EMNLP 2023 论文列表

Lifelong sequence generation (LSG), a problem in continual learning, aims to continually train a model on a sequence of generation tasks to learn constantly emerging new generation patterns while avoiding the forgetting of previous knowledge. Existing LSG methods mainly focus on maintaining old knowledge while paying little attention to knowledge transfer across tasks. In contrast, humans can better learn new tasks by leveraging previously acquired knowledge from similar tasks. Inspired by the learning paradigm of humans, we propose Dynamic Module Expansion and Adaptation (DMEA), which enables the model to dynamically determine the architecture for acquiring new knowledge based on task correlation and select the most similar previous tasks to facilitate adaptation to new tasks. In addition, as the learning process can easily be biased towards the current task which might cause more severe forgetting of previously learned knowledge, we propose dynamic gradient scaling to balance the learning of the current task and replayed tasks. With extensive experiments, we demonstrate that DMEA can consistently outperform existing methods in different LSG settings.

Though majority vote among annotators is typically used for ground truth labels in machine learning, annotator disagreement in tasks such as hate speech detection may reflect systematic differences in opinion across groups, not noise. Thus, a crucial problem in hate speech detection is determining if a statement is offensive to the demographic group that it targets, when that group may be a small fraction of the annotator pool. We construct a model that predicts individual annotator ratings on potentially offensive text and combines this information with the predicted target group of the text to predict the ratings of target group members. We show gains across a range of metrics, including raising performance over the baseline by 22% at predicting individual annotators' ratings and by 33% at predicting variance among annotators, which provides a metric for model uncertainty downstream. We find that annotators' ratings can be predicted using their demographic information as well as opinions on online content, and that non-invasive questions on annotators' online experiences minimize the need to collect demographic information when predicting annotators' opinions.

We explore the possibility of improving probabilistic models in structured prediction. Specifically, we combine the models with constrained decoding approaches in the context of token classification for information extraction. The decoding methods search for constraint-satisfying label-assignments while maximizing the total probability. To do this, we evaluate several existing approaches, as well as propose a novel decoding method called Lazy-$k$. Our findings demonstrate that constrained decoding approaches can significantly improve the models' performances, especially when using smaller models. The Lazy-$k$ approach allows for more flexibility between decoding time and accuracy. The code for using Lazy-$k$ decoding can be found at https://github.com/ArthurDevNL/lazyk.

Recent advancements in natural language processing have demonstrated the efficacy of pre-trained language models for various downstream tasks through prompt-based fine-tuning. In contrast to standard fine-tuning, which relies solely on labeled examples, prompt-based fine-tuning combines a few labeled examples (few shot) with guidance through prompts tailored for the specific language and task. For low-resource languages, where labeled examples are limited, prompt-based fine-tuning appears to be a promising alternative. In this paper, we compare prompt-based and standard fine-tuning for the popular task of text classification in Urdu and Roman Urdu languages. We conduct experiments using five datasets, covering different domains, and pre-trained multilingual transformers. The results reveal that significant improvement of up to 13% in accuracy is achieved by prompt-based fine-tuning over standard fine-tuning approaches. This suggests the potential of prompt-based fine-tuning as a valuable approach for low-resource languages with limited labeled data.

Cross-document relation extraction (CodRED) task aims to infer the relation between two entities mentioned in different documents within a reasoning path. Previous studies have concentrated on merely capturing implicit relations between the entities. However, humans usually utilize explicit information chains such as hyperlinks or additional searches to find the relations between two entities. Inspired by this, we propose Path wIth expLOraTion (PILOT) that provides the enhanced reasoning path by exploring the explicit clue information within the documents. PILOT finds the bridging entities which directly guide the paths between the entities and then employs them as stepstones to navigate desirable paths. We show that models with PILOT outperform the baselines in the CodRED task. Furthermore, we offer a variety of analyses to verify the validity of the reasoning paths constructed through PILOT, including evaluations using large language models such as ChatGPT.

With the rise of powerful closed-sourced LLMs (ChatGPT, GPT-4), there are increasing interests in distilling the capabilies of close-sourced LLMs to smaller open-sourced LLMs. Previous distillation methods usually prompt ChatGPT to generate a set of instructions and answers, for the student model to learn. However, such standard distillation approach neglects the merits and conditions of the student model. Inspired by modern teaching principles, we design a personalised distillation process, in which the student attempts to solve a task first, then the teacher provides an adaptive refinement for the student to improve. Instead of feeding the student with teacher's prior, personalised distillation enables personalised learning for the student model, as it only learns on examples it makes mistakes upon and learns to improve its own solution. On code generation, personalised distillation consistently outperforms standard distillation with only one third of the data. With only 2.5-3K personalised examples that incur a data-collection cost of 4-6\$, we boost CodeGen-mono-16B by 7\% to achieve 36.4\% pass@1 and StarCoder by 12.2\% to achieve 45.8\% pass@1 on HumanEval.\footnote{We will release our models and codes upon acceptance}

Temporal reasoning represents a vital component of human communication and understanding, yet remains an underexplored area within the context of Large Language Models (LLMs). Despite LLMs demonstrating significant proficiency in a range of tasks, a comprehensive, large-scale analysis of their temporal reasoning capabilities is missing. Our paper addresses this gap, presenting the first extensive benchmarking of LLMs on temporal reasoning tasks. We critically evaluate 8 different LLMs across 6 datasets using 3 distinct prompting strategies. Additionally, we broaden the scope of our evaluation by including in our analysis 2 Code Generation LMs. Beyond broad benchmarking of models and prompts, we also conduct a fine-grained investigation of performance across different categories of temporal tasks. We further analyze the LLMs on varying temporal aspects, offering insights into their proficiency in understanding and predicting the continuity, sequence, and progression of events over time. Our findings reveal a nuanced depiction of the capabilities and limitations of the models within temporal reasoning, offering a comprehensive reference for future research in this pivotal domain.

Due to the prohibitively high cost of creating error correction datasets, most Factual Claim Correction methods rely on a powerful verification model to guide the correction process. This leads to a significant drop in performance in domains like Scientific Claim Correction, where good verification models do not always exist. In this work we introduce SciFix, a claim correction system that does not require a verifier but is able to outperform existing methods by a considerable margin — achieving correction accuracy of 84% on the SciFact dataset, 77% on SciFact-Open and 72.75% on the CovidFact dataset, compared to next best accuracies of 7.6%, 5% and 15% on the same datasets respectively. Our method leverages the power of prompting with LLMs during training to create a richly annotated dataset that can be used for fully supervised training and regularization. We additionally use a claim-aware decoding procedure to improve the quality of corrected claims. Our method outperforms the very LLM that was used to generate the annotated dataset — with FewShot Prompting on GPT3.5 achieving 58%, 61% and 64% on the respective datasets, a consistently lower correction accuracy, despite using nearly 800 times as many parameters as our model.

Neural models, including large language models (LLMs), achieve superior performance on multi-hop question-answering. To elicit reasoning capabilities from LLMs, recent works propose using the chain-of-thought (CoT) mechanism to generate both the reasoning chain and the answer, which enhances the model's capabilities in conducting multi-hop reasoning. However, several challenges still remain: such as struggling with inaccurate reasoning, hallucinations, and lack of interpretability. On the other hand, information extraction (IE) identifies entities, relations, and events grounded to the text. The extracted structured information can be easily interpreted by humans and machines (Grishman, 2019). In this work, we investigate constructing and leveraging extracted semantic structures (graphs) for multi-hop question answering, especially the reasoning process. Empirical results and human evaluations show that our framework: generates more faithful reasoning chains and substantially improves the QA performance on two benchmark datasets. Moreover, the extracted structures themselves naturally provide grounded explanations that are preferred by humans, as compared to the generated reasoning chains and saliency-based explanations.

Understanding how styles differ across languages is advantageous for training both humans and computers to generate culturally appropriate text. We introduce an explanation framework to extract stylistic differences from multilingual LMs and compare styles across languages. Our framework (1) generates comprehensive style lexica in any language and (2) consolidates feature importances from LMs into comparable lexical categories. We apply this framework to compare politeness, creating the first holistic multilingual politeness dataset and exploring how politeness varies across four languages. Our approach enables an effective evaluation of how distinct linguistic categories contribute to stylistic variations and provides interpretable insights into how people communicate differently around the world.

Event Causality Extraction (ECE) aims to extract the cause-effect event pairs from the given text, which requires the model to possess a strong reasoning ability to capture event causalities. However, existing works have not adequately exploited the interactions between the cause and effect event that could provide crucial clues for causality reasoning. To this end, we propose an Implicit Cause-Effect interaction (ICE) framework, which formulates ECE as a template-based conditional generation problem. The proposed method captures the implicit intra- and inter-event interactions by incorporating the privileged information (ground truth event types and arguments) for reasoning, and a knowledge distillation mechanism is introduced to alleviate the unavailability of privileged information in the test stage. Furthermore, to facilitate knowledge transfer from teacher to student, we design an event-level alignment strategy named Cause-Effect Optimal Transport (CEOT) to strengthen the semantic interactions of cause-effect event types and arguments. Experimental results indicate that ICE achieves state-of-the-art performance on the ECE-CCKS dataset.

Scholarship on generative pretraining (GPT) remains acutely Anglocentric, leaving serious gaps in our understanding of the whole class of autoregressive models. For example, we have little knowledge about the potential of these models and their societal impacts in diverse linguistic and cultural settings. We alleviate this issue for Arabic, a wide collection of languages and dialectal varieties with more than 400 million population, by introducing JASMINE. JASMINE is a suite of powerful Arabic autoregressive Transformer language models ranging in size between 300 million-6.7 billion parameters pretrained on a large and diverse dataset (~ 235 GB of text). We also carefully design and release a comprehensive benchmark for both automated and human evaluation of Arabic autoregressive models, with coverage of potential social biases, harms, and toxicity. Using our novel benchmark, we evaluate JASMINE extensively showing powerful performance intrinsically as well as in few-shot learning on a wide range of NLP tasks. We aim to responsibly release our models and evaluation benchmark with interested researchers, along with code for experimenting with them.

Scientific literature review generation aims to extract and organize important information from an abundant collection of reference papers and produces corresponding reviews while lacking a clear and logical hierarchy. We observe that a high-quality catalogue-guided generation process can effectively alleviate this problem. Therefore, we present an atomic and challenging task named Hierarchical Catalogue Generation for Literature Review as the first step for review generation, which aims to produce a hierarchical catalogue of a review paper given various references. We construct a novel English Hierarchical Catalogues of Literature Reviews Dataset with 7.6k literature review catalogues and 389k reference papers. To accurately assess the model performance, we design two evaluation metrics for informativeness and similarity to ground truth from semantics and structure. Our extensive analyses verify the high quality of our dataset and the effectiveness of our evaluation metrics. We further benchmark diverse experiments on state-of-the-art summarization models like BART and large language models like ChatGPT to evaluate their capabilities. We further discuss potential directions for this task to motivate future research.

Large language models (LLMs) have showcased remarkable capabilities in complex reasoning through chain of thought (CoT) prompting. Recently, there has been a growing interest in transferring these reasoning abilities from LLMs to smaller models. However, achieving both the diversity and consistency in rationales presents a challenge. In this paper, we focus on enhancing these two aspects and propose Multi-CoT Consistent Knowledge Distillation (MCC-KD) to efficiently distill the reasoning capabilities. In MCC-KD, we generate multiple rationales for each question and enforce consistency among their predictions by minimizing the bidirectional KL-divergence between the answer distributions. We conduct comprehensive experiments to investigate the effectiveness of MCC-KD with different model architectures (LLaMA/FlanT5) and various model scales (3B/7B/11B/13B) on both mathematical reasoning and commonsense reasoning benchmarks. The empirical results demonstrate that MCC-KD achieves superior performance on in-distribution datasets and exhibits a strong generalization ability on out-of-distribution datasets.

While biases disadvantaging African American Language (AAL) have been uncovered in models for tasks such as speech recognition and toxicity detection, there has been little investigation of these biases for language generation models like ChatGPT. We evaluate how well LLMs understand AAL in comparison to White Mainstream English (WME), the encouraged "standard" form of English taught in American classrooms. We measure large language model performance on two tasks: a counterpart generation task, where a model generates AAL given WME and vice versa, and a masked span prediction (MSP) task, where models predict a phrase hidden from their input. Using a novel dataset of AAL texts from a variety of regions and contexts, we present evidence of dialectal bias for six pre-trained LLMs through performance gaps on these tasks.

Text-to-video retrieval (TVR) aims to find the most relevant video in a large video gallery given a query text. The intricate and abundant context of the video challenges the performance and efficiency of TVR. To handle the serialized video contexts, existing methods typically select a subset of frames within a video to represent the video content for TVR. How to select the most representative frames is a crucial issue, whereby the selected frames are required to not only retain the semantic information of the video but also promote retrieval efficiency by excluding temporally redundant frames. In this paper, we make the first empirical study of frame selection for TVR. We systemically classify existing frame selection methods into text-free and text-guided ones, under which we detailedly analyze six different frame selections in terms of effectiveness and efficiency. Among them, two frame selections are first developed in this paper. According to the comprehensive analysis on multiple TVR benchmarks, we empirically conclude that the TVR with proper frame selections can significantly improve the retrieval efficiency without sacrificing the retrieval performance.

To properly explain sentence pairs that provide contradictory (different) information for different conditions, we introduce the task of conditional natural language inference (Cond-NLI) and focus on automatically extracting contradictory aspects and their conditions from a sentence pair. Cond-NLI can help to provide a full spectrum of information, such as when there are multiple answers to a question each addressing a specific condition, or reviews with different opinions for different conditions. We show that widely-used feature-attribution explanation models are not suitable for finding conditions, especially when sentences are long and are written independently. We propose a simple yet effective model for the original NLI task that can successfully extract conditions while not requiring token-level annotations. Our model enhances the interpretability of the NLI task while maintaining comparable accuracy. To evaluate models for the Cond-NLI, we build and release a token-level annotated dataset BioClaim which contains potentially contradictory claims from the biomedical domain. Our experiments show that our proposed model outperforms the full cross-encoder and other baselines in extracting conditions. It also performs on-par with GPT-3 which has an order of magnitude more parameters and trained on a huge amount of data.

Achieving robust language technologies that can perform well across the world's many languages is a central goal of multilingual NLP. In this work, we take stock of and empirically analyse task performance disparities that exist between multilingual task-oriented dialogue (ToD) systems. We first define new quantitative measures of absolute and relative equivalence in system performance, capturing disparities across languages and within individual languages. Through a series of controlled experiments, we demonstrate that performance disparities depend on a number of factors: the nature of the ToD task at hand, the underlying pretrained language model, the target language, and the amount of ToD annotated data. We empirically prove the existence of the adaptation and intrinsic biases in current ToD systems: e.g., ToD systems trained for Arabic or Turkish using annotated ToD data fully parallel to English ToD data still exhibit diminished ToD task performance. Beyond providing a series of insights into the performance disparities of ToD systems in different languages, our analyses offer practical tips on how to approach ToD data collection and system development for new languages.

Phonological reconstruction is one of the central problems in historical linguistics where a proto-word of an ancestral language is determined from the observed cognate words of daughter languages. Computational approaches to historical linguistics attempt to automate the task by learning models on available linguistic data. Several ideas and techniques drawn from computational biology have been successfully applied in this area of computational historical linguistics. Following these lines, we adapt MSA Transformer, a protein language model, to the problem of automated phonological reconstruction. MSA Transformer trains on multiple sequence alignments as input and is, thus, apt for application on aligned cognate words. We, hence, name our model as Cognate Transformer. We also apply the model on another associated task, namely, cognate reflex prediction where a reflex word in a daughter language is predicted based on cognate words from other daughter languages. We show that our model outperforms the existing models on both the tasks, especially when it is pre-trained on masked word prediction task.