EMNLP 2022 Papers

Recent progress regarding the use of language models (LMs) as knowledge bases (KBs) has shown that language models can act as structured knowledge bases for storing relational facts. However, most existing works only considered the LM-as-KB paradigm in a static setting, which ignores the analysis of temporal dynamics of world knowledge. Furthermore, a basic function of KBs, i.e., the ability to store conflicting information (i.e., 1-N, N-1, and N-M relations), is underexplored. In this paper, we formulate two practical requirements for treating LMs as temporal KBs: (i) The capacity to store temporally-scoped knowledge that contains conflicting information and (ii) the ability to use stored knowledge for temporally-scoped knowledge queries. We introduce a new dataset called LAMA-TK which is aimed at probing temporally-scoped knowledge, and investigate the two above requirements to explore the LM-as-KB paradigm in the temporal domain. On the one hand, experiments show that LMs can memorize millions of temporally-scoped facts with relatively high accuracy and transfer stored knowledge to temporal knowledge queries, thereby expanding the LM-as-KB paradigm to the temporal domain. On the other hand, we show that memorizing conflicting information, which has been neglected by previous works, is still challenging for LMs and hinders the memorization of other unrelated one-to-one relationships.

Abstract Meaning Representation (AMR) has been shown to be useful for many downstream tasks. In this work, we explore the use of AMR for legal and logical reasoning. Specifically, we investigate if AMR can help capture logical relationships on multiple choice question answering (MCQA) tasks. We propose neural architectures that utilize linearised AMR graphs in combination with pre-trained language models. While these models are not able to outperform text-only baselines, they correctly solve different instances than the text models, suggesting complementary abilities. Error analysis further reveals that AMR parsing quality is the most prominent challenge, especially regarding inputs with multiple sentences. We conduct a theoretical analysis of how logical relations are represented in AMR and conclude it might be helpful in some logical statements but not for others.

Calibration strengthens the trustworthiness of black-box models by producing better accurate confidence estimates on given examples. However, little is known about if model explanations can help confidence calibration. Intuitively, humans look at important features attributions and decide whether the model is trustworthy. Similarly, the explanations may tell us when the model might know and when it does not. Inspired by this, we propose a method named CME that leverages model explanations to make the model less confident with non-inductive attributions. The idea is that when the model is not highly confident, it is difficult to identify strong indications of any class, and the tokens accordingly do not have high attribution scores for any class and vice versa. We conduct extensive experiments on six datasets with two popular pre-trained language models in the in-domain and out-of-domain settings. The results show that CME improves calibration performance in all settings. The expected calibration errors are further reduced when combined with temperature scaling. Our findings highlight that model explanations can help calibrate posterior estimates.

We investigate model calibration in the setting of zero-shot cross-lingual transfer with large-scale pre-trained language models. The level of model calibration is an important metric for evaluating the trustworthiness of predictive models. There exists an essential need for model calibration when natural language models are deployed in critical tasks. We study different post-training calibration methods in structured and unstructured prediction tasks. We find that models trained with data from the source language become less calibrated when applied to the target language and that calibration errors increase with intrinsic task difficulty and relative sparsity of training data. Moreover, we observe a potential connection between the level of calibration error and an earlier proposed measure of the distance from English to other languages. Finally, our comparison demonstrates that among other methods Temperature Scaling (TS) generalizes well to distant languages, but TS fails to calibrate more complex confidence estimation in structured predictions compared to more expressive alternatives like Gaussian Process Calibration.

Multi-hop Question Answering (QA) is a challenging task since it requires an accurate aggregation of information from multiple context paragraphs and a thorough understanding of the underlying reasoning chains. Recent work in multi-hop QA has shown that performance can be boosted by first decomposing the questions into simpler, single-hop questions. In this paper, we explore one additional utility of the multi-hop decomposition from the perspective of explainable NLP: to create explanation by probing a neural QA model with them. We hypothesize that in doing so, users will be better able to predict when the underlying QA system will give the correct answer. Through human participant studies, we verify that exposing the decomposition probes and answers to the probes to users can increase their ability to predict system performance on a question instance basis. We show that decomposition is an effective form of probing QA systems as well as a promising approach to explanation generation. In-depth analyses show the need for improvements in decomposition systems.

Knowledge Distillation (KD) is an effective method to transfer knowledge from one network (a.k.a. teacher) to another (a.k.a. student). In this paper, we study KD on the emotion-related tasks from a new perspective: calibration. We further explore the impact of the mixup data augmentation technique on the distillation objective and propose to use a simple yet effective mixup method informed by training dynamics for calibrating the student models. Underpinned by the regularization impact of the mixup process by providing better training signals to the student models using training dynamics, our proposed mixup strategy gradually enhances the student model’s calibration while effectively improving its performance. We evaluate the calibration of pre-trained language models through knowledge distillation over three tasks of emotion detection, sentiment analysis, and empathy detection. By conducting extensive experiments on different datasets, with both in-domain and out-of-domain test sets, we demonstrate that student models distilled from teacher models trained using our proposed mixup method obtained the lowest Expected Calibration Errors (ECEs) and best performance on both in-domain and out-of-domain test sets.

Imbalanced data distribution is a practical and common challenge in building production-level machine learning (ML) models in industry, where data usually exhibits long-tail distributions. For instance, in virtual AI Assistants, such as Google Assistant, Amazon Alexa and Apple Siri, the “play music” or “set timer” utterance is exposed to an order of magnitude more traffic than other skills. This can easily cause trained models to overfit to the majority classes, categories or intents, lead to model miscalibration. The uncalibrated models output unreliable (mostly overconfident) predictions, which are at high risk of affecting downstream decision-making systems. In this work, we study the calibration of production models in the industry use-case of predicting product return reason codes in customer service conversations of an online retail store; The returns reasons also exhibit class imbalance. To alleviate the resulting miscalibration in the production ML model, we streamline the model development and deployment using focal loss (CITATION).We empirically show the effectiveness of model training with focal loss in learning better calibrated models, as compared to standard cross-entropy loss. Better calibration, in turn, enables better control of the precision-recall trade-off for the models deployed in production.

Previous literature has proved that Pretrained Language Models (PLMs) can store factual knowledge. However, we find that facts stored in the PLMs are not always correct. It motivates us to explore a fundamental question: How do we calibrate factual knowledge in PLMs without re-training from scratch? In this work, we propose a simple and lightweight method CaliNet to achieve this goal. To be specific, we first detect whether PLMs can learn the right facts via a contrastive score between right and fake facts. If not, we then use a lightweight method to add and adapt new parameters to specific factual texts. Experiments on the knowledge probing task show the calibration effectiveness and efficiency. In addition, through closed-book question answering, we find that the calibrated PLM possesses knowledge generalization ability after finetuning.Beyond the calibration performance, we further investigate and visualize the knowledge calibration mechanism.

Current summarization systems yield generic summaries that are disconnected from users’ preferences and expectations. To address this limitation, we present CTRLsum, a generic framework to control generated summaries through a set of keywords. During training keywords are extracted automatically without requiring additional human annotations. At test time CTRLsum features a control function to map control signal to keywords; through engineering the control function, the same trained model is able to be applied to control summaries on various dimensions, while neither affecting the model training process nor the pretrained models. We additionally explore the combination of keywords and text prompts for more control tasks. Experiments demonstrate the effectiveness of CTRLsum on three domains of summarization datasets and five control tasks: (1) entity-centric and (2) length-controllable summarization, (3) contribution summarization on scientific papers, (4) invention purpose summarization on patent filings, and (5) question-guided summarization on news articles. Moreover, when used in a standard, unconstrained summarization setting, CTRLsum is comparable or better than strong pretrained systems.

Well-designed diagnostic tasks have played a key role in studying the failure of neural nets (NNs) to generalize systematically. Famous examples include SCAN and Compositional Table Lookup (CTL). Here we introduce CTL++, a new diagnostic dataset based on compositions of unary symbolic functions. While the original CTL is used to test length generalization or productivity, CTL++ is designed to test systematicity of NNs, that is, their capability to generalize to unseen compositions of known functions. CTL++ splits functions into groups and tests performance on group elements composed in a way not seen during training. We show that recent CTL-solving Transformer variants fail on CTL++. The simplicity of the task design allows for fine-grained control of task difficulty, as well as many insightful analyses. For example, we measure how much overlap between groups is needed by tested NNs for learning to compose. We also visualize how learned symbol representations in outputs of functions from different groups are compatible in case of success but not in case of failure. These results provide insights into failure cases reported on more complex compositions in the natural language domain. Our code is public.

Named entity recognition (NER) suffers from the scarcity of annotated training data, especially for low-resource languages without labeled data. Cross-lingual NER has been proposed to alleviate this issue by transferring knowledge from high-resource languages to low-resource languages via aligned cross-lingual representations or machine translation results. However, the performance of cross-lingual NER methods is severely affected by the unsatisfactory quality of translation or label projection. To address these problems, we propose a Cross-lingual Entity Projection framework (CROP) to enable zero-shot cross-lingual NER with the help of a multilingual labeled sequence translation model. Specifically, the target sequence is first translated into the source language and then tagged by a source NER model. We further adopt a labeled sequence translation model to project the tagged sequence back to the target language and label the target raw sentence. Ultimately, the whole pipeline is integrated into an end-to-end model by the way of self-training. Experimental results on two benchmarks demonstrate that our method substantially outperforms the previous strong baseline by a large margin of +3 7 F1 scores and achieves state-of-the-art performance.

Videos often capture objects, their visible properties, their motion, and the interactions between different objects. Objects also have physical properties such as mass, which the imaging pipeline is unable to directly capture. However, these properties can be estimated by utilizing cues from relative object motion and the dynamics introduced by collisions. In this paper, we introduce CRIPP-VQA, a new video question answering dataset for reasoning about the implicit physical properties of objects in a scene. CRIPP-VQA contains videos of objects in motion, annotated with questions that involve counterfactual reasoning about the effect of actions, questions about planning in order to reach a goal, and descriptive questions about visible properties of objects. The CRIPP-VQA test set enables evaluation under several out-of-distribution settings – videos with objects with masses, coefficients of friction, and initial velocities that are not observed in the training distribution. Our experiments reveal a surprising and significant performance gap in terms of answering questions about implicit properties (the focus of this paper) and explicit properties of objects (the focus of prior work).

Context-dependent text-to-SQL is the task of translating multi-turn questions into database-related SQL queries. Existing methods typically focus on making full use of history context or previously predicted SQL for currently SQL parsing, while neglecting to explicitly comprehend the schema and conversational dependency, such as co-reference, ellipsis and user focus change. In this paper, we propose CQR-SQL, which uses auxiliary Conversational Question Reformulation (CQR) learning to explicitly exploit schema and decouple contextual dependency for multi-turn SQL parsing. Specifically, we first present a schema enhanced recursive CQR method to produce domain-relevant self-contained questions. Secondly, we train CQR-SQL models to map the semantics of multi-turn questions and auxiliary self-contained questions into the same latent space through schema grounding consistency task and tree-structured SQL parsing consistency task, which enhances the abilities of SQL parsing by adequately contextual understanding. At the time of writing, our CQR-SQL achieves new state-of-the-art results on two context-dependent text-to-SQL benchmarks SParC and CoSQL.

Prompt tuning is a new few-shot transfer learning technique that only tunes the learnable prompt for pre-trained vision and language models such as CLIP. However, existing prompt tuning methods tend to learn spurious or entangled representations, which leads to poor generalization to unseen concepts.Towards non-spurious and efficient prompt learning from limited examples, this paper presents a novel Counterfactual Prompt Learning (CPL) method for vision and language models, which simultaneously employs counterfactual generation and contrastive learning in a joint optimization framework.Particularly, CPL constructs counterfactual by identifying minimal non-spurious feature change between semantically-similar positive and negative samples that causes concept change, and learns more generalizable prompt representation from both factual and counterfactual examples via contrastive learning. Extensive experiments demonstrate that CPL can obtain superior few-shot performance on different vision and language tasks than previous prompt tuning methods on CLIP. On image classification, we achieve 3.55% average relative improvement on unseen classes across seven datasets; on image-text retrieval and visual question answering, we gain up to 4.09% and 25.08% relative improvements across three few-shot scenarios on unseen test sets respectively.

Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity. For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs. However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs. Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers. Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration. Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically. To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer. Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.

Comparative opinion is a common linguistic phenomenon. The opinion is expressed by comparing multiple targets on a shared aspect, e.g., “camera A is better than camera B in picture quality”. Among the various subtasks in opinion mining, comparative opinion classification is relatively less studied. Current solutions use rules or classifiers to identify opinions, i.e., better, worse, or same, through feature engineering. Because the features are directly derived from the input sentence, these solutions are sensitive to the order of the targets mentioned in the sentence. For example, “camera A is better than camera B” means the same as “camera B is worse than camera A”; but the features of these two sentences are completely different. In this paper, we approach comparative opinion classification through prompt learning, taking the advantage of embedded knowledge in pre-trained language model. We design a twin framework with dual prompts, named CORT. This extremely simple model delivers state-of-the-art and robust performance on all benchmark datasets for comparative opinion classification. We believe CORT well serves as a new baseline for comparative opinion classification.

Counterfactual data augmentation (CDA) – i.e., adding minimally perturbed inputs during training – helps reduce model reliance on spurious correlations and improves generalization to out-of-distribution (OOD) data. Prior work on generating counterfactuals only considered restricted classes of perturbations, limiting their effectiveness. We present Counterfactual Generation via Retrieval and Editing (CORE), a retrieval-augmented generation framework for creating diverse counterfactual perturbations for CDA. For each training example, CORE first performs a dense retrieval over a task-related unlabeled text corpus using a learned bi-encoder and extracts relevant counterfactual excerpts. CORE then incorporates these into prompts to a large language model with few-shot learning capabilities, for counterfactual editing. Conditioning language model edits on naturally occurring data results in more diverse perturbations. Experiments on natural language inference and sentiment analysis benchmarks show that CORE counterfactuals are more effective at improving generalization to OOD data compared to other DA approaches. We also show that the CORE retrieval framework can be used to encourage diversity in manually authored perturbations.

Conceptual knowledge is fundamental to human cognition and knowledge bases. However, existing knowledge probing works only focus on evaluating factual knowledge of pre-trained language models (PLMs) and ignore conceptual knowledge. Since conceptual knowledge often appears as implicit commonsense behind texts, designing probes for conceptual knowledge is hard. Inspired by knowledge representation schemata, we comprehensively evaluate conceptual knowledge of PLMs by designing three tasks to probe whether PLMs organize entities by conceptual similarities, learn conceptual properties, and conceptualize entities in contexts, respectively. For the tasks, we collect and annotate 24k data instances covering 393 concepts, which is COPEN, a COnceptual knowledge Probing bENchmark. Extensive experiments on different sizes and types of PLMs show that existing PLMs systematically lack conceptual knowledge and suffer from various spurious correlations. We believe this is a critical bottleneck for realizing human-like cognition in PLMs. COPEN and our codes are publicly released at https://github.com/THU-KEG/COPEN.

Recent work on question generation has largely focused on factoid questions such as who, what,where, when about basic facts. Generating open-ended why, how, what, etc. questions thatrequire long-form answers have proven more difficult. To facilitate the generation of openended questions, we propose CONSISTENT, a new end-to-end system for generating openended questions that are answerable from and faithful to the input text. Using news articles asa trustworthy foundation for experimentation, we demonstrate our model’s strength over several baselines using both automatic and human based evaluations. We contribute an evaluationdataset of expert-generated open-ended questions. We discuss potential downstream applications for news media organizations.