EMNLP 2023 Papers

Graph-to-text (G2T) generation takes a graph as input and aims to generate a fluent and faith- ful textual representation of the information in the graph. The task has many applications, such as dialogue generation and question an- swering. In this work, we investigate to what extent the G2T generation problem is solved for previously studied datasets, and how pro- posed metrics perform when comparing generated texts. To help address their limitations, we propose a new metric that correctly identifies factual faithfulness, i.e., given a triple (subject, predicate, object), it decides if the triple is present in a generated text. We show that our metric FactSpotter achieves the highest correlation with human annotations on data correct- ness, data coverage, and relevance. In addition, FactSpotter can be used as a plug-in feature to improve the factual faithfulness of existing models. Finally, we investigate if existing G2T datasets are still challenging for state-of-the-art models. Our code is available online: https://github.com/guihuzhang/FactSpotter.

Medical systematic reviews play a vital role in healthcare decision making and policy. However, their production is time-consuming, limiting the availability of high-quality and up-to-date evidence summaries. Recent advancements in LLMs offer the potential to automatically generate literature reviews on demand, addressing this issue. However, LLMs sometimes generate inaccurate (and potentially misleading) texts by hallucination or omission. In healthcare, this can make LLMs unusable at best and dangerous at worst. We conducted 16 interviews with international systematic review experts to characterize the perceived utility and risks of LLMs in the specific context of medical evidence reviews. Experts indicated that LLMs can assist in the writing process by drafting summaries, generating templates, distilling information, and crosschecking information. They also raised concerns regarding confidently composed but inaccurate LLM outputs and other potential downstream harms, including decreased accountability and proliferation of low-quality reviews. Informed by this qualitative analysis, we identify criteria for rigorous evaluation of biomedical LLMs aligned with domain expert views.

An end-to-end speech-to-text (S2T) translation model is usually initialized from a pre-trained speech recognition encoder and a pre-trained text-to-text (T2T) translation decoder. Although this straightforward setting has been shown empirically successful, there do not exist clear answers to the research questions: 1) how are speech and text modalities fused in S2T model and 2) how to better fuse the two modalities? In this paper, we take the first step toward understanding the fusion of speech and text features in S2T model. We first design and release a 10GB linguistic probing benchmark, namely Speech-Senteval, to investigate the acoustic and linguistic behaviors of S2T models. Preliminary analysis reveals that the uppermost encoder layers of the S2T model can not learn linguistic knowledge efficiently, which is crucial for accurate translation. Based on the finding, we further propose a simple plug-in prompt-learning strategy on the uppermost encoder layers to broaden the abstract representation power of the encoder of S2T models. We call such a prompt-enhanced S2T model PromptST. Experimental results on four widely-used S2T datasets show that PromptST can deliver significant improvements over a strong baseline by capturing richer linguistic knowledge. Benchmarks, code, and scripts are freely available at https://github.com/ytf-philp/PromptST.

Pixel-based language models process text rendered as images, which allows them to handle any script, making them a promising approach to open vocabulary language modelling. However, recent approaches use text renderers that produce a large set of almost-equivalent input patches, which may prove sub-optimal for downstream tasks, due to redundancy in the input representations. In this paper, we investigate four approaches to rendering text in the PIXEL model (Rust et al., 2023), and find that simple character bigram rendering brings improved performance on sentence-level tasks without compromising performance on token-level or multilingual tasks. This new rendering strategy also makes it possible to train a more compact model with only 22M parameters that performs on par with the original 86M parameter model. Our analyses show that character bigram rendering leads to a consistently better model but with an anisotropic patch embedding space, driven by a patch frequency bias, highlighting the connections between image patch- and tokenization-based language models.

The choice of input text prompt plays a critical role in the performance of Vision-Language Pretrained (VLP) models such as CLIP. We present APoLLo, a unified multi-modal approach that combines Adapter and Prompt learning for Vision-Language models. Our method is designed to substantially improve the generalization capabilities of VLP models when they are fine-tuned in a few-shot setting. We introduce trainable cross-attention-based adapter layers in conjunction with vision and language encoders to strengthen the alignment between the two modalities. We enforce consistency between the respective encoder branches (receiving augmented inputs) to prevent overfitting in downstream tasks. Our method is evaluated on three representative tasks: generalization to novel classes, cross-dataset evaluation, and unseen domain shifts. In practice, APoLLo achieves a relative gain up to 6.03% over MaPLe (SOTA) on novel classes for 10 diverse image recognition datasets.

We propose a novel unsupervised keyphrase extraction approach, called SAMRank, which uses only a self-attention map in a pre-trained language model (PLM) to determine the importance of phrases. Most recent approaches for unsupervised keyphrase extraction mainly utilize contextualized embeddings to capture semantic relevance between words, sentences, and documents. However, due to the anisotropic nature of contextual embeddings, these approaches may not be optimal for semantic similarity measurements. SAMRank as proposed here computes the importance of phrases solely leveraging a self-attention map in a PLM, in this case BERT and GPT-2, eliminating the need to measure embedding similarities. To assess the level of importance, SAMRank combines both global and proportional attention scores through calculations using a self-attention map. We evaluate the SAMRank on three keyphrase extraction datasets: Inspec, SemEval2010, and SemEval2017. The experimental results show that SAMRank outperforms most embedding-based models on both long and short documents and demonstrating that it is possible to use only a self-attention map for keyphrase extraction without relying on embeddings. Source code is available at https://github.com/kangnlp/SAMRank.

Document image translation (DIT) aims to translate text embedded in images from one language to another. It is a challenging task that needs to understand visual layout with text semantics simultaneously. However, existing methods struggle to capture the crucial visual layout in real-world complex document images. In this work, we make the first attempt to incorporate layout knowledge into DIT in an end-to-end way. Specifically, we propose a novel Layout-aware end-to-end Document Image Translation (LayoutDIT) with multi-step conductive decoder. A layout-aware encoder is first introduced to model visual layout relations with raw OCR results. Then a novel multi-step conductive decoder is unified with hidden states conduction across three step-decoders to achieve the document translation step by step. Benefiting from the layout-aware end-to-end joint training, our LayoutDIT outperforms state-of-the-art methods with better parameter efficiency. Besides, we create a new multi-domain document image translation dataset to validate the model’s generalization. Extensive experiments show that LayoutDIT has a good generalization in diverse and complex layout scenes.

Lexical semantic relations (LSRs) characterize meaning relationships between words and play an important role in systematic generalization on lexical inference tasks. Notably, several tasks that require knowledge of hypernymy still pose a challenge for pretrained language models (LMs) such as BERT, underscoring the need to better align their linguistic behavior with our knowledge of LSRs. In this paper, we propose Balaur, a model that addresses this challenge by modeling LSRs directly in the LM's hidden states throughout pretraining. Motivating our approach is the hypothesis that the internal representations of LMs can provide an interface to their observable linguistic behavior, and that by controlling one we can influence the other. We validate our hypothesis and demonstrate that Balaur generally improves the performance of large transformer-based LMs on a comprehensive set of hypernymy-informed tasks, as well as on the original LM objective. Code and data are made available at https://github.com/mirandrom/balaur

Large neural-based dialog generation models have been applied in many real-life scenarios, yet they are prone to hallucination and tend to produce factually inaccurate outputs which raise great concerns. To alleviate this problem, we propose a plug-and-play retrieval-based framework IKA, which leverages in-context learning and retrieval techniques to enhance LLMs on knowledge grounded dialog generation. We design thorough experiments on a large-scale knowledge graph with 1M+ facts to investigate the effectiveness and generalization of our framework. Experiments show that our method surpasses previous training-based SOTA by a large margin, specifically 46.67% in BLEU4, 26.01% in ROUGE-L, 122.90% in BARTScore and 30.50% in Entity Coverage F1. Further analysis show promising abilities of LLMs to perform knowledge-intensive tasks, which is previously considered weak and understudied.

Inference, especially those derived from inductive processes, is a crucial component in our conversation to complement the information implicitly or explicitly conveyed by a speaker. While recent large language models show remarkable advances in inference tasks, their performance in inductive reasoning, where not all information is present in the context, is far behind deductive reasoning. In this paper, we analyze the behavior of the models based on the task difficulty defined by the semantic information gap -- which distinguishes inductive and deductive reasoning. Our analysis reveals that the information gap between dialogue contexts and desired inferences renders the inductive inference process more challenging. To mitigate this information gap, we investigate a contrastive learning approach by feeding negative samples. Our experiments suggest negative samples help models understand what is wrong and improve their inference generations.

Despite the ability to train capable LLMs, the methodology for maintaining their relevancy and rectifying errors remains elusive. To this end, the past few years have witnessed a surge in techniques for editing LLMs, the objective of which is to alter the behavior of LLMs \textbf{efficiently} within a specific domain without negatively impacting performance across other inputs. This paper embarks on a deep exploration of the problems, methods, and opportunities related to model editing for LLMs. In particular, we provide an exhaustive overview of the task definition and challenges associated with model editing, along with an in-depth empirical analysis of the most progressive methods currently at our disposal. We also build a new benchmark dataset to facilitate a more robust evaluation and pinpoint enduring issues intrinsic to existing techniques. Our objective is to provide valuable insights into the effectiveness and feasibility of each editing technique, thereby assisting the community in making informed decisions on the selection of the most appropriate method for a specific task or context\footnote{Code and datasets are in the supplementary and will be released.}.

Graph neural networks (GNNs) have shown promising performance for knowledge graph reasoning. A recent variant of GNN called progressive relational graph neural network (PRGNN), utilizes relational rules to infer missing knowledge in relational digraphs and achieves notable results. However, during reasoning with PRGNN, two important properties are often overlooked: (1) the sequentiality of relation composition, where the order of combining different relations affects the semantics of the relational rules, and (2) the lagged entity information propagation, where the transmission speed of required information lags behind the appearance speed of new entities. Ignoring these properties leads to incorrect relational rule learning and decreased reasoning accuracy. To address these issues, we propose a novel knowledge graph reasoning approach, the Relational rUle eNhanced Graph Neural Network (RUN-GNN). Specifically, RUN-GNN employs a query related fusion gate unit to model the sequentiality of relation composition and utilizes a buffering update mechanism to alleviate the negative effect of lagged entity information propagation, resulting in higher-quality relational rule learning. Experimental results on multiple datasets demonstrate the superiority of RUN-GNN is superior on both transductive and inductive link prediction tasks.

While question answering over knowledge bases (KBQA) has shown progress in addressing factoid questions, KBQA with numerical reasoning remains relatively unexplored. In this paper, we focus on the complex numerical reasoning in KBQA, and propose a new task, NR-KBQA, which necessitates the ability to perform both multi-hop reasoning and numerical reasoning. We also design a logic form in Python format called PyQL to represent the reasoning process of numerical reasoning questions. To facilitate the development of NR-KBQA, we present a large NR-KBQA dataset called MarkQA, which is automatically constructed by a small set of seeds. Each question in MarkQA is annotated with its corresponding SPARQL query, alongside the step-by-step reasoning path in the QDMR format and PyQL program. Experimental results of some state-of-the-art QA methods performed on the MarkQA dataset show that complex numerical reasoning in KBQA faces great challenges.

Studies in bias and fairness in natural language processing have primarily examined social biases within a single language and/or across few attributes (e.g. gender, race). However, biases can manifest differently across various languages for individual attributes. As a result, it is critical to examine biases within each language and attribute. Of equal importance is to study how these biases compare across languages and how the biases are affected when training a model on multilingual data versus monolingual data. We present a bias analysis across Italian, Chinese, English, Hebrew, and Spanish on the downstream sentiment analysis task to observe whether specific demographics are viewed more positively. We study bias similarities and differences across these languages and investigate the impact of multilingual vs. monolingual training data. We adapt existing sentiment bias templates in English to Italian, Chinese, Hebrew, and Spanish for four attributes: race, religion, nationality, and gender. Our results reveal similarities in bias expression such as favoritism of groups that are dominant in each language's culture (e.g. majority religions and nationalities). Additionally, we find an increased variation in predictions across protected groups, indicating bias amplification, after multilingual finetuning in comparison to multilingual pretraining.

Narrative understanding involves capturing the author's cognitive processes, providing insights into their knowledge, intentions, beliefs, and desires. Although large language models (LLMs) excel in generating grammatically coherent text, their ability to comprehend the author's thoughts remains uncertain. This limitation hinders the practical applications of narrative understanding. In this paper, we conduct a comprehensive survey of narrative understanding tasks, thoroughly examining their key features, definitions, taxonomy, associated datasets, training objectives, evaluation metrics, and limitations. Furthermore, we explore the potential of expanding the capabilities of modularized LLMs to address novel narrative understanding tasks. By framing narrative understanding as the retrieval of the author's imaginative cues that outline the narrative structure, our study introduces a fresh perspective on enhancing narrative comprehension.

Conversational Recommender System (CRS) aims to explicitly acquire user preferences towards items and attributes through natural language conversations. However, existing CRS methods ask users to provide explicit answers (yes/no) for each attribute they require, regardless of users' knowledge or interest, which may significantly reduce the user experience and semantic consistency. Furthermore, these methods assume that users like all attributes of the target item and dislike those unrelated to it, which can introduce bias in attribute-level feedback and impede the system's ability to accurately identify the target item. To address these issues, we propose a more realistic, user-friendly, and explainable CRS framework called Hierarchical User-Interest Tracking for Conversational Recommender System (HutCRS). HutCRS portrays the conversation as a hierarchical interest tree that consists of two stages. In stage I, the system identifies the aspects that the user prefers while the system asks about attributes related to these positive aspects or recommends items in stage II. In addition, we develop a Hierarchical-Interest Policy Learning (HIPL) module to integrate the decision-making process of which aspects to ask and when to ask about attributes or recommend items. Moreover, we classify the attribute-level feedback results to further enhance the system’s ability to capture special information, such as attribute instances that are accepted by users but not presented in their historical interactive data. Extensive experiments on four benchmark datasets demonstrate the superiority of our method. The implementation of HutCRS is publicly available at https://github.com/xinle1129/HutCRS.

The tasks of out-of-domain (OOD) intent discovery and generalized intent discovery (GID) aim to extend a closed intent classifier to open-world intent sets, which is crucial to task-oriented dialogue (TOD) systems. Previous methods address them by fine-tuning discriminative models. Recently, although some studies has been exploring the application of large language models (LLMs) represented by ChatGPT to various downstream tasks, it is still unclear for the ability of ChatGPT to discover and incrementally extent OOD intents. In this paper, we comprehensively evaluate ChatGPT on OOD intent discovery and GID, and then outline the strengths and weaknesses of ChatGPT. Overall, ChatGPT exhibits consistent advantages under zero-shot settings, but is still at a disadvantage compared to fine-tuned models. More deeply, through a series of analytical experiments, we summarize and discuss the challenges faced by LLMs including clustering, domain-specific understanding, and cross-domain in-context learning scenarios. Finally, we provide empirical guidance for future directions to address these challenges.

Utterances do not occur in isolation in dialogues; it is essential to have the information of who the speaker of an utterance is to be able to recover the speaker’s intention with respect to the surrounding context. Beyond simply capturing speaker switches, identifying how speakers interact with each other in a dialogue is crucial to understanding conversational flow. This becomes increasingly important and simultaneously difficult to model when more than two interlocutors take part in a conversation. To overcome this challenge, we propose to explicitly add speaker awareness to each utterance representation. To that end, we use a graph neural network to model how each speaker is behaving within the local context of a conversation. The speaker representations learned this way are then used to update their respective utterance representations. We experiment with both multiparticipant and dyadic conversations on the MRDA and SwDA datasets and show the effectiveness of our approach.

Language models can be prompted to perform a wide variety of tasks with zero- and few-shot in-context learning. However, performance varies significantly with the choice of prompt, and we do not yet understand why this happens. In this paper, we analyze the factors that contribute to this variance and establish a new empirical hypothesis: the performance of a prompt is predicted by the extent to which the model is familiar with the language it contains. Over a wide range of tasks, we show that the lower the perplexity of the prompt, the better it is able to perform the task, when considering reasonable prompts that are related to it. As part of our analysis, we also devise a method to automatically extend a small seed set of manually written prompts by paraphrasing with GPT3 and backtranslation. This larger set allows us to verify that perplexity is a strong predictor of the success of a prompt and we show that the lowest perplexity prompts are consistently effective.