EMNLP 2023 Papers

In the aftermath of GPT-3.5, commonly known as ChatGPT, research have attempted to assess its capacity for lowering annotation cost, either by doing zero-shot learning, generating new data, or replacing human annotators. Some studies have also investigated its use for data augmentation (DA), but only in limited contexts, which still leaves the question of how ChatGPT performs compared to state-of-the-art algorithms. In this paper, we use ChatGPT to create new data both with paraphrasing and with zero-shot generation, and compare it to seven other algorithms. We show that while ChatGPT performs exceptionally well on some simpler data, it overall does not perform better than the other algorithms, yet demands a much larger implication from the practitioner due to the ChatGPT often refusing to answer due to sensitive content in the datasets.

Recent years have seen remarkable advances in the field of Simultaneous Machine Translation (SiMT) due to the introduction of innovative policies that dictate whether to READ or WRITE at each step of the translation process. However, a common assumption in many existing studies is that operations are carried out at the subword level, even though the standard unit for input and output in most practical scenarios is typically at the word level. This paper demonstrates that policies devised and validated at the subword level are surpassed by those operating at the word level, which process multiple subwords to form a complete word in a single step. Additionally, we suggest a method to boost SiMT models using language models (LMs), wherein the proposed word-level policy plays a vital role in addressing the subword disparity between LMs and SiMT models. Code is available at https://github.com/xl8-ai/WordSiMT.

Few-shot named entity recognition (NER) systems aim to recognize new classes of entities with limited labeled samples. However, these systems face a significant challenge of overfitting compared to tasks with abundant samples. This overfitting is mainly caused by the spurious correlation resulting from the bias in selecting a few samples. To address this issue, we propose a causal intervention-based few-shot NER method in this paper. Our method, based on the prototypical network, intervenes in the context to block the backdoor path between context and label. In the one-shot scenario, where no additional context is available for intervention, we employ incremental learning to intervene on the prototype, which also helps mitigate catastrophic forgetting. Our experiments on various benchmarks demonstrate that our approach achieves new state-of-the-art results.

News recommendation is one of the widest commercialization in natural language processing research area, which aims to recommend news according to user interests. New recall plays an important role in news recommendation. It is to recall candidates from a very large news database. Recent researches of news recall mostly adopt dual-encoder architecture as it provides a much faster recall scheme, and they encode each word equally. However, these works remain two challenges: irrelevant word distraction and weak dual-encoder interaction. Therefore, we propose a model Topic-aware Attention and powerful Dual-encoder Interaction for Recall in news recommendation (TADI). To avoid irrelevant word distraction, TADI designs a Topic-aware Attention (TA) which weights words according to news topics. To enhance dual-encoder interaction, TADI provides a cheap yet powerful interaction module, namely Dual-encoder Interaction (DI). DI helps dual encoders interact powerfully based on two aux targets. After performance comparisons between TADI and state-of-the-arts in a series of experiments, we verify the effectiveness of TADI.

The information stored in large language models (LLMs) falls out of date quickly, and retraining from scratch is often not an option. This has recently given rise to a range of techniques for injecting new facts through updating model weights. Current evaluation paradigms are extremely limited, mainly validating the recall of edited facts, but changing one fact should cause rippling changes to the model's related beliefs. If we edit the UK Prime Minister to now be Rishi Sunak, then we should get a different answer to Who is married to the British Prime Minister? In this work, we present a benchmark MQuAKE (Multi-hop Question Answering for Knowledge Editing) comprising multi-hop questions that assess whether edited models correctly answer questions where the answer should change as an entailed consequence of edited facts. While we find that current knowledge-editing approaches can recall edited facts accurately, they fail catastrophically on the constructed multi-hop questions. We thus propose a simple memory-based approach, MeLLo, which stores all edited facts externally while prompting the language model iteratively to generate answers that are consistent with the edited facts. While MQuAKE remains challenging, we show that MeLLo scales well with LLMs (up to 175B) and outperforms previous model editors by a large margin.

Previous research on argumentation in online discussions has largely focused on examining individual comments and neglected the interactive nature of discussions. In line with previous work, we represent individual comments as sequences of semantic argumentative unit types. However, because it is intuitively necessary for dialogical argumentation to address the opposing viewpoints, we extend this model by clustering type sequences into different argument arrangement patterns and representing discussions as sequences of these patterns. These sequences of patterns are a symbolic representation of argumentation strategies that capture the overall structure of discussions. Using this novel approach, we conduct an in-depth analysis of the strategies in 34,393 discussions from the online discussion forum Change My View and show that our discussion model is effective for persuasiveness prediction, outperforming LLM-based classifiers on the same data. Our results provide valuable insights into argumentation dynamics in online discussions and, through the presented prediction procedure, are of practical importance for writing assistance and persuasive text generation systems.

Identifying users' stances regarding specific targets/topics is a significant route to learning public opinion from social media platforms. Most existing studies of stance detection strive to learn stance information about specific targets from the context, in order to determine the user's stance on the target. However, in real-world scenarios, we usually have a certain understanding of a target when we express our stance on it. In this paper, we investigate stance detection from a novel perspective, where the background knowledge of the targets is taken into account for better stance detection. To be specific, we categorize background knowledge into two categories: episodic knowledge and discourse knowledge, and propose a novel Knowledge-Augmented Stance Detection (KASD) framework. For episodic knowledge, we devise a heuristic retrieval algorithm based on the topic to retrieve the Wikipedia documents relevant to the sample. Further, we construct a prompt for ChatGPT to filter the Wikipedia documents to derive episodic knowledge. For discourse knowledge, we construct a prompt for ChatGPT to paraphrase the hashtags, references, etc., in the sample, thereby injecting discourse knowledge into the sample. Experimental results on four benchmark datasets demonstrate that our KASD achieves state-of-the-art performance in in-target and zero-shot stance detection.

Evaluating open-domain dialogue systems is currently an open question. Automatic evaluation metrics have shown poor correlation with human assessment in dialogue generation tasks. Human evaluation, which involves annotators for multi-dimension scoring, is trustworthy but time-consuming. In this work, we propose FFAEval, a reliable and efficient human evaluation framework using Free-For-All ranking approach. By sharing the dialogue history, the framework enables annotators to converse with multiple dialogue systems simultaneously in a single-blind, multi-turn manner. The subsequent free-for-all allows annotators to select the most favourable model in each turn from among all the participating dialogue systems. The final performance of each model is represented by calculating the TrueSkill score derived from the free-for-all competition. Our empirical study on English and Chinese dialogue systems demonstrates that FFAEval achieves a strong correlation with score-based human assessment compared to existing evaluation methods. We further prove the efficiency and stability of our framework in additional experiments. The source code and data are available on Github.

The conversational machine reading comprehension (CMRC) task aims to answer questions in conversations, which has been a hot research topic in recent years because of its wide applications. However, existing CMRC benchmarks in which each conversation is assigned a static passage are inconsistent with real scenarios. Thus, model's comprehension ability towards real scenarios are hard to evaluate reasonably. To this end, we propose the first Chinese CMRC benchmark \textbf{Orca} and further provide zero-shot/few-shot settings to evaluate model's generalization ability towards diverse domains. We collect 831 hot-topic driven conversations with 4,742 turns in total. Each turn of a conversation is assigned with a response-related passage, aiming to evaluate model's comprehension ability more reasonably. The topics of conversations are collected from social media platform and cover 33 domains, trying to be consistent with real scenarios. Importantly, answers in Orca are all well-annotated natural responses rather than the specific spans or short phrase in previous datasets. Besides, we implement three strong baselines to tackle the challenge in Orca. The results indicate the great challenge of our CMRC benchmark.

Entities and relationships between entities are vital in the real world. Essentially, we understand the world by understanding entities and relations. For instance, to understand a field, e.g., computer science, we need to understand the relevant concepts, e.g., machine learning, and the relationships between concepts, e.g., machine learning and artificial intelligence. To understand a person, we should first know who he/she is and how he/she is related to others. To understand entities and relations, humans may refer to natural language descriptions. For instance, when learning a new scientific term, people usually start by reading its definition in dictionaries or encyclopedias. To know the relationship between two entities, humans tend to create a sentence to connect them. In this paper, we propose VER: a unified model for Verbalizing Entities and Relations. Specifically, we attempt to build a system that takes any entity or entity set as input and generates a sentence to represent entities and relations. Extensive experiments demonstrate that our model can generate high-quality sentences describing entities and entity relationships and facilitate various tasks on entities and relations, including definition modeling, relation modeling, and generative commonsense reasoning.

Extracting meaningful entities belonging to predefined categories from Visually-rich Form-like Documents (VFDs) is a challenging task. Visual and layout features such as font, background, color, and bounding box location and size provide important cues for identifying entities of the same type. However, existing models commonly train a visual encoder with weak cross-modal supervision signals, resulting in a limited capacity to capture these non-textual features and suboptimal performance. In this paper, we propose a novel Visually-Asymmetric coNsistenCy Learning (VANCL) approach that addresses the above limitation by enhancing the model's ability to capture fine-grained visual and layout features through the incorporation of color priors. Experimental results on benchmark datasets show that our approach substantially outperforms the strong LayoutLM series baseline, demonstrating the effectiveness of our approach. Additionally, we investigate the effects of different color schemes on our approach, providing insights for optimizing model performance. We believe our work will inspire future research on multimodal information extraction.

Social norms fundamentally shape interpersonal communication. We present NormDial, a high-quality dyadic dialogue dataset with turn-by-turn annotations of social norm adherences and violations for Chinese and American cultures. Introducing the task of social norm observance detection, our dataset is synthetically generated in both Chinese and English using a human-in-the-loop pipeline by prompting large language models with a small collection of expert-annotated social norms. We show that our generated dialogues are of high quality through human evaluation and further evaluate the performance of existing large language models on this task. Our findings point towards new directions for understanding the nuances of social norms as they manifest in conversational contexts that span across languages and cultures.

Public and private actors struggle to assess the vast amounts of information about sustainability commitments made by various institutions. To address this problem, we create a novel tool for automatically detecting corporate and national net zero and reduction targets in three steps. First, we introduce an expert-annotated data set with 3.5K text samples. Second, we train and release ClimateBERT-NetZero, a natural language classifier to detect whether a text contains a net zero or reduction target. Third, we showcase its analysis potential with two use cases: We first demonstrate how ClimateBERT-NetZero can be combined with conventional question-answering (Q\&A) models to analyze the ambitions displayed in net zero and reduction targets. Furthermore, we employ the ClimateBERT-NetZero model on quarterly earning call transcripts and outline how communication patterns evolve over time. Our experiments demonstrate promising pathways for extracting and analyzing net zero and emission reduction targets at scale.

In psycholinguistics, surprisal theory posits that the amount of online processing effort expended by a human comprehender per word positively correlates with the surprisal of that word given its preceding context. In addition to this overall correlation, more importantly, the specific quantitative form taken by the processing effort as a function of surprisal offers insights into the underlying cognitive mechanisms of language processing. Focusing on English, previous studies have looked into the linearity of surprisal on reading times. Here, we extend the investigation by examining eyetracking corpora of seven languages: Danish, Dutch, English, German, Japanese, Mandarin, and Russian. We find evidence for superlinearity in some languages, but the results are highly sensitive to which language model is used to estimate surprisal.

Computational inefficiency in transformers has been a long-standing challenge, hindering the deployment in resource-constrained or real-time applications. One promising approach to mitigate this limitation is to progressively remove less significant tokens, given that the sequence length strongly contributes to the inefficiency. However, this approach entails a potential risk of losing crucial information due to the irrevocable nature of token removal. In this paper, we introduce Leap-of-Thought (LoT), a novel token reduction approach that dynamically routes tokens within layers. Unlike previous work that irrevocably discards tokens, LoT enables tokens to `leap' across layers. This ensures that all tokens remain accessible in subsequent layers while reducing the number of tokens processed within layers. We achieve this by pairing the transformer with dynamic token routers, which learn to selectively process tokens essential for the task. Evaluation results clearly show that LoT achieves a substantial improvement in computational efficiency. Specifically, LoT attains up to 25x faster inference time without a significant loss in accuracy

Recent research has shown that evaluating the robustness of natural language processing models using textual attack methods is significant. However, most existing text attack methods only use heuristic replacement strategies or language models to generate replacement words at the word level. The blind pursuit of high attack success rates makes it difficult to ensure the quality of the generated adversarial text. As a result, adversarial text is often difficult for humans to understand. In fact, many methods that perform well in terms of text attacks often generate adversarial text with poor quality. To address this important gap, our work treats black-box text attack as an unsupervised text generation problem and proposes a search and learning framework for Adversarial Text Generation by Search and Learning (ATGSL) and develops three adversarial attack methods (ATGSL-SA, ATGSL-BM, ATGSL-FUSION) for black box text attacks. We first apply a heuristic search attack algorithm (ATGSL-SA) and a linguistic thesaurus to generate adversarial samples with high semantic similarity. After this process, we train a conditional generative model to learn from the search results while smoothing out search noise. Moreover, we design an efficient ATGSL-BM attack algorithm based on the text generator. Furthermore, we propose a hybrid attack method (ATGSL-FUSION) that integrates the advantages of ATGSL-SA and ATGSL-BM to enhance attack effectiveness. Our proposed attack algorithms are significantly superior to the most advanced methods in terms of attack efficiency and adversarial text quality.

Query-focused Summarization (QfS) deals with systems that generate summaries from document(s) based on a query. Motivated by the insight that Reinforcement Learning (RL) provides a generalization to Supervised Learning (SL) for Natural Language Generation, and thereby performs better (empirically) than SL, we use an RL-based approach for this task of QfS. Additionally, we also resolve the conflict of employing RL in Transformers with Teacher Forcing. We develop multiple Policy Gradient networks, trained on various reward signals: ROUGE, BLEU, and Semantic Similarity, which lead to a $\mathit{10}$-point improvement over the State-of-the-Art approach on the ROUGE-L metric for a benchmark dataset (ELI5). We also show performance of our approach in zero-shot setting for another benchmark dataset (DebatePedia) -- our approach leads to results comparable to baselines, which were specifically trained on DebatePedia. To aid the RL training, we propose a better semantic similarity reward, enabled by a novel Passage Embedding scheme developed using Cluster Hypothesis. Lastly, we contribute a gold-standard test dataset to further research in QfS and Long-form Question Answering (LfQA).

In-context learning (ICL) is a new learning paradigm that has gained popularity along with the development of large language models. In this work, we adapt a recently proposed hardness metric, pointwise $\mathcal{V}$-usable information (PVI), to an in-context version (in-context PVI). Compared to the original PVI, in-context PVI is more efficient in that it requires only a few exemplars and does not require fine-tuning. We conducted a comprehensive empirical analysis to evaluate the reliability of in-context PVI. Our findings indicate that in-context PVI estimates exhibit similar characteristics to the original PVI. Specific to the in-context setting, we show that in-context PVI estimates remain consistent across different exemplar selections and numbers of shots. The variance of in-context PVI estimates across different exemplar selections is insignificant, which suggests that in-context PVI estimates are stable. Furthermore, we demonstrate how in-context PVI can be employed to identify challenging instances. Our work highlights the potential of in-context PVI and provides new insights into the capabilities of ICL.

Group fairness is a central research topic in text classification, where reaching fair treatment between sensitive groups (e.g. women vs. men) remains an open challenge. This paper presents a novel method for mitigating biases in neural text classification, agnostic to the model architecture. Considering the difficulty to distinguish fair from unfair information in a text encoder, we take inspiration from adversarial training to induce Wasserstein independence between representations learned to predict our target label and the ones learned to predict some sensitive attribute. Our approach provides two significant advantages. Firstly, it does not require annotations of sensitive attributes in both testing and training data. This is more suitable for real-life scenarios compared to existing methods that require annotations of sensitive attributes at train time. Secondly, our approach exhibits a comparable or better fairness-accuracy trade-off compared to existing methods.