EMNLP 2024 论文列表

Question decomposition has emerged as an effective strategy for prompting Large Language Models (LLMs) to answer complex questions. However, while existing methods primarily focus on unimodal language models, the question decomposition capability of Multimodal Large Language Models (MLLMs) has yet to be explored. To this end, this paper explores visual question decomposition on MLLMs. Specifically, we introduce a systematic evaluation framework including a dataset and several evaluation criteria to assess the quality of the decomposed sub-questions, revealing that existing MLLMs struggle to produce high-quality sub-questions. To address this limitation, we propose a specific finetuning dataset, DecoVQA+, for enhancing the model’s question decomposition capability. Aiming at enabling models to perform appropriate selective decomposition, we propose an efficient finetuning pipeline. The finetuning pipeline consists of our proposed dataset and a training objective for selective decomposition. Finetuned MLLMs demonstrate significant improvements in the quality of sub-questions and the policy of selective question decomposition. Additionally, the models also achieve higher accuracy with selective decomposition on VQA benchmark datasets.

Vision Large Language Models (VLLMs) are transforming the intersection of computer vision and natural language processing; however, the potential of using visual prompts for emotion recognition in these models remains largely unexplored and untapped. Traditional methods in VLLMs struggle with spatial localization and often discard valuable global context. We propose a novel Set-of-Vision prompting (SoV) approach that enhances zero-shot emotion recognition by using spatial information, such as bounding boxes and facial landmarks, to mark targets precisely. SoV improves accuracy in face count and emotion categorization while preserving the enriched image context. Through comprehensive experimentation and analysis of recent commercial or open-source VLLMs, we evaluate the SoV model’s ability to comprehend facial expressions in natural environments. Our findings demonstrate the effectiveness of integrating spatial visual prompts into VLLMs for improving emotion recognition performance.

Unsupervised multimodal machine translation (UMMT) aims to leverage vision information as a pivot between two languages to achieve better performance on low-resource language pairs. However, there is presently a challenge: how to handle alignment between distant language pairs (DLPs) in UMMT. To this end, this paper proposes a visual pivoting UMMT method for DLPs. Specifically, we first construct a dataset containing two DLPs, including English-Uyghur and Chinese-Uyghur. We then apply the visual pivoting method for both to pre-training and fine-tuning, and we observe that the images on the encoder and decoder of UMMT have noticeable effects on DLPs. Finally, we introduce informative multi-granularity image features to facilitate further alignment of the latent space between the two languages. Experimental results show that the proposed method significantly outperforms several baselines on DLPs and close language pairs (CLPs).

We present a practical distillation approach to fine-tune LLMs for invoking tools in real-time applications. We focus on visual editing tasks; specifically, we modify images and videos by interpreting user stylistic requests, specified in natural language (“golden hour”), using an LLM to select the appropriate tools and their parameters to achieve the desired visual effect.We found that proprietary LLMs such as GPT-3.5-Turbo show potential in this task, but their high cost and latency make them unsuitable for real-time applications.In our approach, we fine-tune a (smaller) student LLM with guidance from a (larger) teacher LLM and behavioral signals.We introduce offline metrics to evaluate student LLMs. Both online and offline experiments show that our student models manage to match the performance of our teacher model (GPT-3.5-Turbo), significantly reducing costs and latency.Lastly, we show that fine-tuning was improved by 25% in low-data regimes using augmentation.

Recent advances in fine-tuning Vision-Language Models (VLMs) have witnessed the success of prompt tuning and adapter tuning, while the classic model fine-tuning on inherent parameters seems to be overlooked. It is believed that fine-tuning the parameters of VLMs with few-shot samples corrupts the pre-trained knowledge since fine-tuning the CLIP model even degrades performance. In this paper, we revisit this viewpoint, and propose a new perspective: fine-tuning the specific parameters instead of all will uncover the power of classic model fine-tuning on VLMs. Through our meticulous study, we propose ClipFit, a simple yet effective method to fine-tune CLIP without introducing any overhead of extra parameters. We demonstrate that by only fine-tuning the specific bias terms and normalization layers, ClipFit can improve the performance of zero-shot CLIP by 7.27% average harmonic mean accuracy. Lastly, to understand how fine-tuning in CLIPFit affects the pre-trained models, we conducted extensive experimental analyses w.r.t. changes in internal parameters and representations. We found that low-level text bias layers and the first layer normalization layer change much more than other layers. The code will be released.

Large language models (LLMs) are trained from vast repositories of text authored by millions of distinct authors, reflecting an enormous diversity of human traits. While these models bear the potential to be used as approximations of human subjects in behavioral studies, prior efforts have been limited in steering model responses to match individual human users. In this work, we introduce Anthology, a method for conditioning LLMs to particular virtual personas by harnessing open-ended life narratives, which we refer to as backstories. We show that our methodology enhances the consistency and reliability of experimental outcomes while ensuring better representation of diverse sub-populations. Across three nationally representative human surveys conducted as part of Pew Research Center’s American Trends Panel (ATP), we demonstrate that Anthology achieves up to 18% improvement in matching the response distributions of human respondents and 27% improvement in consistency metrics.

Jailbreak attacks on large language models (LLMs) involve inducing these models to generate harmful content that violates ethics or laws, posing a significant threat to LLM security. Current jailbreak attacks face two main challenges: low success rates due to defensive measures and high resource requirements for crafting specific prompts. This paper introduces Virtual Context, which leverages special tokens, previously overlooked in LLM security, to improve jailbreak attacks. Virtual Context addresses these challenges by significantly increasing the success rates of existing jailbreak methods and requiring minimal background knowledge about the target model, thus enhancing effectiveness in black-box settings without additional overhead. Comprehensive evaluations show that Virtual Context-assisted jailbreak attacks can improve the success rates of four widely used jailbreak methods by approximately 40% across various LLMs. Additionally, applying Virtual Context to original malicious behaviors still achieves a notable jailbreak effect. In summary, our research highlights the potential of special tokens in jailbreak attacks and recommends including this threat in red-teaming testing to comprehensively enhance LLM security.

The recent years have witnessed great advances in video generation. However, the development of automatic video metrics is lagging significantly behind. None of the existing metric is able to provide reliable scores over generated videos. The main barrier is the lack of large-scale human-annotated dataset. In this paper, we release VideoFeedback, the first large-scale dataset containing human-provided multi-aspect score over 37.6K synthesized videos from 11 existing video generative models. We train VideoScore (initialized from Mantis)based on VideoFeedback to enable automatic video quality assessment. Experiments show that the Spearman’s correlation betweenVideoScore and humans can reach 77.1 on VideoFeedback-test, beating the prior best metrics by about 50 points. Further result onother held-out EvalCrafter, GenAI-Bench, and VBench show that VideoScore has consistently much higher correlation with humanjudges than other metrics. Due to these results, we believe VideoScore can serve as a great proxy for human raters to (1) rate different video models to track progress (2) simulate fine-grained human feedback in Reinforcement Learning with Human Feedback (RLHF) to improve current video generation models.

In the video-language domain, recent works in leveraging zero-shot Large Language Model-based reasoning for video understanding have become competitive challengers to previous end-to-end models. However, long video understanding presents unique challenges due to the complexity of reasoning over extended timespans, even for zero-shot LLM-based approaches. The challenge of information redundancy in long videos prompts the question of what specific information is essential for large language models (LLMs) and how to leverage them for complex spatial-temporal reasoning in long-form video analysis. We propose a framework VideoINSTA , i.e. INformative Spatial-TemporAl Reasoning for zero-shot long-form video understanding.VideoINSTA contributes (1) a zero-shot framework for long video understanding using LLMs; (2) an event-based temporalreasoning and content-based spatial reasoning approach for LLMs to reason over spatial-temporal information in videos; (3) a self-reflective information reasoning scheme based on information sufficiency and prediction confidence while balancing temporal factors.Our model significantly improves the state-of-the-art on three long video question-answering benchmarks: EgoSchema, NextQA, and IntentQA, and the open question answering dataset ActivityNetQA. Code is released: https://github.com/mayhugotong/VideoINSTA.

Contrastive Language-Image Pre-training (CLIP) has been widely studied and applied in numerous applications. However, the emphasis on brief summary texts during pre-training prevents CLIP from understanding long descriptions. This issue is particularly acute regarding videos given that videos often contain abundant detailed contents. In this paper, we propose the VideoCLIP-XL (eXtra Length) model, which aims to unleash the long-description understanding capability of video CLIP models. Firstly, we establish an automatic data collection system and gather a large-scale VILD pre-training dataset with VIdeo and Long-Description pairs. Then, we propose Text-similarity-guided Primary Component Matching (TPCM) to better learn the distribution of feature space while expanding the long description capability. We also introduce two new tasks namely Detail-aware Description Ranking (DDR) and Hallucination-aware Description Ranking (HDR) for further understanding improvement. Finally, we construct a Long Video Description Ranking (LVDR) benchmark for evaluating the long-description capability more comprehensively. Extensive experimental results on widely-used text-video retrieval benchmarks with both short and long descriptions and our LVDR benchmark can fully demonstrate the effectiveness of our method.

Weakly-supervised Spatio-Temporal Video Grounding(STVG) aims to localize target object tube given a text query, without densely annotated training data. Existing methods extract each candidate tube feature independently by cropping objects from video frame feature, discarding all contextual information such as position change and inter-entity relationship. In this paper, we propose Video-Text Prompting(VTP) to construct candidate feature. Instead of cropping tube region from feature map, we draw visual markers(e.g. red circle) over objects tubes as video prompts; corresponding text prompt(e.g. in red circle) is also inserted after the subject word of query text to highlight its presence. Nevertheless, each candidate feature may look similar without cropping. To address this, we further propose Contrastive VTP(CVTP) by introducing negative contrastive samples whose candidate object is erased instead of being highlighted; by comparing the difference between VTP candidate and the contrastive sample, the gap of matching score between correct candidate and the rest is enlarged. Extensive experiments and ablations are conducted on several STVG datasets and our results surpass existing weakly-supervised methods by a great margin, demonstrating the effectiveness of our proposed methods.

Large Vision-Language Model (LVLM) has enhanced the performance of various downstream tasks in visual-language understanding. Most existing approaches encode images and videos into separate feature spaces, which are then fed as inputs to large language models. However, due to the lack of unified tokenization for images and videos, namely misalignment before projection, it becomes challenging for a Large Language Model (LLM) to learn multi-modal interactions from several poor projection layers.In this work, we unify visual representation into the language feature space to advance the foundational LLM towards a unified LVLM. As a result, we establish a simple but robust LVLM baseline, Video-LLaVA, which learns from a mixed dataset of images and videos, mutually enhancing each other.As a result, Video-LLaVA outperforms Video-ChatGPT by 5.8%, 9.9%, 18.6%, and 10.1% on MSRVTT, MSVD, TGIF, and ActivityNet, respectively. Additionally, our Video-LLaVA also achieves superior performances on a broad range of 9 image benchmarks.Notably, extensive experiments demonstrate that Video-LLaVA mutually benefits images and videos within a unified visual representation, outperforming models designed specifically for images or videos. We aim for this work to provide modest insights into the multi-modal inputs for the LLM.

This paper tackles a new task: discourse parsing for videos, inspired by text discourse parsing based on Rhetorical Structure Theory (RST). The task aims to construct an RST tree for a video to represent its storyline and illustrate the event relationships. We first construct a benchmark dataset by identifying events with their time spans, providing corresponding captions, and constructing RST trees with events as leaves. We then evaluate baseline approaches to video RST parsing: the ‘parsing after captioning’ framework and parsing via visual features. The results show that a parser using gold captions performed the best, while parsers relying on generated captions performed the worst; a parser using visual features provided intermediate performance. However, we observed that parsing via visual features could be improved by pre-training it with video captioning designed to produce a coherent video story. Furthermore, we demonstrated that RST trees obtained from videos contribute to multimodal summarization consisting of keyframes with texts.

While natural language inference (NLI) has emerged as a prominent task for evaluating a model’s capability to perform natural language understanding, creating large benchmarks for training deep learning models imposes a significant challenge since it requires extensive human annotations. To overcome this, we propose to construct pseudo-generated samples (premise-hypothesis pairs) using class-specific fine-tuned large language models (LLMs) thereby reducing the human effort and the costs in annotating large amounts of data. However, despite the impressive performance of LLMs, it is necessary to verify that the pseudo-generated labels are actually correct. Towards this goal, in this paper, we propose VerifyMatch, a semi-supervised learning (SSL) approach in which the LLM pseudo-labels guide the training of the SSL model and, at the same time, the SSL model acts as a verifier of the LLM-generated data. In our approach, we retain all pseudo-labeled samples, but to ensure unlabeled data quality, we further propose to use MixUp whenever the verifier does not agree with the LLM-generated label or when they both agree on the label but the verifier has a low confidence—lower than an adaptive confidence threshold. We achieve competitive accuracy compared to strong baselines for NLI datasets in low-resource settings.

Natural language explanations represent a proxy for evaluating explanation-based and multi-step Natural Language Inference (NLI) models. However, assessing the validity of explanations for NLI is challenging as it typically involves the crowd-sourcing of apposite datasets, a process that is time-consuming and prone to logical errors. To address existing limitations, this paper investigates the verification and refinement of natural language explanations through the integration of Large Language Models (LLMs) and Theorem Provers (TPs). Specifically, we present a neuro-symbolic framework, named Explanation-Refiner, that integrates TPs with LLMs to generate and formalise explanatory sentences and suggest potential inference strategies for NLI. In turn, the TP is employed to provide formal guarantees on the logical validity of the explanations and to generate feedback for subsequent improvements. We demonstrate how Explanation-Refiner can be jointly used to evaluate explanatory reasoning, autoformalisation, and error correction mechanisms of state-of-the-art LLMs as well as to automatically enhance the quality of explanations of variable complexity in different domains.

Recent advances in Large Language Models (LLM) have led to substantial interest in their application to commonsense reasoning tasks. Despite their potential, LLMs are susceptible to reasoning errors and hallucinations that may be harmful in use cases where accurate reasoning is critical. This challenge underscores the need for verifiable, debuggable, and repairable LLM reasoning. Recent works have made progress toward verifiable reasoning with LLMs by using them as either (i) a reasoner over an axiomatic knowledge base, or (ii) a semantic parser for use in existing logical inference systems. However, both settings are unable to extract commonsense axioms from the LLM that are not already formalized in the knowledge base, and also lack a reliable method to repair missed commonsense inferences. In this work, we present LLM-TRes, a logical reasoning framework based on the notion of “theory resolution” that allows for seamless integration of the commonsense knowledge from LLMs with a verifiable logical reasoning framework that mitigates hallucinations and facilitates debugging of the reasoning procedure as well as repair. We crucially prove that repaired axioms are theoretically guaranteed to be given precedence over flawed ones in our theory resolution inference process. We conclude by evaluating on three diverse language-based reasoning tasks—preference reasoning, deductive reasoning, and causal commonsense reasoning—and demonstrate the superior performance of LLM-TRes vs. state-of-the-art LLM-based reasoning methods in terms of both accuracy and reasoning correctness.

Existing metrics for evaluating the factuality of long-form text, such as FACTSCORE (Min et al., 2023) and SAFE (Wei et al., 2024), decompose an input text into “atomic claims” and verify each against a knowledge base like Wikipedia. These metrics are not suitable for most generation tasks because they assume that every claim is verifiable (i.e., can plausibly be proven true or false). We address this issue with VERISCORE,1 a metric for evaluating factuality in diverse long-form generation tasks that contain both verifiable and unverifiable content. VERISCORE can be effectively implemented with either closed or fine-tuned open-weight language models. Human evaluation confirms that VERISCORE’s extracted claims are more sensible than those from competing methods across eight different long-form tasks. We use VERISCORE to evaluate generations from 16 different models across multiple long-form tasks and find that while GPT-4o is the best-performing model overall, open-weight models such as Mixtral-8×22 are closing the gap. We show that an LM’s VERISCORE on one task (e.g., biography generation) does not necessarily correlate to its VERISCORE on a different task (e.g., long-form QA), highlighting the need for expanding factuality evaluation across tasks with varying fact density.

We introduce a new database of cognate words and etymons for the five main Romance languages, the most comprehensive one to date. We propose a strong benchmark for the automatic reconstruction of protowords for Romance languages, by applying a set of machine learning models and features on these data. The best results reach 90% accuracy in predicting the protoword of a given cognate set, surpassing existing state-of-the-art results for this task and showing that computational methods can be very useful in assisting linguists with protoword reconstruction.

Semantic similarity between two sentences is inherently subjective and can vary significantly based on the specific aspects emphasized. Consequently, traditional sentence encoders must be capable of generating conditioned sentence representations that account for diverse conditions or aspects. In this paper, we propose a novel yet efficient framework based on transformer-style language models that facilitates advanced conditioned sentence representation while maintaining model parameters and computational efficiency. Empirical evaluations on the Conditional Semantic Textual Similarity and Knowledge Graph Completion tasks demonstrate the superiority of our proposed framework.