EMNLP 2023 论文列表

In recent years, there has been a rapid proliferation of AI-generated text, primarily driven by the release of powerful pre-trained language models (PLMs). To address the issue of misuse associated with AI-generated text, various high-performing detectors have been developed, including the OpenAI detector and the Stanford DetectGPT. In our study, we ask how reliable these detectors are. We answer the question by designing a novel approach that can prompt any PLM to generate text that evades these high-performing detectors. The proposed approach suggests a universal evasive prompt, a novel type of soft prompt, which guides PLMs in producing "human-like" text that can mislead the detectors. The novel universal evasive prompt is achieved in two steps: First, we create an evasive soft prompt tailored to a specific PLM through prompt tuning; and then, we leverage the transferability of soft prompts to transfer the learned evasive soft prompt from one PLM to another. Employing multiple PLMs in various writing tasks, we conduct extensive experiments to evaluate the efficacy of the evasive soft prompts in their evasion of state-of-the-art detectors.

Human experts write summaries using different techniques, including extracting a sentence from the document and rewriting it, or fusing various information from the document to abstract it. These techniques are flexible and thus difficult to be imitated by any single method. To address this issue, we propose an adaptive model, GEMINI, that integrates a rewriter and a generator to mimic the sentence rewriting and abstracting techniques, respectively. GEMINI adaptively chooses to rewrite a specific document sentence or generate a summary sentence from scratch. Experiments demonstrate that our adaptive approach outperforms the pure abstractive and rewriting baselines on three benchmark datasets, achieving the best results on WikiHow. Interestingly, empirical results show that the human summary styles of summary sentences are consistently predictable given their context. We release our code and model at \url{https://github.com/baoguangsheng/gemini}.

Research on neural networks has focused on understanding a single model trained on a single dataset. However, relatively little is known about the relationships between different models, particularly those trained or tested on different datasets. We address this by studying how the weight space and the underlying loss landscape of different models are interconnected. Specifically, we demonstrate that finetuned models that were optimized for high performance, reside in well-defined regions in weight space, and vice versa -- that any model that resides anywhere in those regions also exhibits high performance. Notably, we show that language models that have been finetuned on the same dataset form a tight cluster in the weight space, while models finetuned on different datasets from the same underlying task form a looser cluster. Moreover, traversing around the region between the models leads to new models that perform comparably or even better than models obtained via finetuning, even on tasks that the original models were not finetuned on. Our findings provide insight into the relationships between models, demonstrating that a model positioned between two similar models can acquire the knowledge of both. We leverage this and design a method for selecting a better model for efficient finetuning. Specifically, we show that starting from the center of the region is as effective, if not more, than using the pretrained model in 11 out of 12 datasets, resulting in an average accuracy improvement of 3.06.

The NLP community has long advocated for the construction of multi-annotator datasets to better capture the nuances of language interpretation, subjectivity, and ambiguity. This paper conducts a retrospective study to show how performance scores can vary when a dataset expands from a single annotation per instance to multiple annotations. We propose a novel multi-annotator simulation process to generate datasets with varying annotation budgets. We show that similar datasets with the same annotation budget can lead to varying performance gains. Our findings challenge the popular belief that models trained on multi-annotation examples always lead to better performance than models trained on single or few-annotation examples.

In this paper, we address the hallucination problem commonly found in natural language generation tasks. Language models often generate fluent and convincing content but can lack consistency with the provided source, resulting in potential inaccuracies. We propose a new decoding method called Fidelity-Enriched Contrastive Search (FECS), which augments the contrastive search framework with context-aware regularization terms. FECS promotes tokens that are semantically similar to the provided source while penalizing repetitiveness in the generated text. We demonstrate its effectiveness across two tasks prone to hallucination: abstractive summarization and dialogue generation. Results show that FECS consistently enhances faithfulness across various language model sizes while maintaining output diversity comparable to well-performing decoding algorithms.

We investigate the potential misuse of modern Large Language Models (LLMs) for generating credible-sounding misinformation and its subsequent impact on information-intensive applications, particularly Open-Domain Question Answering (ODQA) systems. We establish a threat model and simulate potential misuse scenarios, both unintentional and intentional, to assess the extent to which LLMs can be utilized to produce misinformation. Our study reveals that LLMs can act as effective misinformation generators, leading to a significant degradation (up to 87%) in the performance of ODQA systems. Moreover, we uncover disparities in the attributes associated with persuading humans and machines, presenting an obstacle to current human-centric approaches to combat misinformation. To mitigate the harm caused by LLM-generated misinformation, we propose three defense strategies: misinformation detection, vigilant prompting, and reader ensemble. These approaches have demonstrated promising results, albeit with certain associated costs. Lastly, we discuss the practicality of utilizing LLMs as automatic misinformation generators and provide relevant resources and code to facilitate future research in this area.

We present Dolphin, a novel benchmark that addresses the need for a natural language generation (NLG) evaluation framework dedicated to the wide collection of Arabic languages and varieties. The proposed benchmark encompasses a broad range of 13 different NLG tasks, including dialogue generation, question answering, machine translation, summarization, among others. Dolphin comprises a substantial corpus of 40 diverse and representative public datasets across 50 test splits, carefully curated to reflect real-world scenarios and the linguistic richness of Arabic. It sets a new standard for evaluating the performance and generalization capabilities of Arabic and multilingual models, promising to enable researchers to push the boundaries of current methodologies. We provide an extensive analysis of Dolphin, highlighting its diversity and identifying gaps in current Arabic NLG research. We also offer a public leaderboard that is both interactive and modular and evaluate several Arabic and multilingual models on our benchmark, allowing us to set strong baselines against which researchers can compare.

Aspect-Based Argument Mining (ABAM) is a critical task in computational argumentation. Existing methods have primarily treated ABAM as a nested named entity recognition problem, overlooking the need for tailored strategies to effectively address the specific challenges of ABAM tasks. To this end, we propose a layer-based Hierarchical Enhancement Framework (HEF) for ABAM, and introduce three novel components: the Semantic and Syntactic Fusion (SSF) component, the Batch-level Heterogeneous Graph Attention Network (BHGAT) component, and the Span Mask Interactive Attention (SMIA) component. These components serve the purposes of optimizing underlying representations, detecting argument unit stances, and constraining aspect term recognition boundaries, respectively. By incorporating these components, our framework enables better handling of the challenges and improves the performance and accuracy in argument unit and aspect term recognition. Experiments on multiple datasets and various tasks verify the effectiveness of the proposed framework and components.

Toxic language, such as hate speech, can deter users from participating in online communities and enjoying popular platforms. Previous approaches to detecting toxic language and norm violations have been primarily concerned with conversations from online forums and social media, such as Reddit and Twitter. These approaches are less effective when applied to conversations on live-streaming platforms, such as Twitch and YouTube Live, as each comment is only visible for a limited time and lacks a thread structure that establishes its relationship with other comments. In this work, we share the first NLP study dedicated to detecting norm violations in conversations on live-streaming platforms. We define norm violation categories in live-stream chats and annotate 4,583 moderated comments from Twitch. We articulate several facets of live-stream data that differ from other forums, and demonstrate that existing models perform poorly in this setting. By conducting a user study, we identify the informational context humans use in live-stream moderation, and train models leveraging context to identify norm violations. Our results show that appropriate contextual information can boost moderation performance by 35\%.

Contrastively trained vision-language models have achieved remarkable progress in vision and language representation learning. However, recent research has highlighted severe limitations of these models in their ability to perform compositional reasoning over objects, attributes, and relations. Scene graphs have emerged as an effective way to understand images compositionally. These are graph-structured semantic representations of images that contain objects, their attributes, and relations with other objects in a scene. In this work, we consider the scene graph parsed from text as a proxy for the image scene graph and propose a graph decomposition and augmentation framework along with a coarse-to-fine contrastive learning objective between images and text that aligns sentences of various complexities to the same image. We also introduce novel negative mining techniques in the scene graph space for improving attribute binding and relation understanding. Through extensive experiments, we demonstrate the effectiveness of our approach that significantly improves attribute binding, relation understanding, systematic generalization, and productivity on multiple recently proposed benchmarks (For example, improvements up to $\mathbf{18}$% for systematic generalization, $\mathbf{16.5}$% for relation understanding over a strong baseline), while achieving similar or better performance than CLIP on various general multimodal tasks.

Commonsense norms are defeasible by context: reading books is usually great, but not when driving a car. While contexts can be explicitly described in language, in embodied scenarios, contexts are often provided visually. This type of visually grounded reasoning about defeasible commonsense norms is generally easy for humans, but (as we show) poses a challenge for machines, as it necessitates both visual understanding and reasoning about commonsense norms. We construct a new multimodal benchmark for studying commonsense norms: NormLens. NormLens consists of 10K human judgments accompanied by free-form explanations covering 2K multimodal situations, and serves as a probe to address two questions: (1) to what extent can models align with average human judgment? and (2) how well can models explain their predicted judgments? We find that state-of-the-art model judgments and explanations are not well-aligned with human annotation. Additionally, we present a simple yet effective approach to better align models with humans by distilling social commonsense knowledge from large language models. The data and code will be released.

Large language models (LLMs) have shown nearly saturated performance on many natural language processing (NLP) tasks. As a result, it is natural for people to believe that LLMs have also mastered abilities such as time understanding and reasoning. However, research on the temporal sensitivity of LLMs has been insufficiently emphasized. To fill this gap, this paper constructs Multiple Sensitive Factors Time QA (MenatQA), which encompasses three temporal factors (scope factor, order factor, counterfactual factor) with total 2,853 samples for evaluating the time comprehension and reasoning abilities of LLMs. This paper tests current mainstream LLMs with different parameter sizes, ranging from billions to hundreds of billions. The results show most LLMs fall behind smaller temporal reasoning models with different degree on these factors. In specific, LLMs show a significant vulnerability to temporal biases and depend heavily on the temporal information provided in questions. Furthermore, this paper undertakes a preliminary investigation into potential improvement strategies by devising specific prompts and leveraging external tools. These approaches serve as valuable baselines or references for future research endeavors.

Large Language Models (LLMs) have gained significant popularity for their impressive performance across diverse fields. However, LLMs are prone to hallucinate untruthful or nonsensical outputs that fail to meet user expectations in many real-world applications. Existing works for detecting hallucinations in LLMs either rely on external knowledge for reference retrieval or require sampling multiple responses from the LLM for consistency verification, making these methods costly and inefficient. In this paper, we propose a novel reference-free, uncertainty-based method for detecting hallucinations in LLMs. Our approach imitates human focus in factuality checking from three aspects: 1) focus on the most informative and important keywords in the given text; 2) focus on the unreliable tokens in historical context which may lead to a cascade of hallucinations; and 3) focus on the token properties such as token type and token frequency. Experimental results on relevant datasets demonstrate the effectiveness of our proposed method, which achieves state-of-the-art performance across all the evaluation metrics and eliminates the need for additional information.

Evaluating the factual consistency of automatically generated summaries is essential for the progress and adoption of reliable summarization systems. Despite recent advances, existing factuality evaluation models are not robust, being especially prone to entity and relation errors in new domains. We propose FactKB---a simple new approach to factuality evaluation that is generalizable across domains, in particular with respect to entities and relations. FactKB is based on language models pretrained using facts extracted from external knowledge bases. We introduce three types of complementary factuality pretraining objectives based on entity-specific facts, facts extracted from auxiliary knowledge about entities, and facts constructed compositionally through knowledge base walks. The resulting factuality evaluation model achieves state-of-the-art performance on two in-domain news summarization benchmarks as well as on three out-of-domain scientific literature datasets. Further analysis of FactKB shows improved ability to detect erroneous entities and relations in summaries and is robust and easily generalizable across domains.

The effectiveness of Chain-of-thought prompting (CoT) has been widely recognized, but the underlying mechanisms behind its success, the reason why it just works for a wide range of tasks, remains an open question. To investigate this, we employ a counterfactual prompting approach, systematically manipulating elements of examples used in a few-shot prompt, and testing the consequences on model behavior. This allows us to understand the relative contributions of prompt elements such as symbols (digits, entities) and patterns (equations, sentence structure) on in-context learning. Our experiments with three different large language models (LLMs) reveal several key findings. First, the specific symbols used in the prompt do not significantly impact the model's performance. However, consistent patterns in examples and specifying text in style frequently found on the web are crucial. Second, our findings suggest that the necessity of accurate few-shot examples depends on their role in communicating task understanding. We identify tasks where inaccurate few-shot examples hurt and, surprisingly, tasks where they improve performance. Additionally, we find that the intermediate steps in CoT may not necessarily facilitate learning how to solve a task, but instead efficiently convey task understanding (what) to the model. Furthermore, CoT leverages LLMs to fill in missing commonsense information, particularly helping difficult reasoning problems and long-tail questions.

The need for grounding in language understanding is an active research topic. Previous work has suggested that color perception and color language appear as a suitable test bed to empirically study the problem, given its cognitive significance and showing that there is considerable alignment between a defined color space and the feature space defined by a language model. To further study this issue, we collect a large scale source of colors and their descriptions, containing almost a 1 million examples , and perform an empirical analysis to compare two kinds of alignments: (i) inter-space, by learning a mapping between embedding space and color space, and (ii) intra-space, by means of prompting comparatives between color descriptions. Our results show that while color space alignment holds for monolexemic, highly pragmatic color descriptions, this alignment drops considerably in the presence of examples that exhibit elements of real linguistic usage such as subjectivity and abstractedness, suggesting that grounding may be required in such cases.

Offensive language detection is crucial in natural language processing (NLP). We investigated the importance of context for detecting such language in reply tweets on Twitter, where the use of offensive language is widespread. We collected a Turkish tweet dataset where the target group was unvaccinated people during the Covid period. Tweets in the dataset were enriched with contextual information by adding the original tweet to which a particular tweet was posted as a reply. The dataset, which includes over 28,000 tweet-reply pairs, was manually labeled by human annotators and made publicly available. In addition, we compared the performance of different machine learning models with and without contextual information. Our results show that this type of contextual information was not very useful in improving the performance of the models in general, although it slightly increased the macro-averaged F1-score of certain models.

Modern NLP models are often trained over large untrusted datasets, raising the potential for a malicious adversary to compromise model behaviour. For instance, backdoors can be implanted through crafting training instances with a specific textual trigger and a target label. This paper posits that backdoor poisoning attacks exhibit a spurious correlation between simple text features and classification labels, and accordingly, proposes methods for mitigating spurious correlation as means of defence. Our empirical study reveals that the malicious triggers are highly correlated to their target labels; therefore such correlations are extremely distinguishable compared to those scores of benign features, and can be used to filter out potentially problematic instances. Compared with several existing defences, our defence method significantly reduces attack success rates across backdoor attacks, and in the case of insertion-based attacks, our method provides a near-perfect defence.

We explore using the 6B parameter GPT-J language model to create a plan for a simulated robot to achieve 30 classes of goals in ScienceWorld, a text game simulator for elementary science experiments and for which previously published empirical work has shown large language models (LLM)s to be a poor fit (Wang et al., 2022). Using the Markov assumption, the LLM outperforms the state-of-the-art based on reinforcement learning by a factor of 1.4. When we fill the LLM’s input buffer with as many prior steps as will fit, improvement rises to 3.3x. Even when training on only 6.5% of the training data, we observe a 2.3x improvement over the state-of-the-art. Our experiments show that performance varies widely across the 30 classes of actions, indicating that averaging over tasks can hide significant performance issues.