EMNLP 2024 论文列表

Visual representation learning has been a cornerstone in computer vision, involving typical forms such as visual embeddings, structural symbols, and text-based representations. Despite the success of CLIP-type visual embeddings, they often lack access to world knowledge critical for visual reasoning. In this work, we propose Visual Table, a novel form of visual representation tailored for visual reasoning. Visual tables are constructed as hierarchical descriptions of visual scenes, featuring a scene description and multiple object-centric descriptions covering categories, attributes, and knowledge. Thanks to the structural and textual formats, visual tables offer unique properties over mere visual embeddings, such as explainability and controllable editing. Furthermore, they deliver instance-level world knowledge and detailed attributes that are essential for visual reasoning. To create visual tables, we develop a generator trained on the dataset with collected, small-scale annotations. Extensive results on 11 visual reasoning benchmarks demonstrate that the generated visual tables significantly outperform previous structural and text-based representations. Moreover, they consistently enhance state-of-the-art multi-modal large language models across diverse benchmarks, showcasing their potential for advancing visual reasoning tasks. Our code is available at https://github.com/LaVi-Lab/Visual-Table.

Creating human-like large language model (LLM) agents is crucial for faithful social simulation. Having LLMs role-play based on demographic information sometimes improves human likeness but often does not. This study assessed whether LLM alignment with human behavior can be improved by integrating information from empirically-derived human belief networks. Using data from a human survey, we estimated a belief network encompassing 64 topics loading on nine non-overlapping latent factors. We then seeded LLM-based agents with an opinion on one topic, and assessed the alignment of its expressed opinions on remaining test topics with corresponding human data. Role-playing based on demographic information alone did not align LLM and human opinions, but seeding the agent with a single belief greatly improved alignment for topics related in the belief network, and not for topics outside the network. These results suggest a novel path for human-LLM belief alignment in work seeking to simulate and understand patterns of belief distributions in society.

Machine Translation (MT) evaluation metrics assess translation quality automatically. Recently, researchers have employed MT metrics for various new use cases, such as data filtering and translation re-ranking. However, most MT metrics return assessments as scalar scores that are difficult to interpret, posing a challenge to making informed design choices. Moreover, MT metrics’ capabilities have historically been evaluated using correlation with human judgment, which, despite its efficacy, falls short of providing intuitive insights into metric performance, especially in terms of new metric use cases. To address these issues, we introduce an interpretable evaluation framework for MT metrics. Within this framework, we evaluate metrics in two scenarios that serve as proxies for the data filtering and translation re-ranking use cases. Furthermore, by measuring the performance of MT metrics using Precision, Recall, and F-score, we offer clearer insights into their capabilities than correlation with human judgments. Finally, we raise concerns regarding the reliability of manually curated data following the Direct Assessments+Scalar Quality Metrics (DA+SQM) guidelines, reporting a notably low agreement with Multidimensional Quality Metrics (MQM) annotations.

In this work, we address the challenge of cross-lingual proper noun recognition in automatic speech recognition (ASR), where proper nouns in an utterance may originate from a language different from the language in which the ASR system is trained. We enhance the performance of end-to-end ASR systems by instructing a large language model (LLM) to correct the ASR model’s predictions. The LLM’s context is augmented with a dictionary of cross-lingual words that are phonetically and graphemically similar to the potentially incorrect proper nouns in the ASR predictions. Our dictionary-based method DiP-ASR (Dictionary-based Prompting for Automatic Speech Recognition) significantly reduces word error rates compared to both the end-to-end ASR baseline and instruction-based prompting of the LLM without the dictionary across cross-lingual proper noun recognition tasks involving three secondary languages.

While current Automated Essay Scoring (AES) methods demonstrate high scoring agreement with human raters, their decision-making mechanisms are not fully understood. Our proposed method, using counterfactual intervention assisted by Large Language Models (LLMs), reveals that BERT-like models primarily focus on sentence-level features, whereas LLMs such as GPT-3.5, GPT-4 and Llama-3 are sensitive to conventions & accuracy, language complexity, and organization, indicating a more comprehensive rationale alignment with scoring rubrics. Moreover, LLMs can discern counterfactual interventions when giving feedback on essays. Our approach improves understanding of neural AES methods and can also apply to other domains seeking transparency in model-driven decisions.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks. However, current training approaches combine standard cross-entropy loss with extensive data, human feedback, or ad hoc methods to enhance performance. These solutions are often not scalable or feasible due to their associated costs, complexity, or resource requirements. This study investigates the use of established semantic segmentation loss functions in natural language generation to create a versatile, practical, and scalable solution for fine-tuning different architectures. We evaluate their effectiveness in solving Math Word Problems and question answering across different models of varying sizes. For the analyzed tasks, we found that the traditional Cross-Entropy loss represents a sub-optimal choice, while models trained to minimize alternative (task-dependent) losses, such as Focal or Lovász, achieve a mean improvement of +36% on exact match without requiring additional data or human feedback. These findings suggest a promising pathway for more efficient and accessible training processes.

To ensure large language models contain up-to-date knowledge, they need to be updated regularly. However, model editing is challenging as it might also affect knowledge that is unrelated to the new data. State-of-the-art methods identify parameters associated with specific knowledge and then modify them via direct weight updates. However, these locate-and-edit methods suffer from heavy computational overhead and lack theoretical validation. In contrast, directly fine-tuning the model on requested edits affects the model’s behavior on unrelated knowledge, and significantly damages the model’s generation fluency and consistency. To address these challenges, we propose SAUL, a streamlined model editing method that uses sentence concatenation with augmented random facts for generation regularization. Evaluations on three model editing benchmarks show that is a practical and reliable solution for model editing outperforming state-of-the-art methods while maintaining generation quality and reducing computational overhead.

We propose a new method, instruction back-and-forth translation, to improve the quality of instruction-tuning data used for aligning large language models (LLMs). Given preprocessed texts from an initial web corpus (e.g. Dolma (Soldaini et al., 2024)), we generate synthetic instructions using the backtranslation approach proposed by Li et al., (2023), filter the generated data and rewrite the responses to improve their quality further based on the initial texts. Given similar quantities of instructions, fine-tuning Llama-2 on our (synthetic instruction, rewritten response) pairs yields better AlpacaEval win rates than using other common instruction datasets such as Humpback, ShareGPT, Open Orca, Alpaca-GPT4 and Self-instruct, at both 7B and 70B parameter scales. We also demonstrate that rewriting the responses with an LLM is different from direct distillation: the former process yields better win rate at 70B scale, and the two text distributions exhibit significant distinction in the embedding space. Besides, we provide analyses showing that our backtranslated instructions are of higher quality than other sources of synthetic instructions, while our responses are more diverse and complex than what can be obtained from distillation. Overall we find that instruction back-and-forth translation combines the best of both worlds—making use of the information diversity and quantity found on the web, while ensuring the quality of the responses which is necessary for effective alignment.

Foundation models and vision-language pre-training have notably advanced Vision Language Models (VLMs), enabling multimodal processing of visual and linguistic data. However, their performance has been typically assessed on general scene understanding - recognizing objects, attributes, and actions - rather than cultural comprehension. This study introduces CulturalVQA, a visual question-answering benchmark aimed at assessing VLM’s geo-diverse cultural understanding. We curate a diverse collection of 2,378 image-question pairs with 1-5 answers per question representing cultures from 11 countries across 5 continents. The questions probe understanding of various facets of culture such as clothing, food, drinks, rituals, and traditions. Benchmarking VLMs on CulturalVQA, including GPT-4V and Gemini, reveals disparity in their level of cultural understanding across regions, with strong cultural understanding capabilities for North America while significantly weaker capabilities for Africa. We observe disparity in their performance across cultural facets too, with clothing, rituals, and traditions seeing higher performances than food and drink. These disparities help us identify areas where VLMs lack cultural understanding and demonstrate the potential of CulturalVQA as a comprehensive evaluation set for gauging VLM progress in understanding diverse cultures.

Translating culture-related content is vital for effective cross-cultural communication. However, many culture-specific items (CSIs) often lack literal translation across languages, making it challenging to collect high-quality, diverse parallel corpora with CSI annotations. This difficulty hinders the analysis of cultural awareness of machine translation (MT) systems, including traditional neural MT and the emerging MT paradigm using large language models (LLM). To address this gap, we introduce a novel parallel corpus, enriched with CSI annotations in 6 language pairs for investigating Cultural-Aware Machine Translation—CAMT. Furthermore, we design two evaluation metrics to assess CSI translations, focusing on their pragmatic translation quality. Our findings show the superior ability of LLMs over neural MTs in leveraging external cultural knowledge for translating CSIs, especially those lacking translations in the target culture.

The capability to reason from text is crucial for real-world NLP applications. Real-world scenarios often involve incomplete or evolving data. In response, individuals update their beliefs and understandings accordingly. However, most existing evaluations assume that language models (LMs) operate with consistent information. We introduce Belief-R, a new dataset designed to test LMs’ belief revision ability when presented with new evidence. Inspired by how humans suppress prior inferences, this task assesses LMs within the newly proposed delta reasoning (𝛥 R) framework. Belief-R features sequences of premises designed to simulate scenarios where additional information could necessitate prior conclusions drawn by LMs. We evaluate ~30 LMs across diverse prompting strategies and found that LMs generally struggle to appropriately revise their beliefs in response to new information. Further, models adept at updating often underperformed in scenarios without necessary updates, highlighting a critical trade-off. These insights underscore the importance of improving LMs’ adaptiveness to changing information, a step toward more reliable AI systems.

Large Language Models (LLMs) have gained increasing attention for their remarkable capacity, alongside concerns about safety arising from their potential to produce harmful content. Red teaming aims to find prompts that could elicit harmful responses from LLMs, and is essential to discover and mitigate safety risks before real-world deployment. However, manual red teaming is both time-consuming and expensive, rendering it unscalable. In this paper, we propose RTPE, a scalable evolution framework to evolve red teaming prompts across both breadth and depth dimensions, facilitating the automatic generation of numerous high-quality and diverse red teaming prompts. Specifically, in-breadth evolving employs a novel enhanced in-context learning method to create a multitude of quality prompts, whereas in-depth evolving applies customized transformation operations to enhance both content and form of prompts, thereby increasing diversity. Extensive experiments demonstrate that RTPE surpasses existing representative automatic red teaming methods on both attack success rate and diversity. In addition, based on 4,800 red teaming prompts created by RTPE, we further provide a systematic analysis of 8 representative LLMs across 8 sensitive topics.

For a conversation to help and support, speakers should maintain an “effect-effort” tradeoff. As outlined in the gist of “Cognitive Relevance Principle”, helpful speakers should optimize the “cognitive relevance” through maximizing the “cognitive effects” and minimizing the “processing effort” imposed on listeners. Although preference learning methods have given rise a boon of studies in pursuit of“effect-optimization”, none have delved into the critical “effort-optimiazation” to fully cultivate the awareness of “optimal relevance” into thecognition of conversation agents. To address this gap, we integrate the “Cognitive Relevance Principle” into emotional support agents in the environment of multi-turn conversation. The results demonstrate a significant and robust improvement against the baseline systems with respect to response quality, human-likedness and supportivenss. This study offers compelling evidence for the effectiveness of the “Relevance Principle” in generating human-like, helpful, and harmless emotional support conversations. The source code will be available at https://github.com/CN-Eyetk/VLESA-ORL.git

Large language models (LLMs) can adapt to new tasks through in-context learning (ICL) based on a few examples presented in dialogue history without any model parameter update. Despite such convenience, the performance of ICL heavily depends on the quality of the in-context examples presented, which makes the in-context example selection approach a critical choice. This paper proposes a novel eBayesian in-Context example Selection method (ByCS) for ICL. Extending the inference probability conditioned on in-context examples based on Bayes’ theorem, ByCS focuses on the inverse inference conditioned on test input. Following the assumption that accurate inverse inference probability (likelihood) will result in accurate inference probability (posterior), in-context examples are selected based on their inverse inference results. Diverse and extensive cross-tasking and cross-modality experiments are performed with speech, text, and image examples. Experimental results show the efficacy and robustness of our ByCS method on various models, tasks and modalities.

Recent advances in large language models (LLMs) show the potential of using LLMs as evaluators for assessing the quality of text generations from LLMs. However, applying LLM evaluators naively to compare different systems can lead to unreliable results due to the inaccuracy and intrinsic bias of LLM evaluators. In order to mitigate this problem, we propose two calibration methods, Bayesian Win-Rate Sampling (BWRS) and Bayesian Dawid-Skene, both of which leverage Bayesian inference to more accurately infer the true win rate of generative language models. We empirically validate our methods on six datasets covering story generation, summarization, and instruction following tasks. We show that both our methods are effective in improving the accuracy of win rate estimation using LLMs as evaluators, offering a promising direction for reliable automatic text quality evaluation.

This paper presents BattleAgent, a detailed emulation demonstration system that combines the Large Vision-Language Model (VLM) and Multi-Agent System (MAS). This novel system aims to emulate complex dynamic interactions among multiple agents, as well as between agents and their environments, over a period of time. The emulation showcases the current capabilities of agents, featuring fine-grained multi-modal interactions between agents and landscapes. It develops customizable agent structures to meet specific situational requirements, for example, a variety of battle-related activities like scouting and trench digging. These components collaborate to recreate historical events in a lively and comprehensive manner. This methodology holds the potential to substantially improve visualization of historical events and deepen our understanding of historical events especially from the perspective of decision making. The data and code for this project are accessible at https://github.com/agiresearch/battleagent and the demo is accessible at https://drive.google.com/file/d/1I5B3KWiYCSSP1uMiPGNmXlTmild-MzRJ/view?usp=sharing.

Low-resource languages like Bangla are severely limited by the lack of datasets. Romanized Bangla texts are ubiquitous on the internet, offering a rich source of data for Bangla NLP tasks and extending the available data sources. However, due to the informal nature of romanized text, they often lack the structure and consistency needed to provide insights. We address these challenges by proposing: (1) BanglaTLit, the large-scale Bangla transliteration dataset consisting of 42.7k samples, (2) BanglaTLit-PT, a pre-training corpus on romanized Bangla with 245.7k samples, (3) encoders further-pretrained on BanglaTLit-PT achieving state-of-the-art performance in several romanized Bangla classification tasks, and (4) multiple back-transliteration baseline methods, including a novel encoder-decoder architecture using further pre-trained encoders. Our results show the potential of automated Bangla back-transliteration in utilizing the untapped sources of romanized Bangla to enrich this language. The code and datasets are publicly available: https://github.com/farhanishmam/BanglaTLit.

In the realm of dialogue-to-image retrieval, the primary challenge is to fetch images from a pre-compiled database that accurately reflect the intent embedded within the dialogue history. Existing methods often overemphasize inter-modal alignment, neglecting the nuanced nature of conversational context. Dialogue histories are frequently cluttered with redundant information and often lack direct image descriptions, leading to a substantial disconnect between conversational content and visual representation. This study introduces VCU, a novel framework designed to enhance the comprehension of dialogue history and improve cross-modal matching for image retrieval. VCU leverages large language models (LLMs) to perform a two-step extraction process. It generates precise image-related descriptions from dialogues, while also enhancing visual representation by utilizing object-list texts associated with images. Additionally, auxiliary query collections are constructed to balance the matching process, thereby reducing bias in similarity computations. Experimental results demonstrate that VCU significantly outperforms baseline methods in dialogue-to-image retrieval tasks, highlighting its potential for practical application and effectiveness in bridging the gap between dialogue context and visual content.

Jailbreak attacks enable malicious queries to evade detection by LLMs. Existing attacks focus on meticulously constructing prompts to disguise harmful intentions. However, the incorporation of sophisticated disguising prompts may incur the challenge of “intention shift”. Intention shift occurs when the additional semantics within the prompt distract the LLMs, causing the responses to deviate significantly from the original harmful intentions. In this paper, we propose a novel component, “bait”, to alleviate the effects of intention shift. Bait comprises an initial response to the harmful query, prompting LLMs to rectify or supplement the knowledge within the bait. By furnishing rich semantics relevant to the query, the bait helps LLMs focus on the original intention. To conceal the harmful content within the bait, we further propose a novel attack paradigm, BaitAttack. BaitAttack adaptively generates necessary components to persuade targeted LLMs that they are engaging with a legitimate inquiry in a safe context. Our proposal is evaluated on a popular dataset, demonstrating state-of-the-art attack performance and an exceptional capability for mitigating intention shift. The implementation of BaitAttack is accessible at: https://anonymous.4open.science/r/BaitAttack-D1F5.