EMNLP 2023 论文列表

Referring image segmentation is a task that aims to predict pixel-wise masks corresponding to objects in an image described by natural language expressions. Previous methods for referring image segmentation employ a cascade framework to break down complex problems into multiple stages. However, its defects also obvious: existing methods within the cascade framework may encounter challenges in both maintaining a strong focus on the most relevant information during specific stages of the referring image segmentation process and rectifying errors propagated from early stages, which can ultimately result in sub-optimal performance. To address these limitations, we propose the Joint Mask Contextual Embedding Learning Network (JMCELN). JMCELN is designed to enhance the Cascade Framework by incorporating a Learnable Contextual Embedding and a Progressive Alignment Network (PAN). The Learnable Contextual Embedding module dynamically stores and utilizes reasoning information based on the current mask prediction results, enabling the network to adaptively capture and refine pertinent information for improved mask prediction accuracy. Furthermore, the Progressive Alignment Network (PAN) is introduced as an integral part of JMCELN. PAN leverages the output from the previous layer as a filter for the current output, effectively reducing inconsistencies between predictions from different stages. By iteratively aligning the predictions, PAN guides the Learnable Contextual Embedding to incorporate more discriminative information for reasoning, leading to enhanced prediction quality and a reduction in error propagation. With these methods, we achieved state-of-the-art results on three commonly used benchmarks, especially in more intricate datasets. The code will be released.

Knowledge Distillation (KD) compresses computationally expensive pre-trained language models (PLMs) by transferring their knowledge to smaller models, allowing their use in resource-constrained or real-time settings. However, most smaller models fail to surpass the performance of the original larger model, resulting in sacrificing performance to improve inference speed. To address this issue, we propose Co-Training and Co-Distillation (CTCD), a novel framework that improves performance and inference speed together by co-training two models while mutually distilling knowledge. The CTCD framework successfully achieves this based on two significant findings: 1) Distilling knowledge from the smaller model to the larger model during co-training improves the performance of the larger model. 2) The enhanced performance of the larger model further boosts the performance of the smaller model. The CTCD framework shows promise as it can be combined with existing techniques like architecture design or data augmentation, replacing one-way KD methods, to achieve further performance improvement. Extensive ablation studies demonstrate the effectiveness of CTCD, and the small model distilled by CTCD outperforms the original larger model by a significant margin of 1.66 on the GLUE benchmark.

Large decoder-only language models (LMs) can be largely improved in terms of perplexity by retrieval (e.g., RETRO), but its impact on text generation quality and downstream task accuracy is unclear. Thus, it is still an open question: shall we pretrain large autoregressive LMs with retrieval? To answer it, we perform a comprehensive study on a scalable pre-trained retrieval-augmented LM (i.e., RETRO) compared with standard GPT and retrieval-augmented GPT incorporated at fine-tuning or inference stages. We first provide the recipe to reproduce RETRO up to 9.5B parameters while retrieving a text corpus with 330B tokens. Based on that, we have the following novel findings: i) RETRO outperforms GPT on text generation with much less degeneration (i.e., repetition), moderately higher factual accuracy, and slightly lower toxicity with a nontoxic retrieval database. ii) On the LM Evaluation Harness benchmark, RETRO largely outperforms GPT on knowledge-intensive tasks, but is on par with GPT on other tasks. Furthermore, we introduce a simple variant of the model, RETRO++, which largely improves open-domain QA results of original RETRO (e.g., EM score +8.6 on Natural Question) and significantly outperforms retrieval-augmented GPT across different model sizes. Our findings highlight the promising direction of pretraining autoregressive LMs with retrieval as future foundation models. We release our implementation at: https://github.com/NVIDIA/Megatron-LM/tree/main/tools/retro.

Knowledge probing is a task to assess the knowledge encoded within pre-trained language models (PLMs) by having the PLM complete prompts such as "Italy is located in \_\_,". The model's prediction precision serves as a lower bound for the amount of knowledge it contains. Subsequent works explore training a series of vectors as prompts to guide PLMs towards more accurate predictions. However, these methods compromise the readability of the prompts. We cannot directly understand these prompts from their literal meaning, making it difficult to verify whether they are correct. Consequently, the credibility of probing results derived from these prompts is diminished. To address the issue, we propose a novel method called ReadPrompt, which aims to identify meaningful sentences to serve as prompts. Experiments show that ReadPrompt achieves state-of-the-art performance on the current knowledge probing benchmark. Moreover, since the prompt is readable, we discovered a misalignment between constructed prompts and knowledge, which is also present in current prompting methods verified by an attack experiment. We claim that the probing outcomes of the current prompting methods are unreliable that overestimate the knowledge contained within PLMs.

Current scientific fact-checking benchmarks exhibit several shortcomings, such as biases arising from crowd-sourced claims and an over-reliance on text-based evidence. We present SCITAB, a challenging evaluation dataset consisting of 1.2K expert-verified scientific claims that 1) originate from authentic scientific publications and 2) require compositional reasoning for verification. The claims are paired with evidence-containing scientific tables annotated with labels. Through extensive evaluations, we demonstrate that SCITAB poses a significant challenge to state-of-the-art models, including table-based pretraining models and large language models. All models except GPT-4 achieved performance barely above random guessing. Popular prompting techniques, such as Chain-of-Thought, do not achieve much performance gains on SCITAB. Our analysis uncovers several unique challenges posed by SCITAB, including table grounding, claim ambiguity, and compositional reasoning. Our codes and data are publicly available at https://github.com/XinyuanLu00/SciTab.

Previous entity disambiguation (ED) methods adopt a discriminative paradigm, where prediction is made based on matching scores between mention context and candidate entities using length-limited encoders. However, these methods often struggle to capture explicit discourse-level dependencies, resulting in incoherent predictions at the abstract level (e.g. topic or category). We propose CoherentED, an ED system equipped with novel designs aimed at enhancing the coherence of entity predictions. Our method first introduces an unsupervised variational autoencoder (VAE) to extract latent topic vectors of context sentences. This approach not only allows the encoder to handle longer documents more effectively, conserves valuable input space, but also keeps a topic-level coherence. Additionally, we incorporate an external category memory, enabling the system to retrieve relevant categories for undecided mentions. By employing step-by-step entity decisions, this design facilitates the modeling of entity-entity interactions, thereby maintaining maximum coherence at the category level. We achieve new state-of-the-art results on popular ED benchmarks, with an average improvement of 1.3 F1 points. Our model demonstrates particularly outstanding performance on challenging long-text scenarios.

We investigate the predictability of large language model (LLM) capabilities: given records of past experiments using different model families, numbers of parameters, tasks, and numbers of in-context examples, can we accurately predict LLM performance on new experiment configurations? Answering this question has practical implications for LLM users (e.g., deciding which models to try), developers (e.g., prioritizing evaluation on representative tasks), and the research community (e.g., identifying hard-to-predict capabilities that warrant further investigation). We study the performance prediction problem on experiment records from BIG-bench. On a random train-test split, an MLP-based predictor achieves an $R^2$ score greater than 95%, indicating the presence of learnable patterns within the experiment records. We then formulate the problem of searching for "small-bench," an informative subset of BIG-bench tasks from which the performance on the full set can be maximally recovered. We find a subset as informative as BIG-bench Hard for evaluating new model families, while being $3\times$ smaller. Additionally, we find competitive subsets by clustering task representations learned by our MLP-based predictor and selecting tasks close to cluster centroids, highlighting the importance of task diversity in constructing "small-bench."

Simultaneous Speech Translation (SimulST) is a task focused on ensuring high-quality translation of speech in low-latency situations. Despite this, the modality gap (\emph{e.g.}, unknown word boundaries) between audio and text presents a challenge. This gap hinders the effective application of policies from simultaneous text translation (SimulMT) and compromises the performance of offline speech translation. To address this issue, we first leverage the Montreal Forced Aligner (MFA) and utilize audio transcription pairs in pre-training the acoustic encoder, and introduce a token-level cross-modal alignment that allows the wait-$k$ policy from SimulMT to better adapt to SimulST. This token-level boundary alignment simplifies the decision-making process for predicting read/write actions, as if the decoder were directly processing text tokens. Subsequently, to optimize the SimulST task, we propose a robust and random wait-$k$-tokens strategy. This strategy allows a single model to meet various latency requirements and minimizes error accumulation of boundary alignment during inference. Our experiments on the MuST-C dataset show that our method achieves better trade-off between translation quality and latency.

Embodied language comprehension emphasizes that language understanding is not solely a matter of mental processing in the brain but also involves interactions with the physical and social environment. With the explosive growth of Large Language Models (LLMs) and their already ubiquitous presence in our daily lives, it is becoming increasingly necessary to verify their real-world understanding. Inspired by cognitive theories, we propose POSQA: a Physical Object Size Question Answering dataset with simple size comparison questions to examine the extremity and analyze the potential mechanisms of the embodied comprehension of the latest LLMs. We show that even the largest LLMs today perform poorly under the zero-shot setting. We then push their limits with advanced prompting techniques and external knowledge augmentation. Furthermore, we investigate whether their real-world comprehension primarily derives from contextual information or internal weights and analyse the impact of prompt formats and report bias of different objects. Our results show that real-world understanding that LLMs shaped from textual data can be vulnerable to deception and confusion by the surface form of prompts, which makes it less aligned with human behaviours.

We propose a framework for online multimodal dialog act (DA) classification based on raw audio and ASR-generated transcriptions of current and past utterances. Existing multimodal DA classification approaches are limited by ineffective audio modeling and late-stage fusion. We showcase significant improvements in multimodal DA classification by integrating modalities at a more granular level and incorporating recent advancements in large language and audio models for audio feature extraction. We further investigate the effectiveness of self-attention and cross-attention mechanisms in modeling utterances and dialogs for DA classification. We achieve a substantial increase of 3 percentage points in the F1 score relative to current state-of-the-art models on two prominent DA classification datasets, MRDA and EMOTyDA.

While text style transfer has many applications across natural language processing, the core premise of transferring from a single source style is unrealistic in a real-world setting. In this work, we focus on arbitrary style transfer: rewriting a text from an arbitrary, unknown style to a target style. We propose STEER: Unified Style Transfer with Expert Reinforcement, a unified frame-work developed to overcome the challenge of limited parallel data for style transfer. STEER involves automatically generating a corpus of style-transfer pairs using a product of experts during decoding. The generated offline data is then used to pre-train an initial policy before switching to online, off-policy reinforcement learning for further improvements via fine-grained reward signals. STEER is unified and can transfer to multiple target styles from an arbitrary, unknown source style, making it particularly flexible and efficient. Experimental results on a challenging dataset with text from a diverse set of styles demonstrate state-of-the-art results compared to competitive baselines. Remarkably, STEER outperforms the 175B parameter instruction-tuned GPT-3 on overall style transfer quality, despite being 226 times smaller in size. We also show STEER is robust, maintaining its style transfer capabilities on out-of-domain data, and surpassing nearly all baselines across various styles. The success of our method highlights the potential of RL algorithms when augmented with controllable decoding to overcome the challenge of limited data supervision.

Argument structure extraction (ASE) aims to identify the discourse structure of arguments within documents. Previous research has demonstrated that contextual information is crucial for developing an effective ASE model. However, we observe that merely concatenating sentences in a contextual window does not fully utilize contextual information and can sometimes lead to excessive attention on less informative sentences. To tackle this challenge, we propose an Efficient Context-aware ASE model (ECASE) that fully exploits contextual information by enhancing modeling capacity and augmenting training data. Specifically, we introduce a sequence-attention module and distance-weighted similarity loss to aggregate contextual information and argumentative information. Additionally, we augment the training data by randomly masking discourse markers and sentences, which reduces the model's reliance on specific words or less informative sentences. Our experiments on five datasets from various domains demonstrate that our model achieves state-of-the-art performance. Furthermore, ablation studies confirm the effectiveness of each module in our model.

Large Language Models (LLMs) have shown impressive capabilities in various applications, but they still face various inconsistency issues. Existing works primarily focus on the inconsistency issues within a single LLM, while we complementarily explore the inter-consistency among multiple LLMs for collaboration. To examine whether LLMs can collaborate effectively to achieve a consensus for a shared goal, we focus on commonsense reasoning, and introduce a formal debate framework (FORD) to conduct a three-stage debate among LLMs with real-world scenarios alignment: fair debate, mismatched debate, and roundtable debate. Through extensive experiments on various datasets, LLMs can effectively collaborate to reach a consensus despite noticeable inter-inconsistencies, but imbalances in their abilities can lead to domination by superior LLMs. Leveraging a more advanced LLM like GPT-4 as an authoritative judge can boost collaboration performance. Our work contributes to understanding the inter-consistency among LLMs and lays the foundation for developing future collaboration methods. Codes and data are available at https://github.com/Waste-Wood/FORD.

Humans produce and consume language in a particular cultural context, which includes knowledge about specific norms and practices. A listener's awareness of the cultural context is critical for interpreting the speaker's meaning. A simple expression like *I didn't leave a tip* implies a strong sense of dissatisfaction when tipping is assumed to be the norm. As NLP systems reach users from different cultures, achieving culturally aware language understanding becomes increasingly important. However, current research has focused on building cultural knowledge bases without studying how such knowledge leads to contextualized interpretations of texts. In this work, we operationalize cultural variations in language understanding through a natural language inference (NLI) task that surfaces cultural variations as label disagreement between annotators from different cultural groups. We introduce the first Culturally Aware Natural Language Inference (CALI) dataset with 2.7K premise-hypothesis pairs annotated by two cultural groups located in the U.S. and India. With CALI, we categorize how cultural norms affect language understanding and present an evaluation framework to assess at which levels large language models are culturally aware. Our dataset is available at https://github.com/SALT-NLP/CulturallyAwareNLI.

Detecting negatives (such as non-entailment relationships, unanswerable questions, and false claims) is an important and challenging aspect of many natural language understanding tasks. Though manually collecting challenging negative examples can help models detect them, it is both costly and domain-specific. In this work, we propose Self-labeled Counterfactuals for Extrapolating to Negative Examples (SCENE), an automatic method for synthesizing training data that greatly improves models' ability to detect challenging negative examples. In contrast with standard data augmentation, which synthesizes new examples for existing labels, SCENE can synthesize negative examples zero-shot from only positive ones. Given a positive example, SCENE perturbs it with a mask infilling model, then determines whether the resulting example is negative based on a self-training heuristic. With access to only answerable training examples, SCENE can close 69.6% of the performance gap on SQuAD 2.0, a dataset where half of the evaluation examples are unanswerable, compared to a model trained on SQuAD 2.0. Our method also extends to boolean question answering and recognizing textual entailment, and improves generalization from SQuAD to ACE-whQA, an out-of-domain extractive QA benchmark.

Reply suggestion systems represent a staple component of many instant messaging and email systems. However, the requirement to produce sets of replies, rather than individual replies, makes the task poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. As a result, these system often rely on additional post-processing modules to diversify the outputs. However, these approaches are ultimately bottlenecked by the performance of the initial retriever, which in practice struggles to present a sufficiently diverse range of options to the downstream diversification module, leading to the suggestions being less relevant to the user. In this paper, we consider a novel approach that radically simplifies this pipeline through an autoregressive text-to-text retrieval model, that learns the smart reply task end-to-end from a dataset of (message, reply set) pairs obtained via bootstrapping. Empirical results show this method consistently outperforms a range of state-of-the-art baselines across three datasets, corresponding to a 5.1\%-17.9\% improvement in relevance, and a 0.5\%-63.1\% improvement in diversity compared to the best baseline approach. We make our code publicly available.

The potential of using a large language model (LLM) as a knowledge base (KB) has sparked significant interest. To maintain the knowledge acquired by LLMs, we need to ensure that the editing of learned facts respects internal logical constraints, which are known as dependency of knowledge. Existing work on editing LLMs has partially addressed the issue of dependency, when the editing of a fact should apply to its lexical variations without disrupting irrelevant ones. However, they neglect the dependency between a fact and its logical implications. We propose an evaluation protocol with an accompanying question-answering dataset, StandUp, that provides a comprehensive assessment of the editing process considering the above notions of dependency. Our protocol involves setting up a controlled environment in which we edit facts and monitor their impact on LLMs, along with their implications based on If-Then rules. Extensive experiments on StandUp show that existing knowledge editing methods are sensitive to the surface form of knowledge, and that they have limited performance in inferring the implications of edited facts.

Despite the success of spoken language understanding (SLU) in high-resource languages, achieving similar performance in low-resource settings, such as zero-shot scenarios, remains challenging due to limited labeled training data. To improve zero-shot cross-lingual SLU, recent studies have explored code-switched sentences containing tokens from multiple languages. However, vanilla code-switched sentences often lack semantic and grammatical coherence. We ascribe this lack to two issues: (1) randomly replacing code-switched tokens with equal probability and (2) disregarding token-level dependency within each language. To tackle these issues, in this paper, we propose a novel method termed SoGo, for zero-shot cross-lingual SLU. First, we use a saliency-based substitution approach to extract keywords as substitution options. Then, we introduce a novel token-level alignment strategy that considers the similarity between the context and the code-switched tokens, ensuring grammatical coherence in code-switched sentences. Extensive experiments and analyses demonstrate the superior performance of SoGo across nine languages on MultiATIS++.

Large Language Models (LLMs) are trained on corpora disproportionally weighted in favor of Standard American English. As a result, speakers of other dialects experience significantly more failures when interacting with these technologies. In practice, these speakers often accommodate their speech to be better understood. Our work shares the belief that language technologies should be designed to accommodate the diversity in English dialects and not the other way around. However, prior work on dialect struggle with generalizing to evolving and emerging dialects in a scalable manner. To fill this gap, our method, HyperLoRA, leverages expert linguistic knowledge to enable resource-efficient adaptation via hypernetworks. By disentangling dialect-specific and cross-dialectal information, HyperLoRA improves generalization to unseen dialects in a task-agnostic fashion. Not only is HyperLoRA more scalable in the number of parameters, but it also achieves the best or most competitive performance across 5 dialects in a zero-shot setting. In this way, our approach facilitates access to language technology for billions of English dialect speakers who are traditionally underrepresented.