, and in some cases outperforms, its monolingual counterparts. Finally, we show that our system can even compete with strong baselines, even when fed with information from a variety of languages. Essentially, our system is able to answer a question in any language of our language set using information from many languages, making it the first Language-Agnostic GenQA system.

Recent work finds modern natural language processing (NLP) models relying on spurious features for prediction. Mitigating such effects is thus important. Despite this need, there is no quantitative measure to evaluate or compare the effects of different forms of spurious features in NLP. We address this gap in the literature by quantifying model sensitivity to spurious features with a causal estimand, dubbed CENT, which draws on the concept of average treatment effect from the causality literature. By conducting simulations with four prominent NLP models -- TextRNN, BERT, RoBERTa and XLNet -- we rank the models against their sensitivity to artificial injections of eight spurious features. We further hypothesize and validate that models that are more sensitive to a spurious feature will be less robust against perturbations with this feature during inference. Conversely, data augmentation with this feature improves robustness to similar perturbations. We find statistically significant inverse correlations between sensitivity and robustness, providing empirical support for our hypothesis.

This paper proposes a transformer over transformer framework, called Transformer$^2$, to perform neural text segmentation. It consists of two components: bottom-level sentence encoders using pre-trained transformers, and an upper-level transformer-based segmentation model based on the sentence embeddings. The bottom-level component transfers the pre-trained knowledge learned from large external corpora under both single and pair-wise supervised NLP tasks to model the sentence embeddings for the documents. Given the sentence embeddings, the upper-level transformer is trained to recover the segmentation boundaries as well as the topic labels of each sentence. Equipped with a multi-task loss and the pre-trained knowledge, Transformer$^2$ can better capture the semantic coherence within the same segments. Our experiments show that (1) Transformer$^2$ manages to surpass state-of-the-art text segmentation models in terms of a commonly-used semantic coherence measure; (2) in most cases, both single and pair-wise pre-trained knowledge contribute to the model performance; (3) bottom-level sentence encoders pre-trained on specific languages yield better performance than those pre-trained on specific domains.

Neural topic models (NTMs) apply deep neural networks to topic modelling. Despite their success, NTMs generally ignore two important aspects: (1) only document-level word count information is utilized for the training, while more fine-grained sentence-level information is ignored, and (2) external semantic knowledge regarding documents, sentences and words are not exploited for the training. To address these issues, we propose a variational autoencoder (VAE) NTM model that jointly reconstructs the sentence and document word counts using combinations of bag-of-words (BoW) topical embeddings and pre-trained semantic embeddings. The pre-trained embeddings are first transformed into a common latent topical space to align their semantics with the BoW embeddings. Our model also features hierarchical KL divergence to leverage embeddings of each document to regularize those of their sentences, thereby paying more attention to semantically relevant sentences. Both quantitative and qualitative experiments have shown the efficacy of our model in 1) lowering the reconstruction errors at both the sentence and document levels, and 2) discovering more coherent topics from real-world datasets.

Analysis of vision-and-language models has revealed their brittleness under linguistic phenomena such as paraphrasing, negation, textual entailment, and word substitutions with synonyms or antonyms. While data augmentation techniques have been designed to mitigate against these failure modes, methods that can integrate this knowledge into the training pipeline remain under-explored. In this paper, we present \textbf{SDRO}, a model-agnostic method that utilizes a set linguistic transformations in a distributed robust optimization setting, along with an ensembling technique to leverage these transformations during inference. Experiments on benchmark datasets with images (NLVR$^2$) and video (VIOLIN) demonstrate performance improvements as well as robustness to adversarial attacks. Experiments on binary VQA explore the generalizability of this method to other V\&L tasks.

Neural abstractive summarization models are susceptible to generating factually inconsistent content, a phenomenon known as hallucination. This limits the usability and adoption of these systems in real-world applications. To reduce the presence of hallucination, we propose the Mixture of Factual Experts (MoFE) model, which combines multiple summarization experts that each target a specific type of error. We train our experts using reinforcement learning (RL) to minimize the error defined by two factual consistency metrics: entity overlap and dependency arc entailment. We construct MoFE by combining the experts using two ensembling strategies (weights and logits) and evaluate them on two summarization datasets (XSUM and CNN/DM). Our experiments on BART models show that the MoFE improves performance according to both entity overlap and dependency arc entailment, without a significant performance drop on standard ROUGE metrics. The performance improvement also transfers to unseen factual consistency metrics, such as question answer-based factuality evaluation metric and BERTScore precision with respect to the source document.

The goal of Word Sense Disambiguation (WSD) is to identify the sense of a polysemous word in a specific context. Deep-learning techniques using BERT have achieved very promising results in the field and different methods have been proposed to integrate structured knowledge to enhance performance. At the same time, an increasing number of data augmentation techniques have been proven to be useful for NLP tasks. Building upon previous works leveraging BERT and WordNet knowledge, we explore different data augmentation techniques on context-gloss pairs to improve the performance of WSD. In our experiment, we show that both sentence-level and word-level augmentation methods are effective strategies for WSD. Also, we find out that performance can be improved by adding hypernyms' glosses obtained from a lexical knowledge base. We compare and analyze different context-gloss augmentation techniques, and the results show that applying back translation on gloss performs the best.

The common practice for training commonsense models has gone from-human-to-corpus-to-machine: humans author commonsense knowledge graphs in order to train commonsense models. In this work, we investigate an alternative, from-machine-to-corpus-to-machine: general language models author these commonsense knowledge graphs to train commonsense models. Our study leads to a new framework, Symbolic Knowledge Distillation. As with prior art in Knowledge Distillation (Hinton et al., 2015), our approach uses larger models to teach smaller models. A key difference is that we distill knowledge symbolically-as text-in addition to the neural model. We also distill only one aspect-the commonsense of a general language model teacher, allowing the student to be a different type, a commonsense model. Altogether, we show that careful prompt engineering and a separately trained critic model allow us to selectively distill high-quality causal commonsense from GPT-3, a general language model. Empirical results demonstrate that, for the first time, a human-authored commonsense knowledge graph is surpassed by our automatically distilled variant in all three criteria: quantity, quality, and diversity. In addition, it results in a neural commonsense model that surpasses the teacher model's commonsense capabilities despite its 100x smaller size. We apply this to the ATOMIC resource, and share our new symbolic knowledge graph and commonsense models.

International Phonetic Alphabet (IPA) has been widely used in cross-lingual text-to-speech (TTS) to achieve cross-lingual voice cloning (CL VC). However, IPA itself has been understudied in cross-lingual TTS. In this paper, we report some empirical findings of building a cross-lingual TTS model using IPA as inputs. Experiments show that the way to process the IPA and suprasegmental sequence has a negligible impact on the CL VC performance. Furthermore, we find that using a dataset including one speaker per language to build an IPA-based TTS system would fail CL VC since the language-unique IPA and tone/stress symbols could leak the speaker information. In addition, we experiment with different combinations of speakers in the training dataset to further investigate the effect of the number of speakers on the CL VC performance.

Although large-scale pre-trained neural models have shown impressive performances in a variety of tasks, their ability to generate coherent text that appropriately models discourse phenomena is harder to evaluate and less understood. Given the claims of improved text generation quality across various systems, we consider the coherence evaluation of machine generated text to be one of the principal applications of coherence models that needs to be investigated. We explore training data and self-supervision objectives that result in a model that generalizes well across tasks and can be used off-the-shelf to perform such evaluations. Prior work in neural coherence modeling has primarily focused on devising new architectures, and trained the model to distinguish coherent and incoherent text through pairwise self-supervision on the permuted documents task. We instead use a basic model architecture and show significant improvements over state of the art within the same training regime. We then design a harder self-supervision objective by increasing the ratio of negative samples within a contrastive learning setup, and enhance the model further through automatic hard negative mining coupled with a large global negative queue encoded by a momentum encoder. We show empirically that increasing the density of negative samples improves the basic model, and using a global negative queue further improves and stabilizes the model while training with hard negative samples. We evaluate the coherence model on task-independent test sets that resemble real-world use cases and show significant improvements in coherence evaluations of downstream applications.

Motivated by the success of T5 (Text-To-Text Transfer Transformer) in pre-training natural language processing models, we propose a unified-modal SpeechT5 framework that explores the encoder-decoder pre-training for self-supervised speech/text representation learning. The SpeechT5 framework consists of a shared encoder-decoder network and six modal-specific (speech/text) pre/post-nets. After preprocessing the speech/text input through the pre-nets, the shared encoder-decoder network models the sequence to sequence transformation, and then the post-nets generate the output in the speech/text modality based on the decoder output. Particularly, SpeechT5 can pre-train on a large scale of unlabeled speech and text data to improve the capability of the speech and textual modeling. To align the textual and speech information into a unified semantic space, we propose a cross-modal vector quantization method with random mixing-up to bridge speech and text. Extensive evaluations on a wide variety of spoken language processing tasks, including voice conversion, automatic speech recognition, text to speech, and speaker identification, show the superiority of the proposed SpeechT5 framework.

Pun location is to identify the punning word (usually a word or a phrase that makes the text ambiguous) in a given short text, and pun interpretation is to find out two different meanings of the punning word. Most previous studies adopt limited word senses obtained by WSD(Word Sense Disambiguation) technique or pronunciation information in isolation to address pun location. For the task of pun interpretation, related work pays attention to various WSD algorithms. In this paper, a model called DANN (Dual-Attentive Neural Network) is proposed for pun location, effectively integrates word senses and pronunciation with context information to address two kinds of pun at the same time. Furthermore, we treat pun interpretation as a classification task and construct pungloss pairs as processing data to solve this task. Experiments on the two benchmark datasets show that our proposed methods achieve new state-of-the-art results. Our source code is available in the public code repository.

Recently, sequence-to-sequence (seq-to-seq) models have been successfully applied in text-to-speech (TTS) to synthesize speech for single-language text. To synthesize speech for multiple languages usually requires multi-lingual speech from the target speaker. However, it is both laborious and expensive to collect high-quality multi-lingual TTS data for the target speakers. In this paper, we proposed to use low-quality code-switched found data from the non-target speakers to achieve cross-lingual voice cloning for the target speakers. Experiments show that our proposed method can generate high-quality code-switched speech in the target voices in terms of both naturalness and speaker consistency. More importantly, we find that our method can achieve a comparable result to the state-of-the-art (SOTA) performance in cross-lingual voice cloning.

Recursive processing is considered a hallmark of human linguistic abilities. A recent study evaluated recursive processing in recurrent neural language models (RNN-LMs) and showed that such models perform below chance level on embedded dependencies within nested constructions -- a prototypical example of recursion in natural language. Here, we study if state-of-the-art Transformer LMs do any better. We test four different Transformer LMs on two different types of nested constructions, which differ in whether the embedded (inner) dependency is short or long range. We find that Transformers achieve near-perfect performance on short-range embedded dependencies, significantly better than previous results reported for RNN-LMs and humans. However, on long-range embedded dependencies, Transformers' performance sharply drops below chance level. Remarkably, the addition of only three words to the embedded dependency caused Transformers to fall from near-perfect to below-chance performance. Taken together, our results reveal Transformers' shortcoming when it comes to recursive, structure-based, processing.

Pre-trained language models (PLMs), such as BERT and GPT, have revolutionized the field of NLP, not only in the general domain but also in the biomedical domain. Most prior efforts in building biomedical PLMs have resorted simply to domain adaptation and focused mainly on English. In this work we introduce eHealth, a biomedical PLM in Chinese built with a new pre-training framework. This new framework trains eHealth as a discriminator through both token-level and sequence-level discrimination. The former is to detect input tokens corrupted by a generator and select their original signals from plausible candidates, while the latter is to further distinguish corruptions of a same original sequence from those of the others. As such, eHealth can learn language semantics at both the token and sequence levels. Extensive experiments on 11 Chinese biomedical language understanding tasks of various forms verify the effectiveness and superiority of our approach. The pre-trained model is available to the public at \url{https://github.com/PaddlePaddle/Research/tree/master/KG/eHealth} and the code will also be released later.

Many mispronunciation detection and diagnosis (MD&D) research approaches try to exploit both the acoustic and linguistic features as input. Yet the improvement of the performance is limited, partially due to the shortage of large amount annotated training data at the phoneme level. Phonetic embeddings, extracted from ASR models trained with huge amount of word level annotations, can serve as a good representation of the content of input speech, in a noise-robust and speaker-independent manner. These embeddings, when used as implicit phonetic supplementary information, can alleviate the data shortage of explicit phoneme annotations. We propose to utilize Acoustic, Phonetic and Linguistic (APL) embedding features jointly for building a more powerful MD\&D system. Experimental results obtained on the L2-ARCTIC database show the proposed approach outperforms the baseline by 9.93%, 10.13% and 6.17% on the detection accuracy, diagnosis error rate and the F-measure, respectively.

Recent work (e.g. LAMA (Petroni et al., 2019)) has found that the quality of the factual information extracted from Large Language Models (LLMs) depends on the prompts used to query them. This inconsistency is problematic because different users will query LLMs for the same information using different wording, but should receive the same, accurate responses regardless. In this work we aim to address this shortcoming by introducing P-Adapters: lightweight models that sit between the embedding layer and first attention layer of LLMs. They take LLM embeddings as input and output continuous prompts that are used to query the LLM. Additionally, we investigate Mixture of Experts (MoE) models that learn a set of continuous prompts ("experts") and select one to query the LLM. They require a separate classifier trained on human-annotated data to map natural language prompts to the continuous ones. P-Adapters perform comparably to the more complex MoE models in extracting factual information from BERT and RoBERTa while eliminating the need for additional annotations. P-Adapters show between 12-26% absolute improvement in precision and 36-50% absolute improvement in consistency over a baseline of only using natural language queries. Finally, we investigate what makes a P-adapter successful and conclude that access to the LLM's embeddings of the original natural language prompt, particularly the subject of the entity pair being asked about, is a significant factor.

Existing approaches to lifelong language learning rely on plenty of labeled data for learning a new task, which is hard to obtain in most real scenarios. Considering that humans can continually learn new tasks from a handful of examples, we expect the models also to be able to generalize well on new few-shot tasks without forgetting the previous ones. In this work, we define this more challenging yet practical problem as Lifelong Few-shot Language Learning (LFLL) and propose a unified framework for it based on prompt tuning of T5. Our framework called LFPT5 takes full advantage of PT's strong few-shot learning ability, and simultaneously trains the model as a task solver and a data generator. Before learning a new domain of the same task type, LFPT5 generates pseudo (labeled) samples of previously learned domains, and later gets trained on those samples to alleviate forgetting of previous knowledge as it learns the new domain. In addition, a KL divergence loss is minimized to achieve label consistency between the previous and the current model. While adapting to a new task type, LFPT5 includes and tunes additional prompt embeddings for the new task. With extensive experiments, we demonstrate that LFPT5 can be applied to various different types of tasks and significantly outperform previous methods in different LFLL settings.

Aspect Sentiment Triplet Extraction (ASTE) aims to extract aspect term (aspect), sentiment and opinion term (opinion) triplets from sentences and can tell a complete story, i.e., the discussed aspect, the sentiment toward the aspect, and the cause of the sentiment. ASTE is a charming task, however, one triplet extracted by ASTE only includes one opinion of the aspect, but an aspect in a sentence may have multiple corresponding opinions and one opinion only provides part of the reason why the aspect has this sentiment, as a consequence, some triplets extracted by ASTE are hard to understand, and provide erroneous information for downstream tasks. In this paper, we introduce a new task, named Aspect Sentiment Multiple Opinions Triplet Extraction (ASMOTE). ASMOTE aims to extract aspect, sentiment and multiple opinions triplets. Specifically, one triplet extracted by ASMOTE contains all opinions about the aspect and can tell the exact reason that the aspect has the sentiment. We propose an Aspect-Guided Framework (AGF) to address this task. AGF first extracts aspects, then predicts their opinions and sentiments. Moreover, with the help of the proposed Sequence Labeling Attention(SLA), AGF improves the performance of the sentiment classification using the extracted opinions. Experimental results on multiple datasets demonstrate the effectiveness of our approach.

In this paper, we present an extensive investigation of multi-bridge, many-to-many multilingual NMT models (MB-M2M) ie., models trained on non-English language pairs in addition to English-centric language pairs. In addition to validating previous work which shows that MB-M2M models can overcome zeroshot translation problems, our analysis reveals the following results about multibridge models: (1) it is possible to extract a reasonable amount of parallel corpora between non-English languages for low-resource languages (2) with limited non-English centric data, MB-M2M models are competitive with or outperform pivot models, (3) MB-M2M models can outperform English-Any models and perform at par with Any-English models, so a single multilingual NMT system can serve all translation directions.

Aspect category sentiment analysis has attracted increasing research attention. The dominant methods make use of pre-trained language models by learning effective aspect category-specific representations, and adding specific output layers to its pre-trained representation. We consider a more direct way of making use of pre-trained language models, by casting the ACSA tasks into natural language generation tasks, using natural language sentences to represent the output. Our method allows more direct use of pre-trained knowledge in seq2seq language models by directly following the task setting during pre-training. Experiments on several benchmarks show that our method gives the best reported results, having large advantages in few-shot and zero-shot settings.

Despite the increasing popularity of NLP in the humanities and social sciences, advances in model performance and complexity have been accompanied by concerns about interpretability and explanatory power for sociocultural analysis. One popular model that balances complexity and legibility is Word Mover's Distance (WMD). Ostensibly adapted for its interpretability, WMD has nonetheless been used and further developed in ways which frequently discard its most interpretable aspect: namely, the word-level distances required for translating a set of words into another set of words. To address this apparent gap, we introduce WMDecompose: a model and Python library that 1) decomposes document-level distances into their constituent word-level distances, and 2) subsequently clusters words to induce thematic elements, such that useful lexical information is retained and summarized for analysis. To illustrate its potential in a social scientific context, we apply it to a longitudinal social media corpus to explore the interrelationship between conspiracy theories and conservative American discourses. Finally, because of the full WMD model's high time-complexity, we additionally suggest a method of sampling document pairs from large datasets in a reproducible way, with tight bounds that prevent extrapolation of unreliable results due to poor sampling practices.

Plug-and-play functionality allows deep learning models to adapt well to different tasks without requiring any parameters modified. Recently, prefix-tuning was shown to be a plug-and-play method on various text generation tasks by simply inserting corresponding continuous vectors into the inputs. However, sequence labeling tasks invalidate existing plug-and-play methods since different label sets demand changes to the architecture of the model classifier. In this work, we propose the use of label word prediction instead of classification to totally reuse the architecture of pre-trained models for sequence labeling tasks. Specifically, for each task, a label word set is first constructed by selecting a high-frequency word for each class respectively, and then, task-specific vectors are inserted into the inputs and optimized to manipulate the model predictions towards the corresponding label words. As a result, by simply switching the plugin vectors on the input, a frozen pre-trained language model is allowed to perform different tasks. Experimental results on three sequence labeling tasks show that the performance of the proposed method can achieve comparable performance with standard fine-tuning with only 0.1\% task-specific parameters. In addition, our method is up to 70 times faster than non-plug-and-play methods while switching different tasks under the resource-constrained scenario.

Many legal professionals think that the explosion of information about local, regional, national, and international legislation makes their practice more costly, time-consuming, and even error-prone. The two main reasons for this are that most legislation is usually unstructured, and the tremendous amount and pace with which laws are released causes information overload in their daily tasks. In the case of the legal domain, the research community agrees that a system allowing to generate automatic responses to legal questions could substantially impact many practical implications in daily activities. The degree of usefulness is such that even a semi-automatic solution could significantly help to reduce the workload to be faced. This is mainly because a Question Answering system could be able to automatically process a massive amount of legal resources to answer a question or doubt in seconds, which means that it could save resources in the form of effort, money, and time to many professionals in the legal sector. In this work, we quantitatively and qualitatively survey the solutions that currently exist to meet this challenge.

Recent methods for embodied instruction following are typically trained end-to-end using imitation learning. This requires the use of expert trajectories and low-level language instructions. Such approaches assume learned hidden states will simultaneously integrate semantics from the language and vision to perform state tracking, spatial memory, exploration, and long-term planning. In contrast, we propose a modular method with structured representations that (1) builds a semantic map of the scene, and (2) performs exploration with a semantic search policy, to achieve the natural language goal. Our modular method achieves SOTA performance (24.46%) with a substantial (8.17 % absolute) gap from previous work while using less data by eschewing both expert trajectories and low-level instructions. Leveraging low-level language, however, can further increase our performance (26.49%). Our findings suggest that an explicit spatial memory and a semantic search policy can provide a stronger and more general representation for state-tracking and guidance, even in the absence of expert trajectories or low-level instructions.

We propose a machine-translation approach to automatically generate a playlist title from a set of music tracks. We take a sequence of track IDs as input and a sequence of words in a playlist title as output, adapting the sequence-to-sequence framework based on Recurrent Neural Network (RNN) and Transformer to the music data. Considering the orderless nature of music tracks in a playlist, we propose two techniques that remove the order of the input sequence. One is data augmentation by shuffling and the other is deleting the positional encoding. We also reorganize the existing music playlist datasets to generate phrase-level playlist titles. The result shows that the Transformer models generally outperform the RNN model. Also, removing the order of input sequence improves the performance further.

In medical dialogue summarization, summaries must be coherent and must capture all the medically relevant information in the dialogue. However, learning effective models for summarization require large amounts of labeled data which is especially hard to obtain. We present an algorithm to create synthetic training data with an explicit focus on capturing medically relevant information. We utilize GPT-3 as the backbone of our algorithm and scale 210 human labeled examples to yield results comparable to using 6400 human labeled examples (~30x) leveraging low-shot learning and an ensemble method. In detailed experiments, we show that this approach produces high quality training data that can further be combined with human labeled data to get summaries that are strongly preferable to those produced by models trained on human data alone both in terms of medical accuracy and coherency.

Since the beginning of coronavirus, the disease has spread worldwide and drastically changed many aspects of the human's lifestyle. Twitter as a powerful tool can help researchers measure public health in response to COVID-19. According to the high volume of data production on social networks, automated text mining approaches can help search, read and summarize helpful information. This paper preprocessed the existing medical dataset regarding COVID-19 named CORD-19 and annotated the dataset for supervised classification tasks. At this time of the COVID-19 pandemic, we made a preprocessed dataset for the research community. This may contribute towards finding new solutions for some social interventions that COVID-19 has made. The preprocessed version of the mentioned dataset is publicly available through Github.

We present a memory-based model for context-dependent semantic parsing. Previous approaches focus on enabling the decoder to copy or modify the parse from the previous utterance, assuming there is a dependency between the current and previous parses. In this work, we propose to represent contextual information using an external memory. We learn a context memory controller that manages the memory by maintaining the cumulative meaning of sequential user utterances. We evaluate our approach on three semantic parsing benchmarks. Experimental results show that our model can better process context-dependent information and demonstrates improved performance without using task-specific decoders.

In various natural language processing tasks, passage retrieval and passage re-ranking are two key procedures in finding and ranking relevant information. Since both the two procedures contribute to the final performance, it is important to jointly optimize them in order to achieve mutual improvement. In this paper, we propose a novel joint training approach for dense passage retrieval and passage re-ranking. A major contribution is that we introduce the dynamic listwise distillation, where we design a unified listwise training approach for both the retriever and the re-ranker. During the dynamic distillation, the retriever and the re-ranker can be adaptively improved according to each other's relevance information. We also propose a hybrid data augmentation strategy to construct diverse training instances for listwise training approach. Extensive experiments show the effectiveness of our approach on both MSMARCO and Natural Questions datasets. Our code is available at https://github.com/PaddlePaddle/RocketQA.

We introduce semantic form mid-tuning, an approach for transferring semantic knowledge from semantic meaning representations into transformer-based language encoders. In mid-tuning, we learn to align the text of general sentences -- not tied to any particular inference task -- and structured semantic representations of those sentences. Our approach does not require gold annotated semantic representations. Instead, it makes use of automatically generated semantic representations, such as from off-the-shelf PropBank and FrameNet semantic parsers. We show that this alignment can be learned implicitly via classification or directly via triplet loss. Our method yields language encoders that demonstrate improved predictive performance across inference, reading comprehension, textual similarity, and other semantic tasks drawn from the GLUE, SuperGLUE, and SentEval benchmarks. We evaluate our approach on three popular baseline models, where our experimental results and analysis concludes that current pre-trained language models can further benefit from structured semantic frames with the proposed mid-tuning method, as they inject additional task-agnostic knowledge to the encoder, improving the generated embeddings as well as the linguistic properties of the given model, as evident from improvements on a popular sentence embedding toolkit and a variety of probing tasks.

Deep generative models have been widely used in several areas of NLP, and various techniques have been proposed to augment them or address their training challenges. In this paper, we propose a simple modification to Variational Autoencoders (VAEs) by using an Isotropic Gaussian Posterior (IGP) that allows for better utilisation of their latent representation space. This model avoids the sub-optimal behavior of VAEs related to inactive dimensions in the representation space. We provide both theoretical analysis, and empirical evidence on various datasets and tasks that show IGP leads to consistent improvement on several quantitative and qualitative grounds, from downstream task performance and sample efficiency to robustness. Additionally, we give insights about the representational properties encouraged by IGP and also show that its gain generalises to image domain as well.

Style transfer is the task of rewriting an input sentence into a target style while approximately preserving its content. While most prior literature assumes access to large style-labelled corpora, recent work (Riley et al. 2021) has attempted "few-shot" style transfer using only 3-10 sentences at inference for extracting the target style. In this work we consider one such low resource setting where no datasets are available: style transfer for Indian languages. We find that existing few-shot methods perform this task poorly, with a strong tendency to copy inputs verbatim. We push the state-of-the-art for few-shot style transfer with a new method modeling the stylistic difference between paraphrases. When compared to prior work using automatic and human evaluations, our model achieves 2-3x better performance and output diversity in formality transfer and code-mixing addition across five Indian languages. Moreover, our method is better able to control the amount of style transfer using an input scalar knob. We report promising qualitative results for several attribute transfer directions, including sentiment transfer, text simplification, gender neutralization and text anonymization, all without retraining the model. Finally we found model evaluation to be difficult due to the lack of evaluation datasets and metrics for Indian languages. To facilitate further research in formality transfer for Indic languages, we crowdsource annotations for 4000 sentence pairs in four languages, and use this dataset to design our automatic evaluation suite.

Automated audio captioning (AAC) is the task of automatically generating textual descriptions for general audio signals. A captioning system has to identify various information from the input signal and express it with natural language. Existing works mainly focus on investigating new methods and try to improve their performance measured on existing datasets. Having attracted attention only recently, very few works on AAC study the performance of existing pre-trained audio and natural language processing resources. In this paper, we evaluate the performance of off-the-shelf models with a Transformer-based captioning approach. We utilize the freely available Clotho dataset to compare four different pre-trained machine listening models, four word embedding models, and their combinations in many different settings. Our evaluation suggests that YAMNet combined with BERT embeddings produces the best captions. Moreover, in general, fine-tuning pre-trained word embeddings can lead to better performance. Finally, we show that sequences of audio embeddings can be processed using a Transformer encoder to produce higher-quality captions.

Distantly supervised relation extraction (DS-RE) is generally framed as a multi-instance multi-label (MI-ML) task, where the optimal aggregation of information from multiple instances is of key importance. Intra-bag attention (Lin et al., 2016) is an example of a popularly used aggregation scheme for this framework. Apart from this scheme, however, there is not much to choose from in the DS-RE literature as most of the advances in this field are focused on improving the instance-encoding step rather than the instance-aggregation step. With recent works leveraging large pre-trained language models as encoders, the increased capacity of models might allow for more flexibility in the instance-aggregation step. In this work, we explore this hypothesis and come up with a novel aggregation scheme which we call Passage-Att. Under this aggregation scheme, we combine all instances mentioning an entity pair into a "passage of instances", which is summarized independently for each relation class. These summaries are used to predict the validity of a potential triple. We show that our Passage-Att with BERT as passage encoder achieves state-of-the-art performance in three different settings (monolingual DS, monolingual DS with manually-annotated test set, multilingual DS).

Social concepts referring to non-physical objects--such as revolution, violence, or friendship--are powerful tools to describe, index, and query the content of visual data, including ever-growing collections of art images from the Cultural Heritage (CH) field. While much progress has been made towards complete image understanding in computer vision, automatic detection of social concepts evoked by images is still a challenge. This is partly due to the well-known semantic gap problem, worsened for social concepts given their lack of unique physical features, and reliance on more unspecific features than concrete concepts. In this paper, we propose the translation of recent cognitive theories about social concept representation into a software approach to represent them as multimodal frames, by integrating multisensory data. Our method focuses on the extraction, analysis, and integration of multimodal features from visual art material tagged with the concepts of interest. We define a conceptual model and present a novel ontology for formally representing social concepts as multimodal frames. Taking the Tate Gallery's collection as an empirical basis, we experiment our method on a corpus of art images to provide a proof of concept of its potential. We discuss further directions of research, and provide all software, data sources, and results.

Sensitivity of deep-neural models to input noise is known to be a challenging problem. In NLP, model performance often deteriorates with naturally occurring noise, such as spelling errors. To mitigate this issue, models may leverage artificially noised data. However, the amount and type of generated noise has so far been determined arbitrarily. We therefore propose to model the errors statistically from grammatical-error-correction corpora. We present a thorough evaluation of several state-of-the-art NLP systems' robustness in multiple languages, with tasks including morpho-syntactic analysis, named entity recognition, neural machine translation, a subset of the GLUE benchmark and reading comprehension. We also compare two approaches to address the performance drop: a) training the NLP models with noised data generated by our framework; and b) reducing the input noise with external system for natural language correction. The code is released at https://github.com/ufal/kazitext.

The sparsely-activated models have achieved great success in natural language processing through large-scale parameters and relatively low computational cost, and gradually become a feasible technique for training and implementing extremely large models. Due to the limit of communication cost, activating multiple experts is hardly affordable during training and inference. Therefore, previous work usually activate just one expert at a time to alleviate additional communication cost. Such routing mechanism limits the upper bound of model performance. In this paper, we first investigate a phenomenon that increasing the number of activated experts can boost the model performance with higher sparse ratio. To increase the number of activated experts without an increase in computational cost, we propose SAM (Switch and Mixture) routing, an efficient hierarchical routing mechanism that activates multiple experts in a same device (GPU). Our methods shed light on the training of extremely large sparse models and experiments prove that our models can achieve significant performance gain with great efficiency improvement.

Development of speech and language technology for social good (LT4SG), especially those targeted at the welfare of marginalized communities and speakers of low-resource and under-served languages, has been a prominent theme of research within NLP, Speech, and the AI communities. Researchers have mostly relied on their individual expertise, experiences or ad hoc surveys for prioritization of language technologies that provide social good to the end-users. This has been criticized by several scholars who argue that work on LT4SG must include the target linguistic communities during the design and development process. However, none of the LT4SG work and their critiques suggest principled techniques for prioritization of the technologies and methods for inclusion of the end-user during the development cycle. Drawing inspiration from the fields of Economics, Ethics, Psychology, and Participatory Design, here we chart out a set of methodologies for prioritizing LT4SG that are aligned with the end-user preferences. We then analyze several LT4SG efforts in light of the proposed methodologies and bring out their hidden assumptions and potential pitfalls. While the current study is limited to language technologies, we believe that the principles and prioritization techniques highlighted here are applicable more broadly to AI for Social Good.

When directly using existing text generation datasets for controllable generation, we are facing the problem of not having the domain knowledge and thus the aspects that could be controlled are limited.A typical example is when using CNN/Daily Mail dataset for controllable text summarization, there is no guided information on the emphasis of summary sentences. A more useful text generator should leverage both the input text and control variables to guide the generation, which can only be built with deep understanding of the domain knowledge. Motivated by this vi-sion, our paper introduces a new text generation dataset, named MReD. Our new dataset consists of 7,089 meta-reviews and all its 45k meta-review sentences are manually annotated as one of the carefully defined 9 categories, including abstract, strength, decision, etc. We present experimental results on start-of-the-art summarization models, and propose methods for controlled generation on both extractive and abstractive models using our annotated data. By exploring various settings and analaysing the model behavior with respect to the control inputs, we demonstrate the challenges and values of our dataset. MReD allows us to have a better understanding of the meta-review corpora and enlarge the research room for controllable text generation.

Event extraction is typically modeled as a multi-class classification problem where both event types and argument roles are treated as atomic symbols. These approaches are usually limited to a set of pre-defined types. We propose a novel event extraction framework that takes event types and argument roles as natural language queries to extract candidate triggers and arguments from the input text. With the rich semantics in the queries, our framework benefits from the attention mechanisms to better capture the semantic correlation between the event types or argument roles and the input text. Furthermore, the query-and-extract formulation allows our approach to leverage all available event annotations from various ontologies as a unified model. Experiments on two public benchmarks, ACE and ERE, demonstrate that our approach achieves state-of-the-art performance on each dataset and significantly outperforms existing methods on zero-shot event extraction. We will make all the programs publicly available once the paper is accepted.

In this paper, we present a pre-trained language model (PLM) based framework called RID for conversational recommender system (CRS). RID finetunes the large-scale PLMs such as DialoGPT, together with a pre-trained Relational Graph Convolutional Network (RGCN) to encode the node representations of an item-oriented knowledge graph. The former aims to generate fluent and diverse dialogue responses based on the strong language generation ability of PLMs, while the latter is to facilitate the item recommendation by learning better node embeddings on the structural knowledge base. To unify two modules of dialogue generation and item recommendation into a PLMs-based framework, we expand the generation vocabulary of PLMs to include an extra item vocabulary, and introduces a vocabulary pointer to control when to recommend target items in the generation process. Extensive experiments on the benchmark dataset ReDial show RID significantly outperforms the state-of-the-art methods on both evaluations of dialogue and recommendation.

Nested entities are observed in many domains due to their compositionality, which cannot be easily recognized by the widely-used sequence labeling framework. A natural solution is to treat the task as a span classification problem. To increase performance on span representation and classification, it is crucial to effectively integrate all useful information of different formats, which we refer to heterogeneous factors including tokens, labels, boundaries, and related spans. To fuse these heterogeneous factors, we propose a novel triaffine mechanism including triaffine attention and scoring, which interacts with multiple factors in both the stages of representation and classification. Experiments results show that our proposed method achieves the state-of-the-art F1 scores on four nested NER datasets: ACE2004, ACE2005, GENIA, and KBP2017.

While many studies have shown that linguistic information is encoded in hidden word representations, few have studied individual neurons, to show how and in which neurons it is encoded. Among these, the common approach is to use an external probe to rank neurons according to their relevance to some linguistic attribute, and to evaluate the obtained ranking using the same probe that produced it. We show two pitfalls in this methodology: 1. It confounds distinct factors: probe quality and ranking quality. We separate them and draw conclusions on each. 2. It focuses on encoded information, rather than information that is used by the model. We show that these are not the same. We compare two recent ranking methods and a simple one we introduce, and evaluate them with regard to both of these aspects.

The field of conversational agents is growing fast and there is an increasing need for algorithms that enhance natural interaction. In this work we show how we achieved state of the art results in the Keyword Spotting field by adapting and tweaking the Xception algorithm, which achieved outstanding results in several computer vision tasks. We obtained about 96\% accuracy when classifying audio clips belonging to 35 different categories, beating human annotation at the most complex tasks proposed.

How do we perform efficient inference while retaining high translation quality? Existing neural machine translation models, such as Transformer, achieve high performance, but they decode words one by one, which is inefficient. Recent non-autoregressive translation models speed up the inference, but their quality is still inferior. In this work, we propose DSLP, a highly efficient and high-performance model for machine translation. The key insight is to train a non-autoregressive Transformer with Deep Supervision and feed additional Layer-wise Predictions. We conducted extensive experiments on four translation tasks (both directions of WMT'14 EN-DE and WMT'16 EN-RO). Results show that our approach consistently improves the BLEU scores compared with respective base models. Specifically, our best variant outperforms the autoregressive model on three translation tasks, while being 14.8 times more efficient in inference.

Current open-domain conversational models can easily be made to talk in inadequate ways. Online learning from conversational feedback given by the conversation partner is a promising avenue for a model to improve and adapt, so as to generate fewer of these safety failures. However, current state-of-the-art models tend to react to feedback with defensive or oblivious responses. This makes for an unpleasant experience and may discourage conversation partners from giving feedback in the future. This work proposes SaFeRDialogues, a task and dataset of graceful responses to conversational feedback about safety failures. We collect a dataset of 10k dialogues demonstrating safety failures, feedback signaling them, and a response acknowledging the feedback. We show how fine-tuning on this dataset results in conversations that human raters deem considerably more likely to lead to a civil conversation, without sacrificing engagingness or general conversational ability.

Opinion summarization focuses on generating summaries that reflect popular opinions of multiple reviews for a single entity (e.g., a hotel or a product.) While generated summaries offer general and concise information about a particular entity, the information may be insufficient to help the user compare multiple entities. Thus, the user may still struggle with the question "Which one should I pick?" In this paper, we propose a {\em comparative opinion summarization} task, which is to generate two contrastive summaries and one common summary from two given sets of reviews from different entities. We develop a comparative summarization framework CoCoSum, which consists of two few-shot summarization models that are jointly used to generate contrastive and common summaries. Experimental results on a newly created benchmark CoCoTrip show that CoCoSum can produce high-quality contrastive and common summaries than state-of-the-art opinion summarization models.

Training dense passage representations via contrastive learning (CL) has been shown effective for Open-Domain Passage Retrieval (ODPR). Recent studies mainly focus on optimizing this CL framework by improving the sampling strategy or extra pretraining. Different from previous studies, this work devotes itself to investigating the influence of conflicts in the widely used CL strategy in ODPR, motivated by our observation that a passage can be organized by multiple semantically different sentences, thus modeling such a passage as a unified dense vector is not optimal. We call such conflicts Contrastive Conflicts. In this work, we propose to solve it with a representation decoupling method, by decoupling the passage representations into contextual sentence-level ones, and design specific CL strategies to mediate these conflicts. Experiments on widely used datasets including Natural Questions, Trivia QA, and SQuAD verify the effectiveness of our method, especially on the dataset where the conflicting problem is severe. Our method also presents good transferability across the datasets, which further supports our idea of mediating Contrastive Conflicts.

Neural rationale models are popular for interpretable predictions of NLP tasks. In these, a selector extracts segments of the input text, called rationales, and passes these segments to a classifier for prediction. Since the rationale is the only information accessible to the classifier, it is plausibly defined as the explanation. Is such a characterization unconditionally correct? In this paper, we argue to the contrary, with both philosophical perspectives and empirical evidence suggesting that rationale models are, perhaps, less rational and interpretable than expected. We call for more rigorous and comprehensive evaluations of these models to ensure desired properties of interpretability are indeed achieved. The code can be found at https://github.com/yimingz89/Neural-Rationale-Analysis.

Radiology reports are the main form of communication between radiologists and other clinicians, and contain important information for patient care. However in order to use this information for research it is necessary to convert the raw text into structured data suitable for analysis. Domain specific contextual word embeddings have been shown to achieve impressive accuracy at such natural language processing tasks in medicine. In this work we pre-trained a contextual embedding BERT model using breast radiology reports and developed a classifier that incorporated the embedding with auxiliary global textual features in order to perform a section tokenization task. This model achieved a 98% accuracy at segregating free text reports into sections of information outlined in the Breast Imaging Reporting and Data System (BI-RADS) lexicon, a significant improvement over the Classic BERT model without auxiliary information. We then evaluated whether using section tokenization improved the downstream extraction of the following fields: modality/procedure, previous cancer, menopausal status, purpose of exam, breast density and background parenchymal enhancement. Using the BERT model pre-trained on breast radiology reports combined with section tokenization resulted in an overall accuracy of 95.9% in field extraction. This is a 17% improvement compared to an overall accuracy of 78.9% for field extraction for models without section tokenization and with Classic BERT embeddings. Our work shows the strength of using BERT in radiology report analysis and the advantages of section tokenization in identifying key features of patient factors recorded in breast radiology reports.

Fine-tuning all parameters of a pre-trained model has become the mainstream approach for transfer learning. To increase its efficiency and prevent catastrophic forgetting and interference, techniques like adapters and sparse fine-tuning have been developed. Adapters are modular, as they can be combined to adapt a model towards different facets of knowledge (e.g., dedicated language and/or task adapters). Sparse fine-tuning is expressive, as it controls the behavior of all model components. In this work, we introduce a new fine-tuning method with both these desirable properties. In particular, we learn sparse, real-valued masks based on a simple variant of the Lottery Ticket Hypothesis. Task-specific masks are obtained from annotated data in a source language, and language-specific masks from masked language modeling in a target language. Both these masks can then be composed with the pre-trained model. Unlike adapter-based fine-tuning, this method neither increases the number of parameters at inference time nor alters the original model architecture. Most importantly, it outperforms adapters in zero-shot cross-lingual transfer by a large margin in a series of multilingual benchmarks, including Universal Dependencies, MasakhaNER, and AmericasNLI. Based on an in-depth analysis, we additionally find that sparsity is crucial to prevent both 1) interference between the fine-tunings to be composed and 2) overfitting. We release the code and models at https://github.com/cambridgeltl/composable-sft.

In attempts to develop sample-efficient algorithms, researcher have explored myriad mechanisms for collecting and exploiting feature feedback, auxiliary annotations provided for training (but not test) instances that highlight salient evidence. Examples include bounding boxes around objects and salient spans in text. Despite its intuitive appeal, feature feedback has not delivered significant gains in practical problems as assessed on iid holdout sets. However, recent works on counterfactually augmented data suggest an alternative benefit of supplemental annotations: lessening sensitivity to spurious patterns and consequently delivering gains in out-of-domain evaluations. Inspired by these findings, we hypothesize that while the numerous existing methods for incorporating feature feedback have delivered negligible in-sample gains, they may nevertheless generalize better out-of-domain. In experiments addressing sentiment analysis, we show that feature feedback methods perform significantly better on various natural out-of-domain datasets even absent differences on in-domain evaluation. By contrast, on natural language inference tasks, performance remains comparable. Finally, we compare those tasks where feature feedback does (and does not) help.

Semantic parsing is the task of producing a structured meaning representation for natural language utterances or questions. Recent research has pointed out that the commonly-used sequence-to-sequence (seq2seq) semantic parsers struggle to generalize systematically, i.e. to handle examples that require recombining known knowledge in novel settings. In this work, we show that better systematic generalization can be achieved by producing the meaning representation (MR) directly as a graph and not as a sequence. To this end we propose LAGr, the Labeling Aligned Graphs algorithm that produces semantic parses by predicting node and edge labels for a complete multi-layer input-aligned graph. The strongly-supervised LAGr algorithm requires aligned graphs as inputs, whereas weakly-supervised LAGr infers alignments for originally unaligned target graphs using an approximate MAP inference procedure. On the COGS and CFQ compositional generalization benchmarks the strongly- and weakly- supervised LAGr algorithms achieve significant improvements upon the baseline seq2seq parsers.

What would it take to teach a machine to behave ethically? While broad ethical rules may seem straightforward to state ("thou shalt not kill"), applying such rules to real-world situations is far more complex. For example, while "helping a friend" is generally a good thing to do, "helping a friend spread fake news" is not. We identify four underlying challenges towards machine ethics and norms: (1) an understanding of moral precepts and social norms; (2) the ability to perceive real-world situations visually or by reading natural language descriptions; (3) commonsense reasoning to anticipate the outcome of alternative actions in different contexts; (4) most importantly, the ability to make ethical judgments given the interplay between competing values and their grounding in different contexts (e.g., the right to freedom of expression vs. preventing the spread of fake news).

Our paper begins to address these questions within the deep learning paradigm. Our prototype model, Delphi, demonstrates strong promise of language-based commonsense moral reasoning, with up to 92.1% accuracy vetted by humans. This is in stark contrast to the zero-shot performance of GPT-3 of 52.3%, which suggests that massive scale alone does not endow pre-trained neural language models with human values. Thus, we present Commonsense Norm Bank, a moral textbook customized for machines, which compiles 1.7M examples of people's ethical judgments on a broad spectrum of everyday situations. In addition to the new resources and baseline performances for future research, our study provides new insights that lead to several important open research questions: differentiating between universal human values and personal values, modeling different moral frameworks, and explainable, consistent approaches to machine ethics.

Visually-grounded spoken language datasets can enable models to learn cross-modal correspondences with very weak supervision. However, modern audio-visual datasets contain biases that undermine the real-world performance of models trained on that data. We introduce Spoken ObjectNet, which is designed to remove some of these biases and provide a way to better evaluate how effectively models will perform in real-world scenarios. This dataset expands upon ObjectNet, which is a bias-controlled image dataset that features similar image classes to those present in ImageNet. We detail our data collection pipeline, which features several methods to improve caption quality, including automated language model checks. Lastly, we show baseline results on image retrieval and audio retrieval tasks. These results show that models trained on other datasets and then evaluated on Spoken ObjectNet tend to perform poorly due to biases in other datasets that the models have learned. We also show evidence that the performance decrease is due to the dataset controls, and not the transfer setting.

Conventional fine-tuning of pre-trained language models tunes all model parameters and stores a full model copy for each downstream task, which has become increasingly infeasible as the model size grows larger. Recent parameter-efficient language model tuning (PELT) methods manage to match the performance of fine-tuning with much fewer trainable parameters and perform especially well when the training data is limited. However, different PELT methods may perform rather differently on the same task, making it nontrivial to select the most appropriate method for a specific task, especially considering the fast-growing number of new PELT methods and downstream tasks. In light of model diversity and the difficulty of model selection, we propose a unified framework, UniPELT, which incorporates different PELT methods as submodules and learns to activate the ones that best suit the current data or task setup. Remarkably, on the GLUE benchmark, UniPELT consistently achieves 1~3pt gains compared to the best individual PELT method that it incorporates and even outperforms fine-tuning under different setups. Moreover, UniPELT often surpasses the upper bound when taking the best performance of all its submodules used individually on each task, indicating that a mixture of multiple PELT methods may be inherently more effective than single methods.

Dense retrieval (DR) methods conduct text retrieval by first encoding texts in the embedding space and then matching them by nearest neighbor search. This requires strong locality properties from the representation space, i.e, the close allocations of each small group of relevant texts, which are hard to generalize to domains without sufficient training data. In this paper, we aim to improve the generalization ability of DR models from source training domains with rich supervision signals to target domains without any relevant labels, in the zero-shot setting. To achieve that, we propose Momentum adversarial Domain Invariant Representation learning (MoDIR), which introduces a momentum method in the DR training process to train a domain classifier distinguishing source versus target, and then adversarially updates the DR encoder to learn domain invariant representations. Our experiments show that MoDIR robustly outperforms its baselines on 10+ ranking datasets from the BEIR benchmark in the zero-shot setup, with more than 10% relative gains on datasets with enough sensitivity for DR models' evaluation. Source code of this paper will be released.

One often wants to take an existing, trained NLP model and use it on data from a new domain. While fine-tuning or few-shot learning can be used to adapt the base model, there is no one simple recipe to getting these working; moreover, one may not have access to the original model weights if it is deployed as a black box. To this end, we study how to improve a black box model's performance on a new domain given examples from the new domain by leveraging explanations of the model's behavior. Our approach first extracts a set of features combining human intuition about the task with model attributions generated by black box interpretation techniques, and then uses a simple model to calibrate or rerank the model's predictions based on the features. We experiment with our method on two tasks, extractive question answering and natural language inference, covering adaptation from several pairs of domains. The experimental results across all the domain pairs show that explanations are useful for calibrating these models. We show that the calibration features transfer to some extent between tasks and shed light on how to effectively use them.

Toxic speech, also known as hate speech, is regarded as one of the crucial issues plaguing online social media today. Most recent work on toxic speech detection is constrained to the modality of text with no existing work on toxicity detection from spoken utterances. In this paper, we propose a new Spoken Language Processing task of detecting toxicity from spoken speech. We introduce DeToxy, the first publicly available toxicity annotated dataset for English speech, sourced from various openly available speech databases, consisting of over 2 million utterances. Finally, we also provide analysis on how a spoken speech corpus annotated for toxicity can help facilitate the development of E2E models which better capture various prosodic cues in speech, thereby boosting toxicity classification on spoken utterances.

Contemporary natural language processing (NLP) revolves around learning from latent document representations, generated either implicitly by neural language models or explicitly by methods such as doc2vec or similar. One of the key properties of the obtained representations is their dimension. Whilst the commonly adopted dimensions of 256 and 768 offer sufficient performance on many tasks, it is many times unclear whether the default dimension is the most suitable choice for the subsequent downstream learning tasks. Furthermore, representation dimensions are seldom subject to hyperparameter tuning due to computational constraints. The purpose of this paper is to demonstrate that a surprisingly simple and efficient recursive compression procedure can be sufficient to both significantly compress the initial representation, but also potentially improve its performance when considering the task of text classification. Having smaller and less noisy representations is the desired property during deployment, as orders of magnitude smaller models can significantly reduce the computational overload and with it the deployment costs. We propose CoRe, a straightforward, representation learner-agnostic framework suitable for representation compression. The CoRe's performance is showcased and studied on a collection of 17 real-life corpora from biomedical, news, social media, and literary domains. We explored CoRe's behavior when considering contextual and non-contextual document representations, different compression levels, and 9 different compression algorithms. Current results based on more than 100,000 compression experiments indicate that recursive Singular Value Decomposition offers a very good trade-off between the compression efficiency and performance, making CoRe useful in many existing, representation-dependent NLP pipelines.

Deep NLP models have been shown to learn spurious correlations, leaving them brittle to input perturbations. Recent work has shown that counterfactual or contrastive data -- i.e. minimally perturbed inputs -- can reveal these weaknesses, and that data augmentation using counterfactuals can help ameliorate them. Proposed techniques for generating counterfactuals rely on human annotations, perturbations based on simple heuristics, and meaning representation frameworks. We focus on the task of creating counterfactuals for question answering, which presents unique challenges related to world knowledge, semantic diversity, and answerability. To address these challenges, we develop a Retrieve-Generate-Filter(RGF) technique to create counterfactual evaluation and training data with minimal human supervision. Using an open-domain QA framework and question generation model trained on original task data, we create counterfactuals that are fluent, semantically diverse, and automatically labeled. Data augmentation with RGF counterfactuals improves performance on out-of-domain and challenging evaluation sets over and above existing methods, in both the reading comprehension and open-domain QA settings. Moreover, we find that RGF data leads to significant improvements in a model's robustness to local perturbations.

Prompt tuning, which only tunes continuous prompts with a frozen language model, substantially reduces per-task storage and memory usage at training. However, in the context of NLU, prior work and our results reveal that existing methods of prompt tuning do not perform well for normal-sized pre-trained models and for hard sequence tasks, indicating lack of universality. We present a novel empirical finding that properly-optimized prompt tuning can be universally effective across a wide range of model scales and NLU tasks, where it matches the performance of fine-tuning while having only 0.1\%-3\% tuned parameters. Our method P-Tuning v2 is not a new method but a version of prefix-tuning \cite{li2021prefix} optimized and adapted for NLU. Given the universality and simplicity of P-Tuning v2, we believe it can serve as an alternative for fine-tuning and a strong baseline for future research.

The goal of this paper is to learn strong lip reading models that can recognise speech in silent videos. Most prior works deal with the open-set visual speech recognition problem by adapting existing automatic speech recognition techniques on top of trivially pooled visual features. Instead, in this paper we focus on the unique challenges encountered in lip reading and propose tailored solutions. To that end we make the following contributions: (1) we propose an attention-based pooling mechanism to aggregate visual speech representations; (2) we use sub-word units for lip reading for the first time and show that this allows us to better model the ambiguities of the task; (3) we propose a training pipeline that balances the lip reading performance with other key factors such as data and compute efficiency. Following the above, we obtain state-of-the-art results on the challenging LRS2 and LRS3 benchmarks when training on public datasets, and even surpass models trained on large-scale industrial datasets by using an order of magnitude less data. Our best model achieves 22.6% word error rate on the LRS2 dataset, a performance unprecedented for lip reading models, significantly reducing the performance gap between lip reading and automatic speech recognition.

During the last fifteen years, automatic text scaling has become one of the key tools of the Text as Data community in political science. Prominent text scaling algorithms, however, rely on the assumption that latent positions can be captured just by leveraging the information about word frequencies in documents under study. We challenge this traditional view and present a new, semantically aware text scaling algorithm, SemScale, which combines recent developments in the area of computational linguistics with unsupervised graph-based clustering. We conduct an extensive quantitative analysis over a collection of speeches from the European Parliament in five different languages and from two different legislative terms, and show that a scaling approach relying on semantic document representations is often better at capturing known underlying political dimensions than the established frequency-based (i.e., symbolic) scaling method. We further validate our findings through a series of experiments focused on text preprocessing and feature selection, document representation, scaling of party manifestos, and a supervised extension of our algorithm. To catalyze further research on this new branch of text scaling methods, we release a Python implementation of SemScale with all included data sets and evaluation procedures.

In recent years, Question Answering systems have become more popular and widely used by users. Despite the increasing popularity of these systems, the their performance is not even sufficient for textual data and requires further research. These systems consist of several parts that one of them is the Answer Selection component. This component detects the most relevant answer from a list of candidate answers. The methods presented in previous researches have attempted to provide an independent model to undertake the answer-selection task. An independent model cannot comprehend the syntactic and semantic features of questions and answers with a small training dataset. To fill this gap, language models can be employed in implementing the answer selection part. This action enables the model to have a better understanding of the language in order to understand questions and answers better than previous works. In this research, we will present the "BAS" (BERT Answer Selection) that uses the BERT language model to comprehend language. The empirical results of applying the model on the TrecQA Raw, TrecQA Clean, and WikiQA datasets demonstrate that using a robust language model such as BERT can enhance the performance. Using a more robust classifier also enhances the effect of the language model on the answer selection component. The results demonstrate that language comprehension is an essential requirement in natural language processing tasks such as answer-selection.

Deep neural networks are becoming more and more popular due to their revolutionary success in diverse areas, such as computer vision, natural language processing, and speech recognition. However, the decision-making processes of these models are generally not interpretable to users. In various domains, such as healthcare, finance, or law, it is critical to know the reasons behind a decision made by an artificial intelligence system. Therefore, several directions for explaining neural models have recently been explored. In this thesis, I investigate two major directions for explaining deep neural networks. The first direction consists of feature-based post-hoc explanatory methods, that is, methods that aim to explain an already trained and fixed model (post-hoc), and that provide explanations in terms of input features, such as tokens for text and superpixels for images (feature-based). The second direction consists of self-explanatory neural models that generate natural language explanations, that is, models that have a built-in module that generates explanations for the predictions of the model.

Coupled with the availability of large scale datasets, deep learning architectures have enabled rapid progress on the Question Answering task. However, most of those datasets are in English, and the performances of state-of-the-art multilingual models are significantly lower when evaluated on non-English data. Due to high data collection costs, it is not realistic to obtain annotated data for each language one desires to support.

We propose a method to improve the Cross-lingual Question Answering performance without requiring additional annotated data, leveraging Question Generation models to produce synthetic samples in a cross-lingual fashion. We show that the proposed method allows to significantly outperform the baselines trained on English data only. We report a new state-of-the-art on four multilingual datasets: MLQA, XQuAD, SQuAD-it and PIAF (fr).

Despite considerable advancements with deep neural language models (LMs), neural text generation still suffers from degeneration: the generated text is repetitive, generic, self-contradictory, and often lacks commonsense. Our analyses on sentence-level attention patterns in LMs reveal that neural degeneration may be associated with insufficient learning of task-specific characteristics by the attention mechanism. This finding motivates on-the-fly attention modulation -- a simple but effective method that enables the injection of priors into attention computation during inference. Automatic and human evaluation results on three text generation benchmarks demonstrate that attention modulation helps LMs generate text with enhanced fluency, creativity, and commonsense reasoning, in addition to significantly reduce sentence-level repetition.

The COVID-19 pandemic is a global crisis that has been testing every society and exposing the critical role of local politics in crisis response. In the United States, there has been a strong partisan divide between the Democratic and Republican party's narratives about the pandemic which resulted in polarization of individual behaviors and divergent policy adoption across regions. As shown in this case, as well as in most major social issues, strongly polarized narrative frameworks facilitate such narratives. To understand polarization and other social chasms, it is critical to dissect these diverging narratives. Here, taking the Democratic and Republican political social media posts about the pandemic as a case study, we demonstrate that a combination of computational methods can provide useful insights into the different contexts, framing, and characters and relationships that construct their narrative frameworks which individual posts source from. Leveraging a dataset of tweets from elite politicians in the U.S., we found that the Democrats' narrative tends to be more concerned with the pandemic as well as financial and social support, while the Republicans discuss more about other political entities such as China. We then perform an automatic framing analysis to characterize the ways in which they frame their narratives, where we found that the Democrats emphasize the government's role in responding to the pandemic, and the Republicans emphasize the roles of individuals and support for small businesses. Finally, we present a semantic role analysis that uncovers the important characters and relationships in their narratives as well as how they facilitate a membership categorization process. Our findings concretely expose the gaps in the "elusive consensus" between the two parties. Our methodologies may be applied to computationally study narratives in various domains.

Even to a simple and short news headline, readers react in a multitude of ways: cognitively (e.g., inferring the writer's intent), emotionally (e.g., feeling distrust), and behaviorally (e.g., sharing the news with their friends). Such reactions are instantaneous and yet complex, as they rely on factors that go beyond interpreting the factual content the news headline. Instead, understanding reactions require pragmatic understanding of the news headline, including broader background knowledge about contentious news topics as well as commonsense reasoning about people's intents and emotional reactions. We propose Misinfo Reaction Frames, a pragmatic formalism for modeling how readers might react to a news headline cognitively, emotionally, and behaviorally. We also introduce a Misinfo Reaction Frames corpus, a dataset of over 200k news headline annotated with crowdsourced reactions focusing on global crises: the Covid-19 pandemic, climate change, and cancer. Empirical results confirm that it is indeed possible to learn the prominent patterns of readers' reactions to news headlines. We also find a potentially positive use case of our model; When we present our model generated inferences to people, we find that the machine inferences can increase readers' trust in real news while decreasing their trust in misinformation. Our work demonstrates the feasibility and the importance of pragmatic inferences of news to help enhance AI-guided misinformation detection and mitigation.

The audio segmentation mismatch between training data and those seen at run-time is a major problem in direct speech translation. Indeed, while systems are usually trained on manually segmented corpora, in real use cases they are often presented with continuous audio requiring automatic (and sub-optimal) segmentation. After comparing existing techniques (VAD-based, fixed-length and hybrid segmentation methods), in this paper we propose enhanced hybrid solutions to produce better results without sacrificing latency. Through experiments on different domains and language pairs, we show that our methods outperform all the other techniques, reducing by at least 30% the gap between the traditional VAD-based approach and optimal manual segmentation.

In recent times, we have seen an increased use of text chat for communication on social networks and smartphones. This particularly involves the use of Hindi-English code-mixed text which contains words which are not recognized in English vocabulary. We have worked on detecting emotions in these mixed data and classify the sentences in human emotions which are angry, fear, happy or sad. We have used state of the art natural language processing models and compared their performance on the dataset comprising sentences in this mixed data. The dataset was collected and annotated from sources and then used to train the models.

We present RobeCzech, a monolingual RoBERTa language representation model trained on Czech data. RoBERTa is a robustly optimized Transformer-based pretraining approach. We show that RobeCzech considerably outperforms equally-sized multilingual and Czech-trained contextualized language representation models, surpasses current state of the art in all five evaluated NLP tasks and reaches state-of-the-art results in four of them. The RobeCzech model is released publicly at https://hdl.handle.net/11234/1-3691 and https://huggingface.co/ufal/robeczech-base.

While unbiased machine learning models are essential for many applications, bias is a human-defined concept that can vary across tasks. Given only input-label pairs, algorithms may lack sufficient information to distinguish stable (causal) features from unstable (spurious) features. However, related tasks often share similar biases -- an observation we may leverage to develop stable classifiers in the transfer setting. In this work, we explicitly inform the target classifier about unstable features in the source tasks. Specifically, we derive a representation that encodes the unstable features by contrasting different data environments in the source task. We achieve robustness by clustering data of the target task according to this representation and minimizing the worst-case risk across these clusters. We evaluate our method on both text and image classifications. Empirical results demonstrate that our algorithm is able to maintain robustness on the target task, outperforming the best baseline by 22.9% in absolute accuracy across 12 transfer settings. Our code is available at https://github.com/YujiaBao/Tofu.

Knowledgeable FAQ chatbots are a valuable resource to any organization. While powerful and efficient retrieval-based models exist for English, it is rarely the case for other languages for which the same amount of training data is not available. In this paper, we propose a novel pre-training procedure to adapt ConveRT, an English conversational retriever model, to other languages with less training data available. We apply it for the first time to the task of Dutch FAQ answering related to the COVID-19 vaccine. We show it performs better than an open-source alternative in both a low-data regime and a high-data regime.

Recent researches have demonstrated that BERT shows potential in a wide range of natural language processing tasks. It is adopted as an encoder for many state-of-the-art automatic summarizing systems, which achieve excellent performance. However, so far, there is not much work done for Vietnamese. In this paper, we showcase how BERT can be implemented for extractive text summarization in Vietnamese on multi-document. We introduce a novel comparison between different multilingual and monolingual BERT models. The experiment results indicate that monolingual models produce promising results compared to other multilingual models and previous text summarizing models for Vietnamese.

This paper presents the results and analyses stemming from the first VoicePrivacy 2020 Challenge which focuses on developing anonymization solutions for speech technology. We provide a systematic overview of the challenge design with an analysis of submitted systems and evaluation results. In particular, we describe the voice anonymization task and datasets used for system development and evaluation. Also, we present different attack models and the associated objective and subjective evaluation metrics. We introduce two anonymization baselines and provide a summary description of the anonymization systems developed by the challenge participants. We report objective and subjective evaluation results for baseline and submitted systems. In addition, we present experimental results for alternative privacy metrics and attack models developed as a part of the post-evaluation analysis. Finally, we summarize our insights and observations that will influence the design of the next VoicePrivacy challenge edition and some directions for future voice anonymization research.

Phrase representations derived from BERT often do not exhibit complex phrasal compositionality, as the model relies instead on lexical similarity to determine semantic relatedness. In this paper, we propose a contrastive fine-tuning objective that enables BERT to produce more powerful phrase embeddings. Our approach (Phrase-BERT) relies on a dataset of diverse phrasal paraphrases, which is automatically generated using a paraphrase generation model, as well as a large-scale dataset of phrases in context mined from the Books3 corpus. Phrase-BERT outperforms baselines across a variety of phrase-level similarity tasks, while also demonstrating increased lexical diversity between nearest neighbors in the vector space. Finally, as a case study, we show that Phrase-BERT embeddings can be easily integrated with a simple autoencoder to build a phrase-based neural topic model that interprets topics as mixtures of words and phrases by performing a nearest neighbor search in the embedding space. Crowdsourced evaluations demonstrate that this phrase-based topic model produces more coherent and meaningful topics than baseline word and phrase-level topic models, further validating the utility of Phrase-BERT.

While diverse question answering (QA) datasets have been proposed and contributed significantly to the development of deep learning models for QA tasks, the existing datasets fall short in two aspects. First, we lack QA datasets covering complex questions that involve answers as well as the reasoning processes to get the answers. As a result, the state-of-the-art QA research on numerical reasoning still focuses on simple calculations and does not provide the mathematical expressions or evidences justifying the answers. Second, the QA community has contributed much effort to improving the interpretability of QA models. However, these models fail to explicitly show the reasoning process, such as the evidence order for reasoning and the interactions between different pieces of evidence. To address the above shortcomings, we introduce NOAHQA, a conversational and bilingual QA dataset with questions requiring numerical reasoning with compound mathematical expressions. With NOAHQA, we develop an interpretable reasoning graph as well as the appropriate evaluation metric to measure the answer quality. We evaluate the state-of-the-art QA models trained using existing QA datasets on NOAHQA and show that the best among them can only achieve 55.5 exact match scores, while the human performance is 89.7. We also present a new QA model for generating a reasoning graph where the reasoning graph metric still has a large gap compared with that of humans, e.g., 28 scores.

Biomedical named entity recognition (NER) is a key task in the extraction of information from biomedical literature and electronic health records. For this task, both generic and biomedical BERT models are widely used. Robustness of these models is vital for medical applications, such as automated medical decision making. In this paper we investigate the vulnerability of BERT models to variation in input data for NER through adversarial attack. Since adversarial attack methods for NER are sparse, we propose two black-box methods for NER based on existing methods for classification tasks. Experimental results show that the original as well as the biomedical BERT models are highly vulnerable to entity replacement: They can be fooled in 89.2 to 99.4% of the cases to mislabel previously correct entities. BERT models are also vulnerable to variation in the entity context with 20.2 to 45.0% of entities predicted completely wrong and another 29.3 to 53.3% of entities predicted wrong partially. Often a single change is sufficient to fool the model. BERT models seem most vulnerable to changes in the local context of entities. Of the biomedical BERT models, the vulnerability of BioBERT is comparable to the original BERT model whereas SciBERT is even more vulnerable. Our results chart the vulnerabilities of BERT models for biomedical NER and emphasize the importance of further research into uncovering and reducing these weaknesses.

Deep-learning based Automatic Essay Scoring (AES) systems are being actively used by states and language testing agencies alike to evaluate millions of candidates for life-changing decisions ranging from college applications to visa approvals. However, little research has been put to understand and interpret the black-box nature of deep-learning based scoring algorithms. Previous studies indicate that scoring models can be easily fooled. In this paper, we explore the reason behind their surprising adversarial brittleness. We utilize recent advances in interpretability to find the extent to which features such as coherence, content, vocabulary, and relevance are important for automated scoring mechanisms. We use this to investigate the oversensitivity i.e., large change in output score with a little change in input essay content) and overstability i.e., little change in output scores with large changes in input essay content) of AES. Our results indicate that autoscoring models, despite getting trained as "end-to-end" models with rich contextual embeddings such as BERT, behave like bag-of-words models. A few words determine the essay score without the requirement of any context making the model largely overstable. This is in stark contrast to recent probing studies on pre-trained representation learning models, which show that rich linguistic features such as parts-of-speech and morphology are encoded by them. Further, we also find that the models have learnt dataset biases, making them oversensitive. To deal with these issues, we propose detection-based protection models that can detect oversensitivity and overstability causing samples with high accuracies. We find that our proposed models are able to detect unusual attribution patterns and flag adversarial samples successfully.

We study multilingual AMR parsing from the perspective of knowledge distillation, where the aim is to learn and improve a multilingual AMR parser by using an existing English parser as its teacher. We constrain our exploration in a strict multilingual setting: there is but one model to parse all different languages including English. We identify that noisy input and precise output are the key to successful distillation. Together with extensive pre-training, we obtain an AMR parser whose performances surpass all previously published results on four different foreign languages, including German, Spanish, Italian, and Chinese, by large margins (up to 18.8 \textsc{Smatch} points on Chinese and on average 11.3 \textsc{Smatch} points). Our parser also achieves comparable performance on English to the latest state-of-the-art English-only parser.

Self- and semi-supervised learning methods have been actively investigated to reduce labeled training data or enhance the model performance. However, the approach mostly focus on in-domain performance for public datasets. In this study, we utilize the combination of self- and semi-supervised learning methods to solve unseen domain adaptation problem in a large-scale production setting for online ASR model. This approach demonstrates that using the source domain data with a small fraction of the target domain data (3%) can recover the performance gap compared to a full data baseline: relative 13.5% WER improvement for target domain data.

To address the performance gap of English ASR models on L2 English speakers, we evaluate fine-tuning of pretrained wav2vec 2.0 models (Baevski et al., 2020; Xu et al., 2021) on L2-ARCTIC, a non-native English speech corpus (Zhao et al., 2018) under different training settings. We compare \textbf{(a)} models trained with a combination of diverse accents to ones trained with only specific accents and \textbf{(b)} results from different single-accent models. Our experiments demonstrate the promise of developing ASR models for non-native English speakers, even with small amounts of L2 training data and even without a language model. Our models also excel in the zero-shot setting where we train on multiple L2 datasets and test on a blind L2 test set.

Trained on the large corpus, pre-trained language models (PLMs) can capture different levels of concepts in context and hence generate universal language representations. They can benefit multiple downstream natural language processing (NLP) tasks. Although PTMs have been widely used in most NLP applications, especially for high-resource languages such as English, it is under-represented in Lao NLP research. Previous work on Lao has been hampered by the lack of annotated datasets and the sparsity of language resources. In this work, we construct a text classification dataset to alleviate the resource-scare situation of the Lao language. We additionally present the first transformer-based PTMs for Lao with four versions: BERT-small, BERT-base, ELECTRA-small and ELECTRA-base, and evaluate it over two downstream tasks: part-of-speech tagging and text classification. Experiments demonstrate the effectiveness of our Lao models. We will release our models and datasets to the community, hoping to facilitate the future development of Lao NLP applications.

Our world is constantly evolving, and so is the content on the web. Consequently, our languages, often said to mirror the world, are dynamic in nature. However, most current contextual language models are static and cannot adapt to changes over time. In this work, we propose a temporal contextual language model called TempoBERT, which uses time as an additional context of texts. Our technique is based on modifying texts with temporal information and performing time masking - specific masking for the supplementary time information. We leverage our approach for the tasks of semantic change detection and sentence time prediction, experimenting on diverse datasets in terms of time, size, genre, and language. Our extensive evaluation shows that both tasks benefit from exploiting time masking.

Although pre-trained models (PLMs) have achieved remarkable improvements in a wide range of NLP tasks, they are expensive in terms of time and resources. This calls for the study of training more efficient models with less computation but still ensures impressive performance. Instead of pursuing a larger scale, we are committed to developing lightweight yet more powerful models trained with equal or less computation and friendly to rapid deployment. This technical report releases our pre-trained model called Mengzi, which stands for a family of discriminative, generative, domain-specific, and multimodal pre-trained model variants, capable of a wide range of language and vision tasks. Compared with public Chinese PLMs, Mengzi is simple but more powerful. Our lightweight model has achieved new state-of-the-art results on the widely-used CLUE benchmark with our optimized pre-training and fine-tuning techniques. Without modifying the model architecture, our model can be easily employed as an alternative to existing PLMs. Our sources are available at https://github.com/Langboat/Mengzi.