EMNLP 2023 论文列表

Conventional speech-to-text translation (ST) systems are trained on single-speaker utterances, and they may not generalize to real-life scenarios where the audio contains conversations by multiple speakers. In this paper, we tackle single-channel multi-speaker conversational ST with an end-to-end and multi-task training model, named Speaker-Turn Aware Conversational Speech Translation, that combines automatic speech recognition, speech translation and speaker turn detection using special tokens in a serialized labeling format. We run experiments on the Fisher-CALLHOME corpus, which we adapted by merging the two single-speaker channels into one multi-speaker channel, thus representing the more realistic and challenging scenario with multi-speaker turns and cross-talk. Experimental results across single- and multi-speaker conditions and against conventional ST systems, show that our model outperforms the reference systems on the multi-speaker condition, while attaining comparable performance on the single-speaker condition. We release scripts for data processing and model training.

Countering rather than censoring hate speech has emerged as a promising strategy to address hatred. There are many types of counterspeech in user-generated content: addressing the hateful content or its author, generic requests, well-reasoned counter arguments, insults, etc. The effectiveness of counterspeech, which we define as subsequent incivility, depends on these types. In this paper, we present a theoretically grounded taxonomy of replies to hate speech and a new corpus. We work with real, user-generated hate speech and all the replies it elicits rather than replies generated by a third party. Our analyses provide insights into the content real users reply with as well as which replies are empirically most effective. We also experiment with models to characterize the replies to hate speech, thereby opening the door to estimating whether a reply to hate speech will result in further incivility.

Semi-supervised text classification (SSTC) has gained increasing attention due to its ability to leverage unlabeled data. However, existing approaches based on pseudo-labeling suffer from the issues of pseudo-label bias and error accumulation. In this paper, we propose JointMatch, a holistic approach for SSTC that addresses these challenges by unifying ideas from recent semi-supervised learning and the task of learning with noise. JointMatch adaptively adjusts classwise thresholds based on the learning status of different classes to mitigate model bias towards current easy classes. Additionally, JointMatch alleviates error accumulation by utilizing two differently initialized networks to teach each other in a cross-labeling manner. To maintain divergence between the two networks for mutual learning, we introduce a strategy that weighs more disagreement data while also allowing the utilization of high-quality agreement data for training. Experimental results on benchmark datasets demonstrate the superior performance of JointMatch, achieving a significant 5.13\% improvement on average. Notably, JointMatch delivers impressive results even in the extremely-scarce-label setting, obtaining 86\% accuracy on AG News with only 5 labels per class. We make our code available at https://github.com/HenryPengZou/JointMatch.

User-generated social media data is constantly changing as new trends influence online discussion and personal information is deleted due to privacy concerns. However, traditional NLP models rely on fixed training datasets, which means they are unable to adapt to temporal change---both test distribution shift and deleted training data---without frequent, costly re-training. In this paper, we study temporal adaptation through the task of longitudinal hashtag prediction and propose a non-parametric dense retrieval technique, which does not require re-training, as a simple but effective solution. In experiments on a newly collected, publicly available, year-long Twitter dataset exhibiting temporal distribution shift, our method improves by 64% over the best static parametric baseline while avoiding costly gradient-based re-training. Our approach is also particularly well-suited to dynamically deleted user data in line with data privacy laws, with negligible computational cost/performance loss.

We present PeerSum, a novel dataset for generating meta-reviews of scientific papers. The meta-reviews can be interpreted as abstractive summaries of reviews, multi-turn discussions and the paper abstract. These source documents have a rich inter-document relationship with an explicit hierarchical conversational structure, cross-references and (occasionally) conflicting information. To introduce the structural inductive bias into pre-trained language models, we introduce RAMMER (Relationship-aware Multi-task Meta-review Generator), a model that uses sparse attention based on the conversational structure and a multi-task training objective that predicts metadata features (e.g., review ratings). Our experimental results show that RAMMER outperforms other strong baseline models in terms of a suite of automatic evaluation metrics. Further analyses, however, reveal that RAMMER and other models struggle to handle conflicts in source documents, suggesting meta-review generation is a challenging task and a promising avenue for further research.

In this work, we carry out a data archaeology to infer books that are known to ChatGPT and GPT-4 using a name cloze membership inference query. We find that OpenAI models have memorized a wide collection of copyrighted materials, and that the degree of memorization is tied to the frequency with which passages of those books appear on the web. The ability of these models to memorize an unknown set of books complicates assessments of measurement validity for cultural analytics by contaminating test data; we show that models perform much better on memorized books than on non-memorized books for downstream tasks. We argue that this supports a case for open models whose training data is known.

We study whether linguistic information in pre-trained multilingual language models can be accessed by human language: So far, there is no easy method to directly obtain linguistic information and gain insights into the linguistic principles encoded in such models. We use the technique of prompting and formulate linguistic tasks to test the LM’s access to explicit grammatical principles and study how effective this method is at providing access to linguistic features. Our experiments on German, Icelandic and Spanish show that some linguistic properties can in fact be accessed through prompting, whereas others are harder to capture.

Science progresses by building upon the prior body of knowledge documented in scientific publications. The acceleration of research makes it hard to stay up-to-date with the recent developments and to summarize the ever-growing body of prior work. To address this, the task of citation text generation aims to produce accurate textual summaries given a set of papers-to-cite and the citing paper context. Due to otherwise rare explicit anchoring of cited documents in the citing paper, citation text generation provides an excellent opportunity to study how humans aggregate and synthesize textual knowledge from sources. Yet, existing studies are based upon widely diverging task definitions, which makes it hard to study this task systematically. To address this challenge, we propose CiteBench: a benchmark for citation text generation that unifies multiple diverse datasets and enables standardized evaluation of citation text generation models across task designs and domains. Using the new benchmark, we investigate the performance of multiple strong baselines, test their transferability between the datasets, and deliver new insights into the task definition and evaluation to guide future research in citation text generation. We make the code for CiteBench publicly available at https://github.com/UKPLab/citebench.

Vision-language models (VLMs) have shown remarkable performance on visual reasoning tasks (e.g. attributes, location). While such tasks measure the requisite knowledge to ground and reason over a given visual instance, they do not, however, measure the ability of VLMs to retain and generalize such knowledge. In this work, we evaluate VLMs' ability to acquire "visible" physical knowledge -- the information that is easily accessible from images of static scenes, particularly along the dimensions of object color, size, and space. We build an automatic pipeline to derive a comprehensive knowledge resource for calibrating and probing these models. Our results indicate a severe gap between model and human performance across all three dimensions. Furthermore, we demonstrate that a caption pretrained LM significantly outperforms VLMs on both size and spatial tasks -- highlighting that despite sufficient access to ground language with visual modality, they struggle to retain such knowledge.

Large language models (LLMs) can generate natural language texts for various domains and tasks, but their potential for clinical text mining, a domain with scarce, sensitive, and imbalanced medical data, is under-explored. We investigate whether LLMs can augment clinical data for detecting Alzheimer's Disease (AD)-related signs and symptoms from electronic health records (EHRs), a challenging task that requires high expertise. We create a novel pragmatic taxonomy for AD sign and symptom progression based on expert knowledge and generated three datasets: (1) a gold dataset annotated by human experts on longitudinal EHRs of AD patients; (2) a silver dataset created by the data-to-label method, which labels sentences from a public EHR collection with AD-related signs and symptoms; and (3) a bronze dataset created by the label-to-data method which generates sentences with AD-related signs and symptoms based on the label definition. We train a system to detect AD-related signs and symptoms from EHRs. We find that the silver and bronze datasets improves the system performance, outperforming the system using only the gold dataset. This shows that LLMs can generate synthetic clinical data for a complex task by incorporating expert knowledge, and our label-to-data method can produce datasets that are free of sensitive information, while maintaining acceptable quality.

Pre-trained language models (PLMs) have shown impressive performance in various language tasks. However, they are prone to spurious correlations, and often generate illusory information. In real-world applications, PLMs should justify decisions with formalized, coherent reasoning chains, but this challenge remains under-explored. Cognitive psychology theorizes that humans are capable of utilizing fast and intuitive *heuristic* thinking to make decisions based on past experience, then rationalizing the decisions through slower and deliberative *analytic* reasoning. We incorporate these interlinked dual processes in fine-tuning and in-context learning with PLMs, applying them to two language understanding tasks that require coherent physical commonsense reasoning. We show that our proposed Heuristic-Analytic Reasoning (HAR) strategies drastically improve the coherence of rationalizations for model decisions, yielding state-of-the-art results on Tiered Reasoning for Intuitive Physics (TRIP). We also find that this improved coherence is a direct result of more faithful attention to relevant language context in each step of reasoning. Our findings suggest that human-like reasoning strategies can effectively improve the coherence and reliability of PLM reasoning.

This paper proposes a practical multimodal video summarization task setting and a dataset to train and evaluate the task. The target task involves summarizing a given video into a predefined number of keyframe-caption pairs and displaying them in a listable format to grasp the video content quickly. This task aims to extract crucial scenes from the video in the form of images (keyframes) and generate corresponding captions explaining each keyframe's situation. This task is useful as a practical application and presents a highly challenging problem worthy of study. Specifically, achieving simultaneous optimization of the keyframe selection performance and caption quality necessitates careful consideration of the mutual dependence on both preceding and subsequent keyframes and captions. To facilitate subsequent research in this field, we also construct a dataset by expanding upon existing datasets and propose an evaluation framework. Furthermore, we develop two baseline systems and report their respective performance.

Stylized dialogue generation systems aim to produce coherent and context-aware dialogues while effectively emulating the desired style. Generating stylized dialogue is valuable yet challenging due to the scarce parallel data. Existing methods often synthesize pseudo data through back translation, yet suffer from noisy and context-agnostic style signals caused by insufficient guidance on target style features. To address this, we propose the knowledge-augmented stylized dialogue generation model, which includes a feature-guided style knowledge selection module that utilizes context and response features. Specifically, we retrieve dialogue-related style sentences from style corpus to explicitly provide clear style signals. We design a feature-guided selection module with response-related contrastive learning and style responsiveness Kullback-Leibler losses to enhance generation at both semantic and stylized levels. Our approach demonstrates satisfactory performance on two public stylized dialogue benchmarks in both automatic and human evaluations.

Language models may memorize more than just facts, including entire chunks of texts seen during training. Fair use exemptions to copyright laws typically allow for limited use of copyrighted material without permission from the copyright holder, but typically for extraction of information from copyrighted materials, rather than {\em verbatim} reproduction. This work explores the issue of copyright violations and large language models through the lens of verbatim memorization, focusing on possible redistribution of copyrighted text. We present experiments with a range of language models over a collection of popular books and coding problems, providing a conservative characterization of the extent to which language models can redistribute these materials. Overall, this research highlights the need for further examination and the potential impact on future developments in natural language processing to ensure adherence to copyright regulations. Code is at https://github.com/coastalcph/CopyrightLLMs.

Language modeling is a fundamental task in natural language processing, which has been thoroughly explored with various architectures and hyperparameters. However, few studies focus on the effect of sub-word segmentation on the performance of language models (LMs). In this paper, we compare GPT and BERT models trained with the statistical segmentation algorithm BPE vs. two unsupervised algorithms for morphological segmentation — Morfessor and StateMorph. We train the models for several languages — including ones with very rich morphology — and compare their performance with different segmentation algorithms, vocabulary sizes, and model sizes. The results show that training with morphological segmentation allows the LMs to: (1) achieve lower perplexity, (2) converge more efficiently in terms of training time, and (3) achieve equivalent or better evaluation scores on downstream tasks. Lastly, we show that (4) LMs of smaller size using morphological segmentation can perform comparably to models of larger size trained with BPE — both in terms of (1) perplexity and (3) scores on downstream tasks. Points (2) and (4) impact on sustainability, since they reduce the model cost; and while 2 reduces cost only in the training phase, 4 does so also in the inference phase.

Large language models (LLMs) excel at processing and generating text and code. However, LLMs have had limited applicability in grounded task-oriented dialogue as they are difficult to steer toward task objectives and fail to handle novel grounding. We present a modular and interpretable grounded dialogue system that addresses these shortcomings by composing LLMs with a symbolic planner and grounded code execution. Our system, consists of a reader and planner: the reader leverages an LLM to convert partner utterances into executable code, calling functions that perform grounding. The translated code's output is stored to track dialogue state, while a symbolic planner determines the next appropriate response. We evaluate our system's performance on the demanding OneCommon dialogue task, involving collaborative reference resolution on abstract images of scattered dots. Our system substantially outperforms the previous state-of-the-art, including improving task success in human evaluations from 56\% to 69\% in the most challenging setting.

Large Language Models (LLMs) exhibit impressive performance on a range of NLP tasks, due to the general-purpose linguistic knowledge acquired during pretraining. Existing model interpretability research (Tenney et al., 2019) suggests that a linguistic hierarchy emerges in the LLM layers, with lower layers better suited to solving syntactic tasks and higher layers employed for semantic processing. Yet, little is known about how encodings of different linguistic phenomena interact within the models and to what extent processing of linguistically-related categories relies on the same, shared model representations. In this paper, we propose a framework for testing the joint encoding of linguistic categories in LLMs. Focusing on syntax, we find evidence of joint encoding both at the same (related part-of-speech (POS) classes) and different (POS classes and related syntactic dependency relations) levels of linguistic hierarchy. Our cross-lingual experiments show that the same patterns hold across languages in multilingual LLMs.

A hallmark of modern large language models (LLMs) is their impressive general zero-shot and few-shot abilities, often elicited through in-context learning (ICL) via prompting. However, while highly coveted and being the most general, zero-shot performances in LLMs are still typically weaker due to the lack of guidance and the difficulty of applying existing automatic prompt design methods in general tasks when ground-truth labels are unavailable. In this study, we address this by presenting Universal Self-Adaptive Prompting (USP), an automatic prompt design approach specifically tailored for zero-shot learning (while compatible with few-shot). Requiring only a small amount of unlabeled data and an inference-only LLM, USP is highly versatile: to achieve universal prompting, USP categorizes a possible NLP task into one of the three possible task types and then uses a corresponding selector to select the most suitable queries and zero-shot model-generated responses as pseudo-demonstrations, thereby generalizing ICL to the zero-shot setup in a fully automated way. We evaluate USP with PaLM and PaLM 2 models and demonstrate performances that are considerably stronger than standard zero-shot baselines and often comparable to or even superior to few-shot baselines across more than 40 natural language understanding, natural language generation, and reasoning tasks.

Despite the growing use of the Somali language in various online domains, research on Somali language information retrieval remains limited and primarily relies on query translation due to the lack of a dedicated corpus. To address this problem, we collaborated with language experts and natural language processing (NLP) researchers to create an annotated corpus for Somali information retrieval. This corpus comprises 2335 documents collected from various well-known online sites, such as hiiraan online, dhacdo net, and Somali poetry books. We explain how the corpus was constructed, and develop a Somali language information retrieval system using a pseudo-relevance feedback (PRF) query expansion technique on the corpus. Note that collecting such a data set for the low-resourced Somali language can help overcome NLP barriers, such as the lack of electronically available data sets. Which, if available, can enable the development of various NLP tools and applications such as question-answering and text classification. It also provides researchers with a valuable resource for investigating and developing new techniques and approaches for Somali.