ACL 2022 Papers

Recent work has shown that feed-forward networks (FFNs) in pre-trained Transformers are a key component, storing various linguistic and factual knowledge. However, the computational patterns of FFNs are still unclear. In this work, we study the computational patterns of FFNs and observe that most inputs only activate a tiny ratio of neurons of FFNs. This phenomenon is similar to the sparsity of the human brain, which drives research on functional partitions of the human brain. To verify whether functional partitions also emerge in FFNs, we propose to convert a model into its MoE version with the same parameters, namely MoEfication. Specifically, MoEfication consists of two phases: (1) splitting the parameters of FFNs into multiple functional partitions as experts, and (2) building expert routers to decide which experts will be used for each input. Experimental results show that MoEfication can conditionally use 10% to 30% of FFN parameters while maintaining over 95% original performance for different models on various downstream tasks. Besides, MoEfication brings two advantages: (1) it significantly reduces the FLOPS of inference, i.e., 2x speedup with 25% of FFN parameters, and (2) it provides a fine-grained perspective to study the inner mechanism of FFNs. The source code of this paper can be obtained from https://github.com/thunlp/MoEfication.

Recent work on controlled text generation has either required attribute-based fine-tuning of the base language model (LM), or has restricted the parameterization of the attribute discriminator to be compatible with the base autoregressive LM. In this work, we propose Mix and Match LM, a global score-based alternative for controllable text generation that combines arbitrary pre-trained black-box models for achieving the desired attributes in the generated text without involving any fine-tuning or structural assumptions about the black-box models. We interpret the task of controllable generation as drawing samples from an energy-based model whose energy values are a linear combination of scores from black-box models that are separately responsible for fluency, the control attribute, and faithfulness to any conditioning context. We use a Metropolis-Hastings sampling scheme to sample from this energy-based model using bidirectional context and global attribute features. We validate the effectiveness of our approach on various controlled generation and style-based text revision tasks by outperforming recently proposed methods that involve extra training, fine-tuning, or restrictive assumptions over the form of models.

Neural networks are widely used in various NLP tasks for their remarkable performance. However, the complexity makes them difficult to interpret, i.e., they are not guaranteed right for the right reason. Besides the complexity, we reveal that the model pathology - the inconsistency between word saliency and model confidence, further hurts the interpretability. We show that the pathological inconsistency is caused by the representation collapse issue, which means that the representation of the sentences with tokens in different saliency reduced is somehow collapsed, and thus the important words cannot be distinguished from unimportant words in terms of model confidence changing. In this paper, to mitigate the pathology and obtain more interpretable models, we propose Pathological Contrastive Training (PCT) framework, which adopts contrastive learning and saliency-based samples augmentation to calibrate the sentences representation. Combined with qualitative analysis, we also conduct extensive quantitative experiments and measure the interpretability with eight reasonable metrics. Experiments show that our method can mitigate the model pathology and generate more interpretable models while keeping the model performance. Ablation study also shows the effectiveness.

Language models excel at generating coherent text, and model compression techniques such as knowledge distillation have enabled their use in resource-constrained settings. However, these models can be biased in multiple ways, including the unfounded association of male and female genders with gender-neutral professions. Therefore, knowledge distillation without any fairness constraints may preserve or exaggerate the teacher model’s biases onto the distilled model. To this end, we present a novel approach to mitigate gender disparity in text generation by learning a fair model during knowledge distillation. We propose two modifications to the base knowledge distillation based on counterfactual role reversal—modifying teacher probabilities and augmenting the training set. We evaluate gender polarity across professions in open-ended text generated from the resulting distilled and finetuned GPT–2 models and demonstrate a substantial reduction in gender disparity with only a minor compromise in utility. Finally, we observe that language models that reduce gender polarity in language generation do not improve embedding fairness or downstream classification fairness.

Chatbot models have achieved remarkable progress in recent years but tend to yield contradictory responses. In this paper, we exploit the advantage of contrastive learning technique to mitigate this issue. To endow the model with the ability of discriminating contradictory patterns, we minimize the similarity between the target response and contradiction related negative example. The negative example is generated with learnable latent noise, which receives contradiction related feedback from the pretrained critic. Experimental results show that our method helps to avoid contradictions in response generation while preserving response fluency, outperforming existing methods on both automatic and human evaluation.

Dialogue state tracking (DST) aims to extract essential information from multi-turn dialog situations and take appropriate actions. A belief state, one of the core pieces of information, refers to the subject and its specific content, and appears in the form of domain-slot-value. The trained model predicts “accumulated” belief states in every turn, and joint goal accuracy and slot accuracy are mainly used to evaluate the prediction; however, we specify that the current evaluation metrics have a critical limitation when evaluating belief states accumulated as the dialogue proceeds, especially in the most used MultiWOZ dataset. Additionally, we propose relative slot accuracy to complement existing metrics. Relative slot accuracy does not depend on the number of predefined slots, and allows intuitive evaluation by assigning relative scores according to the turn of each dialog. This study also encourages not solely the reporting of joint goal accuracy, but also various complementary metrics in DST tasks for the sake of a realistic evaluation.

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 factual content of news. We propose Misinfo Reaction Frames (MRF), a pragmatic formalism for modeling how readers might react to a news headline. In contrast to categorical schema, our free-text dimensions provide a more nuanced way of understanding intent beyond being benign or malicious. We also introduce a Misinfo Reaction Frames corpus, a crowdsourced dataset of reactions to over 25k news headlines focusing on global crises: the Covid-19 pandemic, climate change, and cancer. Empirical results confirm that it is indeed possible for neural models to predict the prominent patterns of readers’ reactions to previously unseen news headlines. Additionally, our user study shows that displaying machine-generated MRF implications alongside news headlines to readers can increase their trust in real news while decreasing their trust in misinformation. Our work demonstrates the feasibility and importance of pragmatic inferences on news headlines to help enhance AI-guided misinformation detection and mitigation.

Human languages are full of metaphorical expressions. Metaphors help people understand the world by connecting new concepts and domains to more familiar ones. Large pre-trained language models (PLMs) are therefore assumed to encode metaphorical knowledge useful for NLP systems. In this paper, we investigate this hypothesis for PLMs, by probing metaphoricity information in their encodings, and by measuring the cross-lingual and cross-dataset generalization of this information. We present studies in multiple metaphor detection datasets and in four languages (i.e., English, Spanish, Russian, and Farsi). Our extensive experiments suggest that contextual representations in PLMs do encode metaphorical knowledge, and mostly in their middle layers. The knowledge is transferable between languages and datasets, especially when the annotation is consistent across training and testing sets. Our findings give helpful insights for both cognitive and NLP scientists.

Popular language models (LMs) struggle to capture knowledge about rare tail facts and entities. Since widely used systems such as search and personal-assistants must support the long tail of entities that users ask about, there has been significant effort towards enhancing these base LMs with factual knowledge. We observe proposed methods typically start with a base LM and data that has been annotated with entity metadata, then change the model, by modifying the architecture or introducing auxiliary loss terms to better capture entity knowledge. In this work, we question this typical process and ask to what extent can we match the quality of model modifications, with a simple alternative: using a base LM and only changing the data. We propose metadata shaping, a method which inserts substrings corresponding to the readily available entity metadata, e.g. types and descriptions, into examples at train and inference time based on mutual information. Despite its simplicity, metadata shaping is quite effective. On standard evaluation benchmarks for knowledge-enhanced LMs, the method exceeds the base-LM baseline by an average of 4.3 F1 points and achieves state-of-the-art results. We further show the gains are on average 4.4x larger for the slice of examples containing tail vs. popular entities.

Task weighting, which assigns weights on the including tasks during training, significantly matters the performance of Multi-task Learning (MTL); thus, recently, there has been an explosive interest in it. However, existing task weighting methods assign weights only based on the training loss, while ignoring the gap between the training loss and generalization loss. It degenerates MTL’s performance. To address this issue, the present paper proposes a novel task weighting algorithm, which automatically weights the tasks via a learning-to-learn paradigm, referred to as MetaWeighting. Extensive experiments are conducted to validate the superiority of our proposed method in multi-task text classification.

The goal of meta-learning is to learn to adapt to a new task with only a few labeled examples. Inspired by the recent progress in large language models, we propose in-context tuning (ICT), which recasts task adaptation and prediction as a simple sequence prediction problem: to form the input sequence, we concatenate the task instruction, labeled in-context examples, and the target input to predict; to meta-train the model to learn from in-context examples, we fine-tune a pre-trained language model (LM) to predict the target label given the input sequence on a collection of tasks.We benchmark our method on two collections of text classification tasks: LAMA and BinaryClfs. Compared to MAML which adapts the model through gradient descent, our method leverages the inductive bias of pre-trained LMs to perform pattern matching, and outperforms MAML by an absolute 6% average AUC-ROC score on BinaryClfs, gaining more advantage with increasing model size. Compared to non-fine-tuned in-context learning (i.e. prompting a raw LM), in-context tuning meta-trains the model to learn from in-context examples. On BinaryClfs, ICT improves the average AUC-ROC score by an absolute 10%, and reduces the variance due to example ordering by 6x and example choices by 2x.

Recently, the NLP community has witnessed a rapid advancement in multilingual and cross-lingual transfer research where the supervision is transferred from high-resource languages (HRLs) to low-resource languages (LRLs). However, the cross-lingual transfer is not uniform across languages, particularly in the zero-shot setting. Towards this goal, one promising research direction is to learn shareable structures across multiple tasks with limited annotated data. The downstream multilingual applications may benefit from such a learning setup as most of the languages across the globe are low-resource and share some structures with other languages. In this paper, we propose a novel meta-learning framework (called Meta-XNLG) to learn shareable structures from typologically diverse languages based on meta-learning and language clustering. This is a step towards uniform cross-lingual transfer for unseen languages. We first cluster the languages based on language representations and identify the centroid language of each cluster. Then, a meta-learning algorithm is trained with all centroid languages and evaluated on the other languages in the zero-shot setting. We demonstrate the effectiveness of this modeling on two NLG tasks (Abstractive Text Summarization and Question Generation), 5 popular datasets and 30 typologically diverse languages. Consistent improvements over strong baselines demonstrate the efficacy of the proposed framework. The careful design of the model makes this end-to-end NLG setup less vulnerable to the accidental translation problem, which is a prominent concern in zero-shot cross-lingual NLG tasks.

Meta-learning, or learning to learn, is a technique that can help to overcome resource scarcity in cross-lingual NLP problems, by enabling fast adaptation to new tasks. We apply model-agnostic meta-learning (MAML) to the task of cross-lingual dependency parsing. We train our model on a diverse set of languages to learn a parameter initialization that can adapt quickly to new languages. We find that meta-learning with pre-training can significantly improve upon the performance of language transfer and standard supervised learning baselines for a variety of unseen, typologically diverse, and low-resource languages, in a few-shot learning setup.

State-of-the-art pre-trained language models have been shown to memorise facts and perform well with limited amounts of training data. To gain a better understanding of how these models learn, we study their generalisation and memorisation capabilities in noisy and low-resource scenarios. We find that the training of these models is almost unaffected by label noise and that it is possible to reach near-optimal results even on extremely noisy datasets. However, our experiments also show that they mainly learn from high-frequency patterns and largely fail when tested on low-resource tasks such as few-shot learning and rare entity recognition. To mitigate such limitations, we propose an extension based on prototypical networks that improves performance in low-resource named entity recognition tasks.

We introduce MemSum (Multi-step Episodic Markov decision process extractive SUMmarizer), a reinforcement-learning-based extractive summarizer enriched at each step with information on the current extraction history. When MemSum iteratively selects sentences into the summary, it considers a broad information set that would intuitively also be used by humans in this task: 1) the text content of the sentence, 2) the global text context of the rest of the document, and 3) the extraction history consisting of the set of sentences that have already been extracted. With a lightweight architecture, MemSum obtains state-of-the-art test-set performance (ROUGE) in summarizing long documents taken from PubMed, arXiv, and GovReport. Ablation studies demonstrate the importance of local, global, and history information. A human evaluation confirms the high quality and low redundancy of the generated summaries, stemming from MemSum’s awareness of extraction history.

Being able to reliably estimate self-disclosure – a key component of friendship and intimacy – from language is important for many psychology studies. We build single-task models on five self-disclosure corpora, but find that these models generalize poorly; the within-domain accuracy of predicted message-level self-disclosure of the best-performing model (mean Pearson’s r=0.69) is much higher than the respective across data set accuracy (mean Pearson’s r=0.32), due to both variations in the corpora (e.g., medical vs. general topics) and labeling instructions (target variables: self-disclosure, emotional disclosure, intimacy). However, some lexical features, such as expression of negative emotions and use of first person personal pronouns such as ‘I’ reliably predict self-disclosure across corpora. We develop a multi-task model that yields better results, with an average Pearson’s r of 0.37 for out-of-corpora prediction.

There is a growing interest in the combined use of NLP and machine learning methods to predict gaze patterns during naturalistic reading. While promising results have been obtained through the use of transformer-based language models, little work has been undertaken to relate the performance of such models to general text characteristics. In this paper we report on experiments with two eye-tracking corpora of naturalistic reading and two language models (BERT and GPT-2). In all experiments, we test effects of a broad spectrum of features for predicting human reading behavior that fall into five categories (syntactic complexity, lexical richness, register-based multiword combinations, readability and psycholinguistic word properties). Our experiments show that both the features included and the architecture of the transformer-based language models play a role in predicting multiple eye-tracking measures during naturalistic reading. We also report the results of experiments aimed at determining the relative importance of features from different groups using SP-LIME.

Neural Machine Translation (NMT) systems exhibit problematic biases, such as stereotypical gender bias in the translation of occupation terms into languages with grammatical gender. In this paper we describe a new source of bias prevalent in NMT systems, relating to translations of sentences containing person names. To correctly translate such sentences, a NMT system needs to determine the gender of the name. We show that leading systems are particularly poor at this task, especially for female given names. This bias is deeper than given name gender: we show that the translation of terms with ambiguous sentiment can also be affected by person names, and the same holds true for proper nouns denoting race. To mitigate these biases we propose a simple but effective data augmentation method based on randomly switching entities during translation, which effectively eliminates the problem without any effect on translation quality.

With the rapid growth in language processing applications, fairness has emerged as an important consideration in data-driven solutions. Although various fairness definitions have been explored in the recent literature, there is lack of consensus on which metrics most accurately reflect the fairness of a system. In this work, we propose a new formulation – accumulated prediction sensitivity, which measures fairness in machine learning models based on the model’s prediction sensitivity to perturbations in input features. The metric attempts to quantify the extent to which a single prediction depends on a protected attribute, where the protected attribute encodes the membership status of an individual in a protected group. We show that the metric can be theoretically linked with a specific notion of group fairness (statistical parity) and individual fairness. It also correlates well with humans’ perception of fairness. We conduct experiments on two text classification datasets – Jigsaw Toxicity, and Bias in Bios, and evaluate the correlations between metrics and manual annotations on whether the model produced a fair outcome. We observe that the proposed fairness metric based on prediction sensitivity is statistically significantly more correlated with human annotation than the existing counterfactual fairness metric.