EMNLP 2023 Papers

Research interest in task-oriented dialogs has increased as systems such as Google Assistant, Alexa and Siri have become ubiquitous in everyday life. However, the impact of academic research in this area has been limited by the lack of datasets that realistically capture the wide array of user pain points. To enable research on some of the more challenging aspects of parsing realistic conversations, we introduce PRESTO, a public dataset of over 550K contextual multilingual conversations between humans and virtual assistants. PRESTO contains a diverse array of challenges that occur in real-world NLU tasks such as disfluencies, code-switching, and revisions. It is the only large scale human generated conversational parsing dataset that provides structured context such as a user's contacts and lists for each example. Our mT5 model based baselines demonstrate that the conversational phenomenon present in PRESTO are challenging to model, which is further pronounced in a low-resource setup.

In document-level event extraction (DEE) tasks, a document typically contains many event records with multiple event roles. Therefore, accurately extracting all event records is a big challenge since the number of event records is not given. Previous works present the entity-based directed acyclic graph (EDAG) generation methods to autoregressively generate event roles, which requires a given generation order. Meanwhile, parallel methods are proposed to generate all event roles simultaneously, but suffer from the inadequate training which manifests zero accuracies on some event roles. In this paper, we propose an Iteratively Parallel Generation method with the Pre-Filling strategy (IPGPF). Event roles in an event record are generated in parallel to avoid order selection, and the event records are iteratively generated to utilize historical results. Experiments on two public datasets show our IPGPF improves $11.7$ F1 than previous parallel models and up to $5.1$ F1 than auto-regressive models under the control variable settings. Moreover, our enhanced IPGPF outperforms other entity-enhanced models and achieves new state-of-the-art performance.

Conversational systems based on Large Language Models (LLMs), such as ChatGPT, show exceptional proficiency in context understanding and response generation. However, they still possess limitations, such as failing to ask clarifying questions to ambiguous queries or refuse users' unreasonable requests, both of which are considered as key aspects of a conversational agent's proactivity. This raises the question of whether LLM-based conversational systems are equipped to handle proactive dialogue problems. In this work, we conduct a comprehensive analysis of LLM-based conversational systems, specifically focusing on three key aspects of proactive dialogues: clarification, target-guided, and non-collaborative dialogues. To trigger the proactivity of LLMs, we propose the Proactive Chain-of-Thought prompting scheme, which augments LLMs with the goal planning capability over descriptive reasoning chains. Empirical findings are discussed to promote future studies on LLM-based proactive dialogue systems.

Human label variation, or annotation disagreement, exists in many natural language processing (NLP) tasks, including natural language inference (NLI). To gain direct evidence of how NLI label variation arises, we build LiveNLI, an English dataset of 1,415 ecologically valid explanations (annotators explain the NLI labels they chose) for 122 MNLI items (at least 10 explanations per item). The LiveNLI explanations confirm that people can systematically vary on their interpretation and highlight within-label variation: annotators sometimes choose the same label for different reasons. This suggests that explanations are crucial for navigating label interpretations in general. We few-shot prompt large language models to generate explanations but the results are inconsistent: they sometimes produces valid and informative explanations, but it also generates implausible ones that do not support the label, highlighting directions for improvement.

Recent work has aimed to capture nuances of human behavior by using LLMs to simulate responses from particular demographics in settings like social science experiments and public opinion surveys. However, there are currently no established ways to discuss or evaluate the quality of such LLM simulations. Moreover, there is growing concern that these LLM simulations are flattened caricatures of the personas that they aim to simulate, failing to capture the multidimensionality of people and perpetuating stereotypes. To bridge these gaps, we present CoMPosT, a framework to characterize LLM simulations using four dimensions: Context, Model, Persona, and Topic. We use this framework to measure open-ended LLM simulations’ susceptibility to caricature, defined via two criteria: individuation and exaggeration. We evaluate the level of caricature in scenarios from existing work on LLM simulations. We find that for GPT-4, simulations of certain demographics (political and marginalized groups) and topics (general, uncontroversial) are highly susceptible to caricature.

Semi-inductive link prediction (LP) in knowledge graphs (KG) is the task of predicting facts for new, previously unseen entities based on context information. Although new entities can be integrated by retraining the model from scratch in principle, such an approach is infeasible for large-scale KGs, where retraining is expensive and new entities may arise frequently. In this paper, we propose and describe a large-scale benchmark to evaluate semi-inductive LP models. The benchmark is based on and extends Wikidata5M: It provides transductive, k-shot, and 0-shot LP tasks, each varying the available information from (i) only KG structure, to (ii) including textual mentions, and (iii) detailed descriptions of the entities. We report on a small study of recent approaches and found that semi-inductive LP performance is far from transductive performance on long-tail entities throughout all experiments. The benchmark provides a test bed for further research into integrating context and textual information in semi-inductive LP models.

Existing text summarization systems have made significant progress in recent years, but typically generate summaries in a single step. The one-shot summarization setting is sometimes inadequate, however, as the generated summary may contain hallucinations or overlook important details related to the reader's interests. In this paper, we address this limitation by proposing SummIt, an iterative text summarization framework based on large language models like ChatGPT. Our framework enables the model to refine the generated summary iteratively through self-evaluation and feedback, closely resembling the iterative process humans undertake when drafting and revising summaries. Furthermore, we explore the potential benefits of integrating knowledge and topic extractors into the framework to enhance summary faithfulness and controllability. We evaluate the performance of our framework on three benchmark summarization datasets through empirical and qualitative analyses. We also conduct a human evaluation to validate the effectiveness of the model's refinements and find a potential issue of over-correction.

Evaluating conversational information retrieval (CIR) systems is a challenging task that requires a significant amount of human labor for annotation. It is imperative to invest significant effort into researching more labor-effective methods for evaluating CIR systems. To touch upon this challenge, we take the first step to involve active testing in CIR evaluation and propose a novel method, called HomCoE. It strategically selects a few data for human annotation, then calibrates the evaluation results to eliminate evaluation biases. As such, it makes an accurate evaluation of the CIR system at low human labor. We experimentally reveal that it consumes less than 1\% of human labor and achieves a consistency rate of 95\%-99\% with human evaluation results. This emphasizes the superiority of our method over other baselines.

In this paper, we leverage existing corpora, WordNet, and dependency parsing to build the first Galician dataset for training semantic role labeling systems in an effort to expand available NLP resources. Additionally, we introduce verb indexing, a new pre-processing method, which helps increase the performance when semantically parsing highly-complex sentences. We use transfer-learning to test both the resource and the verb indexing method. Our results show that the effects of verb indexing were amplified in scenarios where the model was both pre-trained and fine-tuned on datasets utilizing the method, but improvements are also noticeable when only used during fine-tuning. The best-performing Galician SRL model achieved an f1 score of 0.74, introducing a baseline for future Galician SRL systems. We also tested our method on Spanish where we achieved an f1 score of 0.83, outperforming the baseline set by the 2009 CoNLL Shared Task by 0.025 showing the merits of our verb indexing method for pre-processing.

Benefiting from massive corpora and advanced hardware, large language models (LLMs) exhibit remarkable capabilities in language understanding and generation. However, their performance degrades in scenarios where multiple tasks are encountered sequentially, also known as catastrophic forgetting. In this paper, we propose orthogonal low-rank adaptation (O-LoRA), a simple and efficient approach for continual learning in language models, effectively mitigating catastrophic forgetting while learning new tasks. Specifically, O-LoRA learns tasks in different (low-rank) vector subspaces that are kept orthogonal to each other in order to minimize interference. Our method induces only marginal additional parameter costs and requires no user data storage for replay. Experimental results on continual learning benchmarks show that our method outperforms state-of-the-art methods. Furthermore, compared to previous approaches, our method excels in preserving the generalization ability of LLMs on unseen tasks.

Software version migration and program translation are an important and costly part of the lifecycle of large codebases. Traditional machine translation relies on parallel corpora for supervised translation, which is not feasible for program translation due to a dearth of aligned data. Recent unsupervised neural machine translation techniques have overcome data limitations by included techniques such as back translation and low level compiler intermediate representations (IR). These methods face significant challenges due to the noise in code snippet alignment and the diversity of IRs respectively. In this paper we propose a novel model called Code Distillation (CoDist) whereby we capture the semantic and structural equivalence of code in a language agnostic intermediate representation. Distilled code serves as a translation pivot for any programming language, leading by construction to parallel corpora which scale to all available source code by simply applying the distillation compiler. We demonstrate that our approach achieves state-of-the-art performance on CodeXGLUE and TransCoder GeeksForGeeks translation benchmarks, with an average absolute increase of 12.7% on the TransCoder GeeksforGeeks translation benchmark compare to TransCoder-ST.

Recent studies have revealed that Backdoor Attacks can threaten the safety of natural language processing (NLP) models. Investigating the strategies of backdoor attacks will help to understand the model's vulnerability. Most existing textual backdoor attacks focus on generating stealthy triggers or modifying model weights. In this paper, we directly target the interior structure of neural networks and the backdoor mechanism. We propose a novel Trojan Attention Loss (TAL), which enhances the Trojan behavior by directly manipulating the attention patterns. Our loss can be applied to different attacking methods to boost their attack efficacy in terms of attack successful rates and poisoning rates. It applies to not only traditional dirty-label attacks, but also the more challenging clean-label attacks. We validate our method on different backbone models (BERT, RoBERTa, and DistilBERT) and various tasks (Sentiment Analysis, Toxic Detection, and Topic Classification).

Theory of Mind (ToM) is the ability to reason about one's own and others' mental states. ToM plays a critical role in the development of intelligence, language understanding, and cognitive processes. While previous work has primarily focused on first and second-order ToM, we explore higher-order ToM, which involves recursive reasoning on others' beliefs. %We also incorporate a new deception mechanism in ToM reasoning. We introduce Hi-ToM, a Higher Order Theory of Mind benchmark. Our experimental evaluation using various Large Language Models (LLMs) indicates a decline in performance on higher-order ToM tasks, demonstrating the limitations of current LLMs. We conduct a thorough analysis of different failure cases of LLMs, and share our thoughts on the implications of our findings on the future of NLP.

In this paper, we propose SRL4NLP, a new approach for data augmentation by drawing an analogy between image and text processing: Super-resolution learning. This method is based on using high-resolution images to overcome the problem of low resolution images. While this technique is a common usage in image processing when images have a low resolution or are too noisy, it has never been used in NLP. We therefore propose the first adaptation of this method for text classification and evaluate its effectiveness on urgency detection from tweets posted in crisis situations, a very challenging task where messages are scarce and highly imbalanced. We show that this strategy is efficient when compared to competitive state-of-the-art data augmentation techniques on several benchmarks datasets in two languages.

Deep neural classifiers trained with cross-entropy loss (CE loss) often suffer from poor calibration, necessitating the task of out-of-distribution (OOD) detection. Traditional supervised OOD detection methods require expensive manual annotation of in-distribution and OOD samples. To address the annotation bottleneck, we introduce SELFOOD, a self-supervised OOD detection method that requires only in-distribution samples as supervision. We cast OOD detection as an inter-document intra-label (IDIL) ranking problem and train the classifier with our pairwise ranking loss, referred to as IDIL loss. Specifically, given a set of in-distribution documents and their labels, for each label, we train the classifier to rank the softmax scores of documents belonging to that label to be higher than the scores of documents that belong to other labels. Unlike CE loss, our IDIL loss function reaches zero when the desired confidence ranking is achieved and gradients are backpropagated to decrease probabilities associated with incorrect labels rather than continuously increasing the probability of the correct label. Extensive experiments with several classifiers on multiple classification datasets demonstrate the effectiveness of our method in both coarse- and fine-grained settings.

Knowing how to end and resume conversations over time is a natural part of communication, allowing for discussions to span weeks, months, or years. The duration of gaps between conversations dictates which topics are relevant and which questions to ask, and dialogue systems which do not explicitly model time may generate responses that are unnatural. In this work we explore the idea of making dialogue models aware of time, and present GapChat, a multi-session dialogue dataset in which the time between each session varies. While the dataset is constructed in real-time, progress on events in speakers' lives is simulated in order to create realistic dialogues occurring across a long timespan. We expose time information to the model and compare different representations of time and event progress. In human evaluation we show that time-aware models perform better in metrics that judge the relevance of the chosen topics and the information gained from the conversation.

Summaries of medical text shall be faithful by being consistent and factual with source inputs, which is an important but understudied topic for safety and efficiency in healthcare. In this paper, we investigate and improve faithfulness in summarization on a broad range of medical summarization tasks. Our investigation reveals that current summarization models often produce unfaithful outputs for medical input text. We then introduce FaMeSumm, a framework to improve faithfulness by fine-tuning pre-trained language models based on medical knowledge. FaMeSumm performs contrastive learning on designed sets of faithful and unfaithful summaries, and it incorporates medical terms and their contexts to encourage faithful generation of medical terms. We conduct comprehensive experiments on three datasets in two languages: health question and radiology report summarization datasets in English, and a patient-doctor dialogue dataset in Chinese. Results demonstrate that FaMeSumm is flexible and effective by delivering consistent improvements over mainstream language models such as BART, T5, mT5, and PEGASUS, yielding state-of-the-art performances on metrics for faithfulness and general quality. Human evaluation by doctors also shows that FaMeSumm generates more faithful outputs. Our code is available at https://github.com/psunlpgroup/FaMeSumm.

Despite their impressive performance, large language models (LMs) still struggle with reliably generating complex output structures when not finetuned to follow the required output format exactly. To address this issue, grammar-constrained decoding (GCD) can be used to control the generation of LMs, guaranteeing that the output follows a given structure. Most existing GCD methods are, however, limited to specific tasks, such as parsing or code generation. In this work, we demonstrate that formal grammars can describe the output space for a much wider range of tasks and argue that GCD can serve as a unified framework for structured NLP tasks in general. For increased flexibility, we introduce input-dependent grammars, which allow the grammar to depend on the input and thus enable the generation of different output structures for different inputs. We then empirically demonstrate the power and flexibility of GCD-enhanced LMs on (1) information extraction, (2) entity disambiguation, and (3) constituency parsing. Our results indicate that grammar-constrained LMs substantially outperform unconstrained LMs or even beat task-specific finetuned models. Grammar constraints thus hold great promise for harnessing off-the-shelf LMs for a wide range of structured NLP tasks, especially where training data is scarce or finetuning is expensive. Code and data: https://github.com/epfl-dlab/GCD.

Bilingual lexicon induction (BLI) is the task of inducing word translations with a learned mapping function that aligns monolingual word embedding spaces in two different languages. However, most previous methods treat word embeddings as isolated entities and fail to jointly consider both the intra-space and inter-space topological relations between words. This limitation makes it challenging to align words from embedding spaces with distinct topological structures, especially when the assumption of isomorphism may not hold. To this end, we propose a novel approach called the Structure-Aware Generative Adversarial Network (SA-GAN) model to explicitly capture multiple topological structure information to achieve accurate BLI. Our model first incorporates two lightweight graph convolutional networks (GCNs) to leverage intra-space topological correlations between words for generating source and target embeddings. We then employ a GAN model to explore inter-space topological structures by learning a global mapping function that initially maps the source embeddings to the target embedding space. To further align the coarse-grained structures, we develop a pair-wised local mapping (PLM) strategy that enables word-specific transformations in an unsupervised manner. Extensive experiments conducted on public datasets, including languages with both distant and close etymological relationships, demonstrate the effectiveness of our proposed SA-GAN model.