Jackie Chi Kit Cheung

Kushal Arora, Layla El Asri, Hareesh Bahuleyan et al. · mila

Current language generation models suffer from issues such as repetition, incoherence, and hallucinations. An often-repeated hypothesis is that this brittleness of generation models is caused by the training and the generation procedure mismatch, also referred to as exposure bias. In this paper, we verify this hypothesis by analyzing exposure bias from an imitation learning perspective. We show that exposure bias leads to an accumulation of errors, analyze why perplexity fails to capture this accumulation, and empirically show that this accumulation results in poor generation quality. Source code to reproduce these experiments is available at https://github.com/kushalarora/quantifying_exposure_bias

Meng Cao, Mehdi Fatemi, Jackie Chi Kit Cheung et al. · microsoft-research

With adversarial or otherwise normal prompts, existing large language models (LLM) can be pushed to generate toxic discourses. One way to reduce the risk of LLMs generating undesired discourses is to alter the training of the LLM. This can be very restrictive due to demanding computation requirements. Other methods rely on rule-based or prompt-based token elimination, which are limited as they dismiss future tokens and the overall meaning of the complete discourse. Here, we center detoxification on the probability that the finished discourse is ultimately considered toxic. That is, at each point, we advise against token selections proportional to how likely a finished text from this point will be toxic. To this end, we formally extend the dead-end theory from the recent reinforcement learning (RL) literature to also cover uncertain outcomes. Our approach, called rectification, utilizes a separate but significantly smaller model for detoxification, which can be applied to diverse LLMs as long as they share the same vocabulary. Importantly, our method does not require access to the internal representations of the LLM, but only the token probability distribution at each decoding step. This is crucial as many LLMs today are hosted in servers and only accessible through APIs. When applied to various LLMs, including GPT-3, our approach significantly improves the generated discourse compared to the base LLMs and other techniques in terms of both the overall language and detoxification performance.

Yu Lu Liu, Meng Cao, Su Lin Blodgett et al. · microsoft-research

AI and NLP publication venues have increasingly encouraged researchers to reflect on possible ethical considerations, adverse impacts, and other responsible AI issues their work might engender. However, for specific NLP tasks our understanding of how prevalent such issues are, or when and why these issues are likely to arise, remains limited. Focusing on text summarization -- a common NLP task largely overlooked by the responsible AI community -- we examine research and reporting practices in the current literature. We conduct a multi-round qualitative analysis of 333 summarization papers from the ACL Anthology published between 2020-2022. We focus on how, which, and when responsible AI issues are covered, which relevant stakeholders are considered, and mismatches between stated and realized research goals. We also discuss current evaluation practices and consider how authors discuss the limitations of both prior work and their own work. Overall, we find that relatively few papers engage with possible stakeholders or contexts of use, which limits their consideration of potential downstream adverse impacts or other responsible AI issues. Based on our findings, we make recommendations on concrete practices and research directions.

Ian Porada, Alexandra Olteanu, Kaheer Suleman et al. · mila

It is increasingly common to evaluate the same coreference resolution (CR) model on multiple datasets. Do these multi-dataset evaluations allow us to draw meaningful conclusions about model generalization? Or, do they rather reflect the idiosyncrasies of a particular experimental setup (e.g., the specific datasets used)? To study this, we view evaluation through the lens of measurement modeling, a framework commonly used in the social sciences for analyzing the validity of measurements. By taking this perspective, we show how multi-dataset evaluations risk conflating different factors concerning what, precisely, is being measured. This in turn makes it difficult to draw more generalizable conclusions from these evaluations. For instance, we show that across seven datasets, measurements intended to reflect CR model generalization are often correlated with differences in both how coreference is defined and how it is operationalized; this limits our ability to draw conclusions regarding the ability of CR models to generalize across any singular dimension. We believe the measurement modeling framework provides the needed vocabulary for discussing challenges surrounding what is actually being measured by CR evaluations.

Meng Cao, Yue Dong, Jingyi He et al. · mila

State-of-the-art abstractive summarization systems frequently hallucinate content that is not supported by the source document, mainly due to noise in the training dataset. Existing methods opt to drop the noisy samples or tokens from the training set entirely, reducing the effective training set size and creating an artificial propensity to copy words from the source. In this work, we propose a training objective for abstractive summarization based on rejection learning, in which the model learns whether or not to reject potentially noisy tokens. We further propose a regularized decoding objective that penalizes non-factual candidate summaries during inference by using the rejection probability learned during training. We show that our method considerably improves the factuality of generated summaries in automatic and human evaluations when compared to five baseline models and that it does so while increasing the abstractiveness of the generated summaries.

Yu Lu Liu, Rachel Bawden, Thomas Scialom et al.

In text summarization and simplification, system outputs must be evaluated along multiple dimensions such as relevance, factual consistency, fluency, and grammaticality, and a wide range of possible outputs could be of high quality. These properties make the development of an adaptable, reference-less evaluation metric both necessary and challenging. We introduce MaskEval, a reference-less metric for text summarization and simplification that operates by performing masked language modeling (MLM) on the concatenation of the candidate and the source texts. It features an attention-like weighting mechanism to modulate the relative importance of each MLM step, which crucially allows it to be adapted to evaluate different quality dimensions. We demonstrate its effectiveness on English summarization and simplification in terms of correlations with human judgments, and explore transfer scenarios between the two tasks.

Akshatha Arodi, Martin Pömsl, Kaheer Suleman et al.

Many state-of-the-art natural language understanding (NLU) models are based on pretrained neural language models. These models often make inferences using information from multiple sources. An important class of such inferences are those that require both background knowledge, presumably contained in a model's pretrained parameters, and instance-specific information that is supplied at inference time. However, the integration and reasoning abilities of NLU models in the presence of multiple knowledge sources have been largely understudied. In this work, we propose a test suite of coreference resolution subtasks that require reasoning over multiple facts. These subtasks differ in terms of which knowledge sources contain the relevant facts. We also introduce subtasks where knowledge is present only at inference time using fictional knowledge. We evaluate state-of-the-art coreference resolution models on our dataset. Our results indicate that several models struggle to reason on-the-fly over knowledge observed both at pretrain time and at inference time. However, with task-specific training, a subset of models demonstrates the ability to integrate certain knowledge types from multiple sources. Still, even the best performing models seem to have difficulties with reliably integrating knowledge presented only at inference time.

Meng Cao, Mehdi Fatemi, Jackie Chi Kit Cheung et al.

While large language models (LLMs) have achieved impressive performance in generating fluent and realistic text, controlling the generated text so that it exhibits properties such as safety, factuality, and non-toxicity remains challenging. % such as DExperts, GeDi, and rectification Existing decoding-based methods are static in terms of the dimension of control; if the target subject is changed, they require new training. Moreover, it can quickly become prohibitive to concurrently control multiple subjects. In this work, we introduce SF-GEN, which is grounded in two primary concepts: successor features (SFs) to decouple the LLM's dynamics from task-specific rewards, and language model rectification to proportionally adjust the probability of selecting a token based on the likelihood that the finished text becomes undesired. SF-GEN seamlessly integrates the two to enable dynamic steering of text generation with no need to alter the LLM's parameters. Thanks to the decoupling effect induced by successor features, our method proves to be memory-wise and computationally efficient for training as well as decoding, especially when dealing with multiple target subjects. To the best of our knowledge, our research represents the first application of successor features in text generation. In addition to its computational efficiency, the resultant language produced by our method is comparable to the SOTA (and outperforms baselines) in both control measures as well as language quality, which we demonstrate through a series of experiments in various controllable text generation tasks.

Jie Gao, Yongan Yu, Junzhu Su et al.

Adaptive learning refers to educational technologies that track learners' learning progress and adapt the instructional process based on individual learners' learning performance. It is increasingly recognized as critical for developing an effective learning support tool. Vision language models (VLMs) have seen adoption in mathematics education, and students have been using them as learning aids for personalized instruction. However, it is unknown whether VLMs have the ability to adapt to different learner profiles when providing mathematical instructions. Current VLMs lack a systematic evaluation framework for this adaptivity to different learner profiles in mathematics tutoring tasks. To address this gap, we draw on the learner model from the adaptive learning framework (Shute and Towle, 2018) and propose a learner model-based rubric. Our rubric formalizes adaptivity assessment into three aspects: cognitive aspects, motivational aspects, and complexity. We also evaluate two additional dimensions of VLM responses: correctness (of answers and solutions) and quality (of the response itself). Our experimental results show measurable differences in adaptivity across models and also reveal that current VLMs struggle to consistently produce learner model-based instructional responses, especially when receiving limited learner information.

Lorenzo Jaime Yu Flores, Cesare Spinoso di-Piano, Ori Ernst et al.

Active learning (AL) is a training paradigm for selecting unlabeled samples for annotation to improve model performance on a test set, which is useful when only a limited number of samples can be annotated. These algorithms often work by optimizing for the informativeness and diversity of the training data to be annotated. Recent work found that AL strategies fail to outperform random sampling on various language generation tasks when using 100-500 samples. To understand AL's poor performance when only using few samples, we investigate whether the core assumptions underlying AL strategies hold. We find that neither the informativeness nor diversity of the training data, which AL strategies optimize for, are correlated with test set performance. Instead, factors like the ordering of the training samples and interactions with pre-training data have a larger impact on performance. This suggests that future AL methods must take these factors into account in order to work with very few samples.

Lorenzo Jaime Yu Flores, Cesare Spinoso di-Piano, Jackie Chi Kit Cheung

Uncertainty quantification is a set of techniques that measure confidence in language models. They can be used, for example, to detect hallucinations or alert users to review uncertain predictions. To be useful, these confidence scores must be correlated with the quality of the output. However, recent work found that fine-tuning can affect the correlation between confidence scores and quality. Hence, we investigate the underlying behavior of confidence scores to understand its sensitivity to supervised fine-tuning (SFT). We find that post-SFT, the correlation of various confidence scores degrades, which can stem from changes in confidence scores due to factors other than the output quality, such as the output's similarity to the training distribution. We demonstrate via a case study how failing to address this miscorrelation reduces the usefulness of the confidence scores on a downstream task. Our findings show how confidence metrics cannot be used off-the-shelf without testing, and motivate the need for developing metrics which are more robust to fine-tuning.

Yu Bai, Xiyuan Zou, Heyan Huang et al.

Long sequence modeling has gained broad interest as large language models (LLMs) continue to advance. Recent research has identified that a large portion of hidden states within the key-value caches of Transformer models can be discarded (also termed evicted) without affecting the perplexity performance in generating long sequences. However, we show that these methods, despite preserving perplexity performance, often drop information that is important for solving downstream tasks, a problem which we call information neglect. To address this issue, we introduce Chunked Instruction-aware State Eviction (CItruS), a novel modeling technique that integrates the attention preferences useful for a downstream task into the eviction process of hidden states. In addition, we design a method for chunked sequence processing to further improve efficiency. Our training-free method exhibits superior performance on long sequence comprehension and retrieval tasks over several strong baselines under the same memory budget, while preserving language modeling perplexity. The code and data have been released at https://github.com/ybai-nlp/CItruS.

Lei Yu, Meng Cao, Jackie Chi Kit Cheung et al.

State-of-the-art language models (LMs) sometimes generate non-factual hallucinations that misalign with world knowledge. To explore the mechanistic causes of these hallucinations, we create diagnostic datasets with subject-relation queries and adapt interpretability methods to trace hallucinations through internal model representations. We discover two general and distinct mechanistic causes of hallucinations shared across LMs (Llama-2, Pythia, GPT-J): 1) knowledge enrichment hallucinations: insufficient subject attribute knowledge in lower layer MLPs, and 2) answer extraction hallucinations: failure to select the correct object attribute in upper layer attention heads. We also found these two internal mechanistic causes of hallucinations are reflected in external manifestations. Based on insights from our mechanistic analysis, we propose a novel hallucination mitigation method through targeted restoration of the LM's internal fact recall pipeline, demonstrating superior performance compared to baselines.

Khaoula Chehbouni, Mohammed Haddou, Jackie Chi Kit Cheung et al.

Evaluating natural language generation (NLG) systems remains a core challenge of natural language processing (NLP), further complicated by the rise of large language models (LLMs) that aims to be general-purpose. Recently, large language models as judges (LLJs) have emerged as a promising alternative to traditional metrics, but their validity remains underexplored. This position paper argues that the current enthusiasm around LLJs may be premature, as their adoption has outpaced rigorous scrutiny of their reliability and validity as evaluators. Drawing on measurement theory from the social sciences, we identify and critically assess four core assumptions underlying the use of LLJs: their ability to act as proxies for human judgment, their capabilities as evaluators, their scalability, and their cost-effectiveness. We examine how each of these assumptions may be challenged by the inherent limitations of LLMs, LLJs, or current practices in NLG evaluation. To ground our analysis, we explore three applications of LLJs: text summarization, data annotation, and safety alignment. Finally, we highlight the need for more responsible evaluation practices in LLJs evaluation, to ensure that their growing role in the field supports, rather than undermines, progress in NLG.

Ian Porada, Xiyuan Zou, Jackie Chi Kit Cheung · mila

All state-of-the-art coreference resolution (CR) models involve finetuning a pretrained language model. Whether the superior performance of one CR model over another is due to the choice of language model or other factors, such as the task-specific architecture, is difficult or impossible to determine due to lack of a standardized experimental setup. To resolve this ambiguity, we systematically evaluate five CR models and control for certain design decisions including the pretrained language model used by each. When controlling for language model size, encoder-based CR models outperform more recent decoder-based models in terms of both accuracy and inference speed. Surprisingly, among encoder-based CR models, more recent models are not always more accurate, and the oldest CR model that we test generalizes the best to out-of-domain textual genres. We conclude that controlling for the choice of language model reduces most, but not all, of the increase in F1 score reported in the past five years.

Maxime Darrin, Philippe Formont, Jackie Chi Kit Cheung et al.

Assessing the quality of summarizers poses significant challenges. In response, we propose a novel task-oriented evaluation approach that assesses summarizers based on their capacity to produce summaries that are useful for downstream tasks, while preserving task outcomes. We theoretically establish a direct relationship between the resulting error probability of these tasks and the mutual information between source texts and generated summaries. We introduce $\texttt{COSMIC}$ as a practical implementation of this metric, demonstrating its strong correlation with human judgment-based metrics and its effectiveness in predicting downstream task performance. Comparative analyses against established metrics like $\texttt{BERTScore}$ and $\texttt{ROUGE}$ highlight the competitive performance of $\texttt{COSMIC}$.

Ziling Cheng, Meng Cao, Leila Pishdad et al.

Final-answer-based metrics are commonly used for evaluating large language models (LLMs) on math word problems, often taken as proxies for reasoning ability. However, such metrics conflate two distinct sub-skills: abstract formulation (capturing mathematical relationships using expressions) and arithmetic computation (executing the calculations). Through a disentangled evaluation on GSM8K and SVAMP, we find that the final-answer accuracy of Llama-3 and Qwen2.5 (1B-32B) without CoT is overwhelmingly bottlenecked by the arithmetic computation step and not by the abstract formulation step. Contrary to the common belief, we show that CoT primarily aids in computation, with limited impact on abstract formulation. Mechanistically, we show that these two skills are composed conjunctively even in a single forward pass without any reasoning steps via an abstract-then-compute mechanism: models first capture problem abstractions, then handle computation. Causal patching confirms these abstractions are present, transferable, composable, and precede computation. These behavioural and mechanistic findings highlight the need for disentangled evaluation to accurately assess LLM reasoning and to guide future improvements.

Behzad Shayegh, Hobie H. -B. Lee, Xiaodan Zhu et al.

We address unsupervised dependency parsing by building an ensemble of diverse existing models through post hoc aggregation of their output dependency parse structures. We observe that these ensembles often suffer from low robustness against weak ensemble components due to error accumulation. To tackle this problem, we propose an efficient ensemble-selection approach that considers error diversity and avoids error accumulation. Results demonstrate that our approach outperforms each individual model as well as previous ensemble techniques. Additionally, our experiments show that the proposed ensemble-selection method significantly enhances the performance and robustness of our ensemble, surpassing previously proposed strategies, which have not accounted for error diversity.

Ziling Cheng, Meng Cao, Marc-Antoine Rondeau et al.

The widespread success of large language models (LLMs) on NLP benchmarks has been accompanied by concerns that LLMs function primarily as stochastic parrots that reproduce texts similar to what they saw during pre-training, often erroneously. But what is the nature of their errors, and do these errors exhibit any regularities? In this work, we examine irrelevant context hallucinations, in which models integrate misleading contextual cues into their predictions. Through behavioral analysis, we show that these errors result from a structured yet flawed mechanism that we term class-based (mis)generalization, in which models combine abstract class cues with features extracted from the query or context to derive answers. Furthermore, mechanistic interpretability experiments on Llama-3, Mistral, and Pythia across 39 factual recall relation types reveal that this behavior is reflected in the model's internal computations: (i) abstract class representations are constructed in lower layers before being refined into specific answers in higher layers, (ii) feature selection is governed by two competing circuits -- one prioritizing direct query-based reasoning, the other incorporating contextual cues -- whose relative influences determine the final output. Our findings provide a more nuanced perspective on the stochastic parrot argument: through form-based training, LLMs can exhibit generalization leveraging abstractions, albeit in unreliable ways based on contextual cues -- what we term stochastic chameleons.

Ian Porada, Jackie Chi Kit Cheung · mila

Challenge sets such as the Winograd Schema Challenge (WSC) are used to benchmark systems' ability to resolve ambiguities in natural language. If one assumes as in existing work that solving a given challenge set is at least as difficult as solving some more general task, then high performance on the challenge set should indicate high performance on the general task overall. However, we show empirically that this assumption of difficulty does not always hold. In particular, we demonstrate that despite the strong performance of prompted language models (LMs) on the WSC and its variants, these same modeling techniques perform relatively poorly at resolving certain pronominal ambiguities attested in OntoNotes and related datasets that are perceived to be easier. Motivated by these findings, we propose a method for ensembling a prompted LM with a supervised, task-specific system that is overall more accurate at resolving pronominal coreference across datasets. Finally, we emphasize that datasets involving the same linguistic phenomenon draw on distinct, but overlapping, capabilities, and evaluating on any one dataset alone does not provide a complete picture of a system's overall capability.

Cesare Spinoso-Di Piano, David Austin, Pablo Piantanida et al.

Figurative language (e.g., irony, hyperbole, understatement) is ubiquitous in human communication, resulting in utterances where the literal and the intended meanings do not match. The Rational Speech Act (RSA) framework, which explicitly models speaker intentions, is the most widespread theory of probabilistic pragmatics, but existing implementations are either unable to account for figurative expressions or require modeling the implicit motivations for using figurative language (e.g., to express joy or annoyance) in a setting-specific way. In this paper, we introduce the Rhetorical-Strategy-Aware RSA $(RSA)^2$ framework which models figurative language use by considering a speaker's employed rhetorical strategy. We show that $(RSA)^2$ enables human-compatible interpretations of non-literal utterances without modeling a speaker's motivations for being non-literal. Combined with LLMs, it achieves state-of-the-art performance on the ironic split of PragMega+, a new irony interpretation dataset introduced in this study.

Lorenzo Jaime Yu Flores, Ori Ernst, Jackie Chi Kit Cheung

Well-calibrated model confidence scores can improve the usefulness of text generation models. For example, users can be prompted to review predictions with low confidence scores, to prevent models from returning bad or potentially dangerous predictions. However, confidence metrics are not always well calibrated in text generation. One reason is that in generation, there can be many valid answers, which previous methods do not always account for. Hence, a confident model could distribute its output probability among multiple sequences because they are all valid. We propose task-agnostic confidence metrics suited to generation, which rely solely on the probabilities associated with the model outputs without the need for further fine-tuning or heuristics. Using these, we are able to improve the calibration of BART and Flan-T5 on summarization, translation, and QA datasets.

Steven Koniaev, Ori Ernst, Jackie Chi Kit Cheung

Despite recent advancements in automatic summarization, state-of-the-art models do not summarize all documents equally well, raising the question: why? While prior research has extensively analyzed summarization models, little attention has been given to the role of document characteristics in influencing summarization performance. In this work, we explore two key research questions. First, do documents exhibit consistent summarization quality across multiple systems? If so, can we predict a document's summarization performance without generating a summary? We answer both questions affirmatively and introduce PreSumm, a novel task in which a system predicts summarization performance based solely on the source document. Our analysis sheds light on common properties of documents with low PreSumm scores, revealing that they often suffer from coherence issues, complex content, or a lack of a clear main theme. In addition, we demonstrate PreSumm's practical utility in two key applications: improving hybrid summarization workflows by identifying documents that require manual summarization and enhancing dataset quality by filtering outliers and noisy documents. Overall, our findings highlight the critical role of document properties in summarization performance and offer insights into the limitations of current systems that could serve as the basis for future improvements.

Cesare Spinoso-Di Piano, Jackie Chi Kit Cheung

Query-focused summarization (QFS) is the task of generating a summary in response to a user-written query. Despite its user-oriented nature, there has been limited work in QFS in explicitly considering a user's understanding of a generated summary, potentially causing QFS systems to underperform at inference time. In this paper, we adapt the Rational Speech Act (RSA) framework, a model of human communication, to explicitly model a reader's understanding of a query-focused summary and integrate it within the generation method of existing QFS systems. In particular, we introduce the answer reconstruction objective which approximates a reader's understanding of a summary by their ability to use it to reconstruct the answer to their initial query. Using this objective, we are able to re-rank candidate summaries generated by existing QFS systems and select summaries that better align with their corresponding query and reference summary. More generally, our study suggests that a simple and effective way of improving a language generation system designed for a user-centered task may be to explicitly incorporate its user requirements into the system's generation procedure.

Yu Lu Liu, Su Lin Blodgett, Jackie Chi Kit Cheung et al.

Benchmarking is seen as critical to assessing progress in NLP. However, creating a benchmark involves many design decisions (e.g., which datasets to include, which metrics to use) that often rely on tacit, untested assumptions about what the benchmark is intended to measure or is actually measuring. There is currently no principled way of analyzing these decisions and how they impact the validity of the benchmark's measurements. To address this gap, we draw on evidence-centered design in educational assessments and propose Evidence-Centered Benchmark Design (ECBD), a framework which formalizes the benchmark design process into five modules. ECBD specifies the role each module plays in helping practitioners collect evidence about capabilities of interest. Specifically, each module requires benchmark designers to describe, justify, and support benchmark design choices -- e.g., clearly specifying the capabilities the benchmark aims to measure or how evidence about those capabilities is collected from model responses. To demonstrate the use of ECBD, we conduct case studies with three benchmarks: BoolQ, SuperGLUE, and HELM. Our analysis reveals common trends in benchmark design and documentation that could threaten the validity of benchmarks' measurements.

Yu Bai, Heyan Huang, Cesare Spinoso-Di Piano et al.

In-context learning (ICL) has emerged as an effective solution for few-shot learning with large language models (LLMs). However, how LLMs leverage demonstrations to specify a task and learn a corresponding computational function through ICL is underexplored. Drawing from the way humans learn from content-label mappings in demonstrations, we categorize the tokens in an ICL prompt into content, stopword, and template tokens. Our goal is to identify the types of tokens whose representations directly influence LLM's performance, a property we refer to as being performance-critical. By ablating representations from the attention of the test example, we find that the representations of informative content tokens have less influence on performance compared to template and stopword tokens, which contrasts with the human attention to informative words. We give evidence that the representations of performance-critical tokens aggregate information from the content tokens. Moreover, we demonstrate experimentally that lexical meaning, repetition, and structural cues are the main distinguishing characteristics of these tokens. Our work sheds light on how large language models learn to perform tasks from demonstrations and deepens our understanding of the roles different types of tokens play in large language models.

Rahul Aralikatte, Ziling Cheng, Sumanth Doddapaneni et al.

We present Vārta, a large-scale multilingual dataset for headline generation in Indic languages. This dataset includes 41.8 million news articles in 14 different Indic languages (and English), which come from a variety of high-quality sources. To the best of our knowledge, this is the largest collection of curated articles for Indic languages currently available. We use the data collected in a series of experiments to answer important questions related to Indic NLP and multilinguality research in general. We show that the dataset is challenging even for state-of-the-art abstractive models and that they perform only slightly better than extractive baselines. Owing to its size, we also show that the dataset can be used to pretrain strong language models that outperform competitive baselines in both NLU and NLG benchmarks.

Ian Porada, Alessandro Sordoni, Jackie Chi Kit Cheung

Transformer models pre-trained with a masked-language-modeling objective (e.g., BERT) encode commonsense knowledge as evidenced by behavioral probes; however, the extent to which this knowledge is acquired by systematic inference over the semantics of the pre-training corpora is an open question. To answer this question, we selectively inject verbalized knowledge into the minibatches of a BERT model during pre-training and evaluate how well the model generalizes to supported inferences. We find generalization does not improve over the course of pre-training, suggesting that commonsense knowledge is acquired from surface-level, co-occurrence patterns rather than induced, systematic reasoning.

Meng Cao, Yue Dong, Jackie Chi Kit Cheung

State-of-the-art abstractive summarization systems often generate \emph{hallucinations}; i.e., content that is not directly inferable from the source text. Despite being assumed incorrect, we find that much hallucinated content is factual, namely consistent with world knowledge. These factual hallucinations can be beneficial in a summary by providing useful background information. In this work, we propose a novel detection approach that separates factual from non-factual hallucinations of entities. Our method utilizes an entity's prior and posterior probabilities according to pre-trained and finetuned masked language models, respectively. Empirical results suggest that our approach vastly outperforms two baselines %in both accuracy and F1 scores and strongly correlates with human judgments. % on factuality classification tasks. Furthermore, we show that our detector, when used as a reward signal in an off-line reinforcement learning (RL) algorithm, significantly improves the factuality of summaries while maintaining the level of abstractiveness.

Ian Porada, Kaheer Suleman, Adam Trischler et al.

Understanding natural language requires common sense, one aspect of which is the ability to discern the plausibility of events. While distributional models -- most recently pre-trained, Transformer language models -- have demonstrated improvements in modeling event plausibility, their performance still falls short of humans'. In this work, we show that Transformer-based plausibility models are markedly inconsistent across the conceptual classes of a lexical hierarchy, inferring that "a person breathing" is plausible while "a dentist breathing" is not, for example. We find this inconsistency persists even when models are softly injected with lexical knowledge, and we present a simple post-hoc method of forcing model consistency that improves correlation with human plausibility judgements.

Michaela Socolof, Jackie Chi Kit Cheung, Michael Wagner et al.

Idioms are unlike most phrases in two important ways. First, the words in an idiom have non-canonical meanings. Second, the non-canonical meanings of words in an idiom are contingent on the presence of other words in the idiom. Linguistic theories differ on whether these properties depend on one another, as well as whether special theoretical machinery is needed to accommodate idioms. We define two measures that correspond to the properties above, and we implement them using BERT (Devlin et al., 2019) and XLNet(Yang et al., 2019). We show that idioms fall at the expected intersection of the two dimensions, but that the dimensions themselves are not correlated. Our results suggest that special machinery to handle idioms may not be warranted.

Malik H. Altakrori, Jackie Chi Kit Cheung, Benjamin C. M. Fung

Authorship attribution is the problem of identifying the most plausible author of an anonymous text from a set of candidate authors. Researchers have investigated same-topic and cross-topic scenarios of authorship attribution, which differ according to whether new, unseen topics are used in the testing phase. However, neither scenario allows us to explain whether errors are caused by a failure to capture authorship writing style or by a topic shift. Motivated by this, we propose the \emph{topic confusion} task where we switch the author-topic configuration between the training and testing sets. This setup allows us to distinguish two types of errors: those caused by the topic shift and those caused by the features' inability to capture the writing styles. We show that stylometric features with part-of-speech tags are the least susceptible to topic variations. We further show that combining them with other features leads to significantly lower topic confusion and higher attribution accuracy. Finally, we show that pretrained language models such as BERT and RoBERTa perform poorly on this task and are surpassed by simple features such as word-level $n$-grams.

Jiapeng Wu, Yishi Xu, Yingxue Zhang et al.

Reasoning in a temporal knowledge graph (TKG) is a critical task for information retrieval and semantic search. It is particularly challenging when the TKG is updated frequently. The model has to adapt to changes in the TKG for efficient training and inference while preserving its performance on historical knowledge. Recent work approaches TKG completion (TKGC) by augmenting the encoder-decoder framework with a time-aware encoding function. However, naively fine-tuning the model at every time step using these methods does not address the problems of 1) catastrophic forgetting, 2) the model's inability to identify the change of facts (e.g., the change of the political affiliation and end of a marriage), and 3) the lack of training efficiency. To address these challenges, we present the Time-aware Incremental Embedding (TIE) framework, which combines TKG representation learning, experience replay, and temporal regularization. We introduce a set of metrics that characterizes the intransigence of the model and propose a constraint that associates the deleted facts with negative labels. Experimental results on Wikidata12k and YAGO11k datasets demonstrate that the proposed TIE framework reduces training time by about ten times and improves on the proposed metrics compared to vanilla full-batch training. It comes without a significant loss in performance for any traditional measures. Extensive ablation studies reveal performance trade-offs among different evaluation metrics, which is essential for decision-making around real-world TKG applications.

Yue Dong, Chandra Bhagavatula, Ximing Lu et al.

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.

Peng Xu, Dhruv Kumar, Wei Yang et al.

It is a common belief that training deep transformers from scratch requires large datasets. Consequently, for small datasets, people usually use shallow and simple additional layers on top of pre-trained models during fine-tuning. This work shows that this does not always need to be the case: with proper initialization and optimization, the benefits of very deep transformers can carry over to challenging tasks with small datasets, including Text-to-SQL semantic parsing and logical reading comprehension. In particular, we successfully train $48$ layers of transformers, comprising $24$ fine-tuned layers from pre-trained RoBERTa and $24$ relation-aware layers trained from scratch. With fewer training steps and no task-specific pre-training, we obtain the state-of-the-art performance on the challenging cross-domain Text-to-SQL parsing benchmark Spider. We achieve this by deriving a novel Data-dependent Transformer Fixed-update initialization scheme (DT-Fixup), inspired by the prior T-Fixup work. Further error analysis shows that increasing depth can help improve generalization on small datasets for hard cases that require reasoning and structural understanding.

Kian Kenyon-Dean, Edward Newell, Jackie Chi Kit Cheung

Word embeddings are reliable feature representations of words used to obtain high quality results for various NLP applications. Uncontextualized word embeddings are used in many NLP tasks today, especially in resource-limited settings where high memory capacity and GPUs are not available. Given the historical success of word embeddings in NLP, we propose a retrospective on some of the most well-known word embedding algorithms. In this work, we deconstruct Word2vec, GloVe, and others, into a common form, unveiling some of the common conditions that seem to be required for making performant word embeddings. We believe that the theoretical findings in this paper can provide a basis for more informed development of future models.

Ali Emami, Adam Trischler, Kaheer Suleman et al.

The Winograd Schema Challenge (WSC) and variants inspired by it have become important benchmarks for common-sense reasoning (CSR). Model performance on the WSC has quickly progressed from chance-level to near-human using neural language models trained on massive corpora. In this paper, we analyze the effects of varying degrees of overlap between these training corpora and the test instances in WSC-style tasks. We find that a large number of test instances overlap considerably with the corpora on which state-of-the-art models are (pre)trained, and that a significant drop in classification accuracy occurs when we evaluate models on instances with minimal overlap. Based on these results, we develop the KnowRef-60K dataset, which consists of over 60k pronoun disambiguation problems scraped from web data. KnowRef-60K is the largest corpus to date for WSC-style common-sense reasoning and exhibits a significantly lower proportion of overlaps with current pretraining corpora.

Jingyi He, KC Tsiolis, Kian Kenyon-Dean et al.

Word embeddings are trained to predict word cooccurrence statistics, which leads them to possess different lexical properties (syntactic, semantic, etc.) depending on the notion of context defined at training time. These properties manifest when querying the embedding space for the most similar vectors, and when used at the input layer of deep neural networks trained to solve downstream NLP problems. Meta-embeddings combine multiple sets of differently trained word embeddings, and have been shown to successfully improve intrinsic and extrinsic performance over equivalent models which use just one set of source embeddings. We introduce word prisms: a simple and efficient meta-embedding method that learns to combine source embeddings according to the task at hand. Word prisms learn orthogonal transformations to linearly combine the input source embeddings, which allows them to be very efficient at inference time. We evaluate word prisms in comparison to other meta-embedding methods on six extrinsic evaluations and observe that word prisms offer improvements in performance on all tasks.

Meng Cao, Yue Dong, Jiapeng Wu et al.

Neural abstractive summarization systems have achieved promising progress, thanks to the availability of large-scale datasets and models pre-trained with self-supervised methods. However, ensuring the factual consistency of the generated summaries for abstractive summarization systems is a challenge. We propose a post-editing corrector module to address this issue by identifying and correcting factual errors in generated summaries. The neural corrector model is pre-trained on artificial examples that are created by applying a series of heuristic transformations on reference summaries. These transformations are inspired by an error analysis of state-of-the-art summarization model outputs. Experimental results show that our model is able to correct factual errors in summaries generated by other neural summarization models and outperforms previous models on factual consistency evaluation on the CNN/DailyMail dataset. We also find that transferring from artificial error correction to downstream settings is still very challenging.

Jiapeng Wu, Meng Cao, Jackie Chi Kit Cheung et al.

Inferring missing facts in temporal knowledge graphs (TKGs) is a fundamental and challenging task. Previous works have approached this problem by augmenting methods for static knowledge graphs to leverage time-dependent representations. However, these methods do not explicitly leverage multi-hop structural information and temporal facts from recent time steps to enhance their predictions. Additionally, prior work does not explicitly address the temporal sparsity and variability of entity distributions in TKGs. We propose the Temporal Message Passing (TeMP) framework to address these challenges by combining graph neural networks, temporal dynamics models, data imputation and frequency-based gating techniques. Experiments on standard TKG tasks show that our approach provides substantial gains compared to the previous state of the art, achieving a 10.7% average relative improvement in Hits@10 across three standard benchmarks. Our analysis also reveals important sources of variability both within and across TKG datasets, and we introduce several simple but strong baselines that outperform the prior state of the art in certain settings.

Yue Dong, Shuohang Wang, Zhe Gan et al.

Pre-trained neural abstractive summarization systems have dominated extractive strategies on news summarization performance, at least in terms of ROUGE. However, system-generated abstractive summaries often face the pitfall of factual inconsistency: generating incorrect facts with respect to the source text. To address this challenge, we propose Span-Fact, a suite of two factual correction models that leverages knowledge learned from question answering models to make corrections in system-generated summaries via span selection. Our models employ single or multi-masking strategies to either iteratively or auto-regressively replace entities in order to ensure semantic consistency w.r.t. the source text, while retaining the syntactic structure of summaries generated by abstractive summarization models. Experiments show that our models significantly boost the factual consistency of system-generated summaries without sacrificing summary quality in terms of both automatic metrics and human evaluation.

Edward Newell, Kian Kenyon-Dean, Jackie Chi Kit Cheung

Uncontextualized word embeddings are reliable feature representations of words used to obtain high quality results for various NLP applications. Given the historical success of word embeddings in NLP, we propose a retrospective on some of the most well-known word embedding algorithms. In this work, we deconstruct Word2vec, GloVe, and others, into a common form, unveiling some of the necessary and sufficient conditions required for making performant word embeddings. We find that each algorithm: (1) fits vector-covector dot products to approximate pointwise mutual information (PMI); and, (2) modulates the loss gradient to balance weak and strong signals. We demonstrate that these two algorithmic features are sufficient conditions to construct a novel word embedding algorithm, Hilbert-MLE. We find that its embeddings obtain equivalent or better performance against other algorithms across 17 intrinsic and extrinsic datasets.

Ian Porada, Kaheer Suleman, Jackie Chi Kit Cheung

Modeling semantic plausibility requires commonsense knowledge about the world and has been used as a testbed for exploring various knowledge representations. Previous work has focused specifically on modeling physical plausibility and shown that distributional methods fail when tested in a supervised setting. At the same time, distributional models, namely large pretrained language models, have led to improved results for many natural language understanding tasks. In this work, we show that these pretrained language models are in fact effective at modeling physical plausibility in the supervised setting. We therefore present the more difficult problem of learning to model physical plausibility directly from text. We create a training set by extracting attested events from a large corpus, and we provide a baseline for training on these attested events in a self-supervised manner and testing on a physical plausibility task. We believe results could be further improved by injecting explicit commonsense knowledge into a distributional model.

Teng Long, Yanshuai Cao, Jackie Chi Kit Cheung

Variational autoencoders (VAEs) hold great potential for modelling text, as they could in theory separate high-level semantic and syntactic properties from local regularities of natural language. Practically, however, VAEs with autoregressive decoders often suffer from posterior collapse, a phenomenon where the model learns to ignore the latent variables, causing the sequence VAE to degenerate into a language model. In this paper, we argue that posterior collapse is in part caused by the lack of dispersion in encoder features. We provide empirical evidence to verify this hypothesis, and propose a straightforward fix using pooling. This simple technique effectively prevents posterior collapse, allowing model to achieve significantly better data log-likelihood than standard sequence VAEs. Comparing to existing work, our proposed method is able to achieve comparable or superior performances while being more computationally efficient.

Matt Grenander, Yue Dong, Jackie Chi Kit Cheung et al.

Sentence position is a strong feature for news summarization, since the lead often (but not always) summarizes the key points of the article. In this paper, we show that recent neural systems excessively exploit this trend, which although powerful for many inputs, is also detrimental when summarizing documents where important content should be extracted from later parts of the article. We propose two techniques to make systems sensitive to the importance of content in different parts of the article. The first technique employs 'unbiased' data; i.e., randomly shuffled sentences of the source document, to pretrain the model. The second technique uses an auxiliary ROUGE-based loss that encourages the model to distribute importance scores throughout a document by mimicking sentence-level ROUGE scores on the training data. We show that these techniques significantly improve the performance of a competitive reinforcement learning based extractive system, with the auxiliary loss being more powerful than pretraining.

Meng Cao, Jackie Chi Kit Cheung

Referring Expression Generation (REG) is the task of generating contextually appropriate references to entities. A limitation of existing REG systems is that they rely on entity-specific supervised training, which means that they cannot handle entities not seen during training. In this study, we address this in two ways. First, we propose task setups in which we specifically test a REG system's ability to generalize to entities not seen during training. Second, we propose a profile-based deep neural network model, ProfileREG, which encodes both the local context and an external profile of the entity to generate reference realizations. Our model generates tokens by learning to choose between generating pronouns, generating from a fixed vocabulary, or copying a word from the profile. We evaluate our model on three different splits of the WebNLG dataset, and show that it outperforms competitive baselines in all settings according to automatic and human evaluations.

Yue Dong, Zichao Li, Mehdi Rezagholizadeh et al.

We present the first sentence simplification model that learns explicit edit operations (ADD, DELETE, and KEEP) via a neural programmer-interpreter approach. Most current neural sentence simplification systems are variants of sequence-to-sequence models adopted from machine translation. These methods learn to simplify sentences as a byproduct of the fact that they are trained on complex-simple sentence pairs. By contrast, our neural programmer-interpreter is directly trained to predict explicit edit operations on targeted parts of the input sentence, resembling the way that humans might perform simplification and revision. Our model outperforms previous state-of-the-art neural sentence simplification models (without external knowledge) by large margins on three benchmark text simplification corpora in terms of SARI (+0.95 WikiLarge, +1.89 WikiSmall, +1.41 Newsela), and is judged by humans to produce overall better and simpler output sentences.

Peng Xu, Jackie Chi Kit Cheung, Yanshuai Cao

The variational autoencoder (VAE) can learn the manifold of natural images on certain datasets, as evidenced by meaningful interpolating or extrapolating in the continuous latent space. However, on discrete data such as text, it is unclear if unsupervised learning can discover similar latent space that allows controllable manipulation. In this work, we find that sequence VAEs trained on text fail to properly decode when the latent codes are manipulated, because the modified codes often land in holes or vacant regions in the aggregated posterior latent space, where the decoding network fails to generalize. Both as a validation of the explanation and as a fix to the problem, we propose to constrain the posterior mean to a learned probability simplex, and performs manipulation within this simplex. Our proposed method mitigates the latent vacancy problem and achieves the first success in unsupervised learning of controllable representations for text. Empirically, our method outperforms unsupervised baselines and strong supervised approaches on text style transfer, and is capable of performing more flexible fine-grained control over text generation than existing methods.

Peng Xu, Hamidreza Saghir, Jin Sung Kang et al.

Coherence is an important aspect of text quality and is crucial for ensuring its readability. One important limitation of existing coherence models is that training on one domain does not easily generalize to unseen categories of text. Previous work advocates for generative models for cross-domain generalization, because for discriminative models, the space of incoherent sentence orderings to discriminate against during training is prohibitively large. In this work, we propose a local discriminative neural model with a much smaller negative sampling space that can efficiently learn against incorrect orderings. The proposed coherence model is simple in structure, yet it significantly outperforms previous state-of-art methods on a standard benchmark dataset on the Wall Street Journal corpus, as well as in multiple new challenging settings of transfer to unseen categories of discourse on Wikipedia articles.

Kian Kenyon-Dean, Andre Cianflone, Lucas Page-Caccia et al.

The standard loss function used to train neural network classifiers, categorical cross-entropy (CCE), seeks to maximize accuracy on the training data; building useful representations is not a necessary byproduct of this objective. In this work, we propose clustering-oriented representation learning (COREL) as an alternative to CCE in the context of a generalized attractive-repulsive loss framework. COREL has the consequence of building latent representations that collectively exhibit the quality of natural clustering within the latent space of the final hidden layer, according to a predefined similarity function. Despite being simple to implement, COREL variants outperform or perform equivalently to CCE in a variety of scenarios, including image and news article classification using both feed-forward and convolutional neural networks. Analysis of the latent spaces created with different similarity functions facilitates insights on the different use cases COREL variants can satisfy, where the Cosine-COREL variant makes a consistently clusterable latent space, while Gaussian-COREL consistently obtains better classification accuracy than CCE.