EMNLP 2023 论文列表

We investigate the problem of generating instructions to guide humans to navigate in simulated residential environments. A major issue with current models is hallucination: they generate references to actions or objects that are inconsistent with what a human follower would perform or encounter along the described path. We develop a model that detects these hallucinated references by adopting a model pre-trained on a large corpus of image-text pairs, and fine-tuning it with a contrastive loss that separates correct instructions from instructions containing synthesized hallucinations. Our final model outperforms several baselines, including using word probability estimated by the instruction-generation model, and supervised models based on LSTM and Transformer.

Large language models have recently risen in popularity due to their ability to perform many natural language tasks without requiring any fine-tuning. In this work, we focus on two novel ideas: (1) generating definitions from examples and using them for zero-shot classification, and (2) investigating how an LLM makes use of the definitions. We thoroughly analyze the performance of GPT-3 model for fine-grained multi-label conspiracy theory classification of tweets using zero-shot labeling. In doing so, we asses how to improve the labeling by providing minimal but meaningful context in the form of the definitions of the labels. We compare descriptive noun phrases, human-crafted definitions, introduce a new method to help the model generate definitions from examples, and propose a method to evaluate GPT-3's understanding of the definitions. We demonstrate that improving definitions of class labels has a direct consequence on the downstream classification results.

We introduce ECHo (Event Causality Inference via Human-Centric Reasoning), a diagnostic dataset of event causality inference grounded in visio-linguistic social scenarios. ECHo employs real-world human-centric deductive information building on a television crime drama. ECHo requires the Theory-of-Mind (ToM) ability to understand and reason about social interactions based on multimodal information. Using ECHo, we propose a unified Chain-of-Thought (CoT) framework to assess the reasoning capability of current AI systems. Our ToM-enhanced CoT pipeline accommodates various large foundation models in both zero-shot and few-shot visio-linguistic reasoning. We use this framework to scrutinize recent large foundation models such as InstructGPT and MiniGPT-4 on three diagnostic human-centric tasks. Further analysis demonstrates ECHo as a challenging dataset to expose imperfections and inconsistencies in reasoning. Our data and code are publicly available at [https://github.com/YuxiXie/ECHo](https://github.com/YuxiXie/ECHo).

The success of ChatGPT validates the potential of large language models (LLMs) in artificial general intelligence (AGI). Subsequently, the release of LLMs has sparked the open-source community's interest in instruction-tuning, which is deemed to accelerate ChatGPT's replication process. However, research on instruction-tuning LLMs in Chinese, the world's most spoken language, is still in its early stages. Therefore, this paper makes an in-depth empirical study of instruction-tuning LLMs in Chinese, which can serve as a cookbook that provides valuable findings for effectively customizing LLMs that can better respond to Chinese instructions. Specifically, we systematically explore the impact of LLM bases, parameter-efficient methods, instruction data types, which are the three most important elements for instruction-tuning. Besides, we also conduct experiment to study the impact of other factors, e.g., chain-of-thought data and human-value alignment. We hope that this empirical study can make a modest contribution to the open Chinese version of ChatGPT. This paper will release a powerful Chinese LLM that is comparable to ChatGLM. The code and data are available at https: //github.com/PhoebusSi/Alpaca-CoT.

Large language models (LLMs) have been shown to possess impressive capabilities, while also raising crucial concerns about the faithfulness of their responses. A primary issue arising in this context is the management of (un)answerable queries by LLMs, which often results in hallucinatory behavior due to overconfidence. In this paper, we explore the behavior of LLMs when presented with (un)answerable queries. We ask: do models \textit{represent} the fact that the question is (un)answerable when generating a hallucinatory answer? Our results show strong indications that such models encode the answerability of an input query, with the representation of the first decoded token often being a strong indicator. These findings shed new light on the spatial organization within the latent representations of LLMs, unveiling previously unexplored facets of these models. Moreover, they pave the way for the development of improved decoding techniques with better adherence to factual generation, particularly in scenarios where query (un)answerability is a concern.

Most existing debiasing methods for multimodal models, including causal intervention and inference methods, utilize approximate heuristics to represent the biases, such as shallow features from early stages of training or unimodal features for multimodal tasks like VQA, etc., which may not be accurate. In this paper, we study bias arising from confounders in a causal graph for multimodal data, and examine a novel approach that leverages causally-motivated information minimization to learn the confounder representations. Robust predictive features contain diverse information that helps a model generalize to out-of-distribution data. Hence, minimizing the information content of features obtained from a pretrained biased model helps learn the simplest predictive features that capture the underlying data distribution. We treat these features as confounder representations and use them via methods motivated by causal theory to remove bias from models. We find that the learned confounder representations indeed capture dataset biases and the proposed debiasing methods improve out-of-distribution (OOD) performance on multiple multimodal datasets without sacrificing in-distribution performance. Additionally, we introduce a novel metric to quantify the sufficiency of spurious features in models' predictions that further demonstrates the effectiveness of our proposed methods.

Emotion arcs capture how an individual (or a population) feels over time. They are widely used in industry and research; however, there is little work on evaluating the automatically generated arcs. This is because of the difficulty of establishing the true (gold) emotion arc. Our work, for the first time, systematically and quantitatively evaluates automatically generated emotion arcs. We also compare two common ways of generating emotion arcs: Machine-Learning (ML) models and Lexicon-Only (LexO) methods. By running experiments on 18 diverse datasets in 9 languages, we show that despite being markedly poor at instance level emotion classification, LexO methods are highly accurate at generating emotion arcs when aggregating information from hundreds of instances. We also show, through experiments on six indigenous African languages, as well as Arabic, and Spanish, that automatic translations of English emotion lexicons can be used to generate high-quality emotion arcs in less-resource languages. This opens up avenues for work on emotions in languages from around the world; which is crucial for commerce, public policy, and health research in service of speakers often left behind. Code and resources: https://github.com/dteodore/EmotionArcs

With the rapid progress of large language models (LLMs), many downstream NLP tasks can be well solved given appropriate prompts. Though model developers and researchers work hard on dialog safety to avoid generating harmful content from LLMs, it is still challenging to steer AI-generated content (AIGC) for the human good. As powerful LLMs are devouring existing text data from various domains (e.g., GPT-3 is trained on 45TB texts), it is natural to doubt whether the private information is included in the training data and what privacy threats can these LLMs and their downstream applications bring. In this paper, we study the privacy threats from OpenAI's ChatGPT and the New Bing enhanced by ChatGPT and show that application-integrated LLMs may cause new privacy threats. To this end, we conduct extensive experiments to support our claims and discuss LLMs' privacy implications.

Stance detection on social media is challenging for Large Language Models (LLMs), as emerging slang and colloquial language in online conversations often contain deeply implicit stance labels. Chain-of-Thought (COT) prompting has recently been shown to improve performance on stance detection tasks --- alleviating some of these issues. However, COT prompting still struggles with implicit stance identification. This challenge arises because many samples are initially challenging to comprehend before a model becomes familiar with the slang and evolving knowledge related to different topics, all of which need to be acquired through the training data. In this study, we address this problem by introducing COT Embeddings which improve COT performance on stance detection tasks by embedding COT reasonings and integrating them into a traditional RoBERTa-based stance detection pipeline. Our analysis demonstrates that 1) text encoders can leverage COT reasonings with minor errors or hallucinations that would otherwise distort the COT output label. 2) Text encoders can overlook misleading COT reasoning when a sample's prediction heavily depends on domain-specific patterns. Our model achieves SOTA performance on multiple stance detection datasets collected from social media.

News articles are driven by the informational sources journalists use in reporting. Modeling when, how and why sources get used together in stories can help us better understand the information we consume and even help journalists with the task of producing it. In this work, we take steps toward this goal by constructing the largest and widest-ranging annotated dataset, to date, of informational sources used in news writing. We first show that our dataset can be used to train high-performing models for information detection and source attribution. Then, we introduce a novel task, source prediction, to study the compositionality of sources in news articles -- i.e. how they are chosen to complement each other. We show good modeling performance on this task, indicating that there is a pattern to the way different sources are used \textit{together} in news storytelling. This insight opens the door for a focus on sources in narrative science (i.e. planning-based language generation) and computational journalism (i.e. a source-recommendation system to aid journalists writing stories). All data and model code can be found at https://github.com/alex2awesome/source-exploration.

Understanding and characterizing the discus- sions around key events in news streams is important for analyzing political discourse. In this work, we study the problem of identification of such key events and the news articles associated with those events from news streams. We propose a generic framework for news stream clustering that analyzes the temporal trend of news articles to automatically extract the underlying key news events that draw significant media attention. We characterize such key events by generating event summaries, based on which we form document clusters in an unsupervised fashion. We evaluate our simple yet effective framework, and show that it produces more coherent event-focused clusters. To demonstrate the utility of our approach, and facilitate future research along the line, we use our framework to construct KeyEvents, a dataset of 40k articles with 611 key events from 11 topics.

We present a novel retrofitting method to induce emotion aspects into pre-trained language models (PLMs) such as BERT and RoBERTa. Our method updates pre-trained network weights using contrastive learning so that the text fragments exhibiting similar emotions are encoded nearby in the representation space, and the fragments with different emotion content are pushed apart. While doing so, it also ensures that the linguistic knowledge already present in PLMs is not inadvertently perturbed. The language models retrofitted by our method, i.e., BERTEmo and RoBERTaEmo, produce emotion-aware text representations, as evaluated through different clustering and retrieval metrics. For the downstream tasks on sentiment analysis and sarcasm detection, they perform better than their pre-trained counterparts (about 1% improvement in F1-score) and other existing approaches. Additionally, a more significant boost in performance is observed for the retrofitted models over pre-trained ones in few-shot learning setting.

Pre-trained language models (PLMs) have achieved the preeminent position in dense retrieval due to their powerful capacity in modeling intrinsic semantics. However, most existing PLM-based retrieval models encounter substantial computational costs and are infeasible for processing long documents. In this paper, a novel retrieval model Longtriever is proposed to embrace three core challenges of long document retrieval: substantial computational cost, incomprehensive document understanding, and scarce annotations. Longtriever splits long documents into short blocks and then efficiently models the local semantics within a block and the global context semantics across blocks in a tightly-coupled manner. A pre-training phase is further proposed to empower Longtriever to achieve a better understanding of underlying semantic correlations. Experimental results on two popular benchmark datasets demonstrate the superiority of our proposal.

Target-Oriented Multimodal Sentiment Classification (TMSC) aims to perform sentiment polarity on a target jointly considering its corresponding multiple modalities including text, image, and others. Current researches mainly work on either of two types of targets in a decentralized manner. One type is entity, such as a person name, a location name, etc. and the other is aspect, such as 'food', 'service', etc. We believe that this target type based division in task modelling is not necessary because the sentiment polarity of the specific target is not governed by its type but its context. For this reason, we propose a unified model for target-oriented multimodal sentiment classification, so called UnifiedTMSC. It is prompt-based language modelling and performs well on four datasets spanning the above two target types. Specifically, we design descriptive prompt paraphrasing to reformulate TMSC task via (1) task paraphrasing, which obtains paraphrased prompts based on the task description through a paraphrasing rule, and (2) image prefix tuning, which optimizes a small continuous image vector throughout the multimodal representation space of text and images. Conducted on two entity-level multimodal datasets: Twitter-2015 and Twitter-2017, and two aspect-level multimodal datasets: Multi-ZOL and MASAD, the experimental results show the effectiveness of our UnifiedTMSC.

Persuasive dialogue aims to persuade users to achieve some targets by conversations. While previous persuasion models have achieved notable successes, they mostly base themselves on utterance semantic matching, and an important aspect has been ignored, that is, the strategy of the conversations, for example, the agent can choose an \textit{emotional-appeal} strategy to impress users. Compared with utterance semantics, conversation strategies are high-level concepts, which can be informative and provide complementary information to achieve effective persuasions. In this paper, we propose to build a persuasion model by jointly modeling the conversation semantics and strategies, where we design a BERT-like module and an auto-regressive predictor to match the semantics and strategies, respectively. Experimental results indicate that our proposed approach can significantly improve the state-of-the-art baseline by 5\% on a small dataset and 37\% on a large dataset in terms of Recall@1. Detailed analyses show that the auto-regressive predictor contributes most to the final performance.

Existing sentence ordering approaches generally employ encoder-decoder frameworks with the pointer net to recover the coherence by recurrently predicting each sentence step-by-step. Such an autoregressive manner only leverages unilateral dependencies during decoding and cannot fully explore the semantic dependency between sentences for ordering. To overcome these limitations, in this paper, we propose a novel Non-Autoregressive Ordering Network, dubbed \textit{NAON}, which explores bilateral dependencies between sentences and predicts the sentence for each position in parallel. We claim that the non-autoregressive manner is not just applicable but also particularly suitable to the sentence ordering task because of two peculiar characteristics of the task: 1) each generation target is in deterministic length, and 2) the sentences and positions should match exclusively. Furthermore, to address the repetition issue of the naive non-autoregressive Transformer, we introduce an exclusive loss to constrain the exclusiveness between positions and sentences. To verify the effectiveness of the proposed model, we conduct extensive experiments on several common-used datasets and the experimental results show that our method outperforms all the autoregressive approaches and yields competitive performance compared with the state-of-the-arts. The codes are available at: \url{https://github.com/steven640pixel/nonautoregressive-sentence-ordering}.

With the recent undeniable advancement in reasoning abilities in large language models (LLMs) like ChatGPT and GPT-4, there is a growing trend for using LLMs on various tasks. One area where LLMs can be employed is as an alternative evaluation metric for complex generative tasks, which generally demands expensive human judges to complement the traditional automatic metrics for various evaluation dimensions such as fluency and consistency. In this work, we conduct extensive analysis to investigate the stability and reliability of LLMs as automatic evaluators for abstractive summarization. We found that while ChatGPT and GPT-4 outperform the commonly used automatic metrics, they are not ready as human replacements due to significant limitations. That is, LLM evaluators rate each candidate system inconsistently and are dimension-dependent. They also struggle to compare candidates with close performance and become more unreliable with higher-quality summaries by obtaining a lower correlation with humans. In other words, with better abstractive summarization systems being introduced at a fast pace, LLMs may result in misleading and unreliable evaluations.

The de-identification task aims to detect and remove the protected health information from electronic medical records (EMRs). Previous studies generally focus on the within-hospital setting and achieve great successes, while the cross-hospital setting has been overlooked. This study introduces a new de-identification dataset comprising EMRs from three hospitals in China, creating a benchmark for evaluating both within- and cross-hospital generalization. We find significant domain discrepancy between hospitals. A model with almost perfect within-hospital performance struggles when transferred across hospitals. Further experiments show that pretrained language models and some domain generalization methods can alleviate this problem. We believe that our data and findings will encourage investigations on the generalization of medical NLP models.

Large Language Models (LLMs) prompted to generate chain-of-thought (CoT) exhibit impressive reasoning capabilities. Recent attempts at prompt decomposition toward solving complex, multi-step reasoning problems depend on the ability of the LLM to simultaneously decompose and solve the problem. A significant disadvantage is that foundational LLMs are typically not available for fine-tuning, making adaptation computationally prohibitive. We believe (and demonstrate) that problem decomposition and solution generation are distinct capabilites, better addressed in separate modules, than by one monolithic LLM. We introduce DaSLaM, which uses a decomposition generator to decompose complex problems into subproblems that require fewer reasoning steps. These subproblems are answered by a solver. We use a relatively small (13B parameters) LM as the decomposition generator, which we train using policy gradient optimization to interact with a solver LM (regarded as black-box) and guide it through subproblems, thereby rendering our method solver-agnostic. Evaluation on multiple different reasoning datasets reveal that with our method, a 175 billion parameter LM (text-davinci-003) can produce competitive or even better performance, compared to its orders-of-magnitude larger successor, GPT-4. Additionally, we show that DaSLaM is not limited by the solver's capabilities as a function of scale; e.g., solver LMs with diverse sizes give significant performance improvement with our solver-agnostic decomposition technique. Exhaustive ablation studies evince the superiority of our modular finetuning technique over exorbitantly large decomposer LLMs, based on prompting alone.