EMNLP 2024 Papers

This paper explores the pitfalls in evaluating multilingual automatic speech recognition (ASR) models, with a particular focus on Indic language scripts. We investigate the text normalization routine employed by leading ASR models, including OpenAI Whisper, Meta’s MMS, Seamless, and Assembly AI’s Conformer, and their unintended consequences on performance metrics. Our research reveals that current text normalization practices, while aiming to standardize ASR outputs for fair comparison, by removing inconsistencies such as variations in spelling, punctuation, and special characters, are fundamentally flawed when applied to Indic scripts. Through empirical analysis using text similarity scores and in-depth linguistic examination, we demonstrate that these flaws lead to artificially improved performance metrics for Indic languages. We conclude by proposing a shift towards developing text normalization routines that leverage native linguistic expertise, ensuring more robust and accurate evaluations of multilingual ASR models.

This paper presents a way of enhancing the reliability of Large Multi-modal Models (LMMs) in addressing hallucination, where the models generate cross-modal inconsistent responses. Without additional training, we propose Counterfactual Inception, a novel method that implants counterfactual thinking into LMMs using self-generated counterfactual keywords. Our method is grounded in the concept of counterfactual thinking, a cognitive process where human considers alternative realities, enabling more extensive context exploration. Bridging the human cognition mechanism into LMMs, we aim for the models to engage with and generate responses that span a wider contextual scene understanding, mitigating hallucinatory outputs. We further introduce Plausibility Verification Process (PVP), a simple yet robust keyword constraint that effectively filters out sub-optimal keywords to enable the consistent triggering of counterfactual thinking in the model responses. Comprehensive analyses across various LMMs, including both open-source and proprietary models, corroborate that counterfactual thinking significantly reduces hallucination and helps to broaden contextual understanding based on true visual clues.

Leveraging large language models (LLMs) for various natural language processing tasks has led to superlative claims about their performance. For the evaluation of machine translation (MT), existing research shows that LLMs are able to achieve results comparable to fine-tuned multilingual pre-trained language models. In this paper, we explore what translation information, such as the source, reference, translation errors and annotation guidelines, is needed for LLMs to evaluate MT quality. In addition, we investigate prompting techniques such as zero-shot, Chain of Thought (CoT) and few-shot prompting for eight language pairs covering high-, medium- and low-resource languages, leveraging varying LLM variants. Our findings indicate the importance of reference translations for an LLM-based evaluation. While larger models do not necessarily fare better, they tend to benefit more from CoT prompting, than smaller models. We also observe that LLMs do not always provide a numerical score when generating evaluations, which poses a question on their reliability for the task. Our work presents a comprehensive analysis for resource-constrained and training-less LLM-based evaluation of machine translation. We release the accrued prompt templates, code and data publicly for reproducibility.

Fully end-to-end task-oriented dialogue (EToD) systems have shown excellent performance, which requires the ability to retrieve entities accurately for generation. Existing methods improve the accuracy of entity retrieval and construct data flows between retrieval results and response generator, achieving promising results. However, most of them suffer from the following issues: (1) The entity is retrieved by directly interacting with the context at a coarse-grained level, so the similarity score may be disturbed by irrelevant attributes; (2) The generator pays equal attention to retrieved entities and the context and does not learn the generation preferences for the current turn. In this paper, we propose a framework called Regulating Preferences of Generator (RPG) based on retrieval results, which includes a generator preference extractor, an entity retriever, and a generator with the gate-controlled preference regulator. The generator preference extractor not only improves the entity retriever by filtering the interference of irrelevant attributes but also provides more focused guidance to the generator by performing inter-turn attribute prediction. Experiments and analyses on three standard benchmarks show that our framework outperforms existing methods and improves the quality of the dialogue.

Existing script event prediction task forcasts the subsequent event based on an event script chain. However, the evolution of historical events are more complicated in real world scenarios and the limited information provided by the event script chain also make it difficult to accurately predict subsequent events. This paper introduces a Causality Graph Event Prediction(CGEP) task that forecasting consequential event based on an Event Causality Graph (ECG). We propose a Semantic Enhanced Distance-sensitive Graph Prompt Learning (SeDGPL) Model for the CGEP task. In SeDGPL, (1) we design a Distance-sensitive Graph Linearization (DsGL) module to reformulate the ECG into a graph prompt template as the input of a PLM; (2) propose an Event-Enriched Causality Encoding (EeCE) module to integrate both event contextual semantic and graph schema information; (3) propose a Semantic Contrast Event Prediction (ScEP) module to enhance the event representation among numerous candidate events and predict consequential event following prompt learning paradigm. Experiment results validate our argument our proposed SeDGPL model outperforms the advanced competitors for the CGEP task.

Language models often struggle with handling factual knowledge, exhibiting factual hallucination issue. This makes it vital to evaluate the models’ ability to recall its parametric knowledge about facts. In this study, we introduce a knowledge probing benchmark, BELIEF(ICL), to evaluate the knowledge recall ability of both encoder- and decoder-based pre-trained language models (PLMs) from diverse perspectives. BELIEFs utilize a multi-prompt dataset to evaluate PLM’s accuracy, consistency, and reliability in factual knowledge recall. To enable a more reliable evaluation with BELIEFs, we semi-automatically create MyriadLAMA, which has massively diverse prompts. We validate the effectiveness of BELIEFs in comprehensively evaluating PLM’s knowledge recall ability on diverse PLMs, including recent large language models (LLMs). We then investigate key factors in memorizing and recalling facts in PLMs, such as model size, pretraining strategy and corpora, instruction-tuning process and in-context learning settings. Finally, we reveal the limitation of the prompt-based knowledge probing. The MyriadLAMA is publicized.

Language models (LM) are capable of remarkably complex linguistic tasks; however, numerical reasoning is an area in which they frequently struggle. An important but rarely evaluated form of reasoning is understanding probability distributions. In this paper, we focus on evaluating the probabilistic reasoning capabilities of LMs using idealized and real-world statistical distributions. We perform a systematic evaluation of state-of-the-art LMs on three tasks: estimating percentiles, drawing samples, and calculating probabilities. We evaluate three ways to provide context to LMs 1) anchoring examples from within a distribution or family of distributions, 2) real-world context, 3) summary statistics on which to base a Normal approximation. Models can make inferences about distributions, and can be further aided by the incorporation of real-world context, example shots and simplified assumptions, even if these assumptions are incorrect or misspecified. To conduct this work, we developed a comprehensive benchmark distribution dataset with associated question-answer pairs that we have released publicly.

Web agents are emerging as powerful tools capable of performing complex tasks across diverse web environments. The rapid development of large multimodal models is further enhancing this advancement. However, there is a lack of standardized and user-friendly tools for research and development, as well as experimental platforms on live websites. To address this challenge, we present WebOlympus, an open platform for web agents operating on live websites. WebOlympus offers a Chrome extension-based UI, enabling users without programming experience to easily utilize the platform. It allows users to run web agents with various designs using only a few lines of code or simple clicks on the Chrome extension. To ensure the trustworthiness of web agents, a safety monitor module that prevents harmful actions through human supervision or model-based control is incorporated. WebOlympus supports diverse applications, including annotation interfaces for web agent trajectories and data crawling.

When large language models (LLMs) surpass human capabilities, supervising them effectively becomes difficult. Weak-to-strong learning, where a less capable model enhances a stronger one, proves valuable in this context. Yet, the efficacy of this paradigm for complex reasoning tasks is still unexplored. In this paper, we introduce a progressive weak-to-strong reasoning framework that enables the strong model to autonomously refine its training data, maximizing the use of weak signals and unlocking its latent abilities. This framework begins with supervised fine-tuning on a selective small but high-quality dataset, followed by preference optimization on contrastive samples identified by the strong model itself. Experiments on the GSM8K and MATH datasets verify that our method can effectively improve the reasoning capabilities of Llama2-70b using three separate weak models. This work paves the way for a more scalable and sophisticated strategy to enhance AI reasoning powers. All relevant code and resources are available in https://github.com/GAIR-NLP/weak-to-strong-reasoning.

The inherent ambiguity of cause and effect boundaries poses a challenge in evaluating causal event extraction tasks. Traditional metrics like Exact Match and BertScore poorly reflect model performance, so we trained evaluation models to approximate human evaluation, achieving high agreement. We used them to perform Reinforcement Learning with extraction models to align them with human preference, prioritising semantic understanding. We successfully explored our approach through multiple datasets, including transferring an evaluator trained on one dataset to another as a way to decrease the reliance on human-annotated data. In that vein, we also propose a weak-to-strong supervision method that uses a fraction of the annotated data to train an evaluation model while still achieving high performance in training an RL model.

Recent advancements in large language models (LLMs) have expanded their scope in natural language processing (NLP) to encompass multimodal functions. However, integrating listening capabilities effectively remains a significant challenge for generalization and complex auditory task execution. In this work, we introduce WavLLM, a robust and adaptive speech large language model featuring dual encoders—a Whisper encoder for semantics and a WavLM encoder for speaker characteristics. Within the two-stage curriculum learning framework, WavLLM first builds its foundational capabilities by optimizing on mixed elementary single tasks, followed by advanced multi-task training on more complex tasks such as combinations of the elementary tasks. To enhance the flexibility and adherence to different tasks and instructions, a prompt-aware LoRA weight adapter is introduced in the second advanced multi-task training stage. We validate the proposed model on universal speech benchmarks and also apply it to specialized speech-question-answer (SQA) dataset, and speech Chain-of-Thought (CoT) evaluation set. Experiments demonstrate that the proposed model achieves state-of-the-art performance across a range of speech tasks on the same model size, exhibiting robust generalization capabilities in executing complex tasks using CoT approach. The codes, models, audio samples, and SQA evaluation set can be accessed at https://github.com/microsoft/SpeechT5/tree/main/WavLLM.

Protecting intellectual property (IP) of text such as articles and code is increasingly important, especially as sophisticated attacks become possible, such as paraphrasing by large language models (LLMs) or even unauthorized training of LLMs on copyrighted text to infringe such IP. However, existing text watermarking methods are not robust enough against such attacks nor scalable to millions of users for practical implementation. In this paper, we propose Waterfall, the first training-free framework for robust and scalable text watermarking applicable across multiple text types (e.g., articles, code) and languages supportable by LLMs, for general text and LLM data provenance. Waterfall comprises several key innovations, such as being the first to use LLM as paraphrasers for watermarking along with a novel combination of techniques that are surprisingly effective in achieving robust verifiability and scalability. We empirically demonstrate that Waterfall achieves significantly better scalability, robust verifiability, and computational efficiency compared to SOTA article-text watermarking methods, and also showed how it could be directly applied to the watermarking of code.

Large language model agents have exhibited exceptional performance across a range of complex interactive tasks. Recent approaches have utilized tuning with expert trajectories to enhance agent performance, yet they primarily concentrate on outcome rewards, which may lead to errors or suboptimal actions due to the absence of process supervision signals. In this paper, we introduce the **I**terative step-level **P**rocess **R**efinement **(IPR)** framework, which provides detailed step-by-step guidance to enhance agent training. Specifically, we adopt the Monte Carlo method to estimate step-level rewards. During each iteration, the agent explores along the expert trajectory and generates new actions. These actions are then evaluated against the corresponding step of expert trajectory using step-level rewards. Such comparison helps identify discrepancies, yielding contrastive action pairs that serve as training data for the agent. Our experiments on three complex agent tasks demonstrate that our framework outperforms a variety of strong baselines. Moreover, our analytical finds highlight the effectiveness of IPR in augmenting action efficiency and its applicability to diverse models.

WalledEval is a comprehensive AI safety testing toolkit designed to evaluate large language models (LLMs). It accommodates a diverse range of models, including both open-weight and API-based ones, and features over 35 safety benchmarks covering areas such as multilingual safety, exaggerated safety, and prompt injections. The framework supports both LLM and judge benchmarking, and incorporates custom mutators to test safety against various text-style mutations such as future tense and paraphrasing. Additionally, WalledEval introduces WalledGuard, a new, small and performant content moderation tool, and SGXSTest, a benchmark for assessing exaggerated safety in cultural contexts. We make WalledEval publicly available at https://github.com/walledai/walledeval with a demonstration video at https://youtu.be/50Zy97kj1MA.

The common toxicity and societal bias in contents generated by large language models (LLMs) necessitate strategies to reduce harm. Present solutions often demand white-box access to the model or substantial training, which is impractical for cutting-edge commercial LLMs. Moreover, prevailing prompting methods depend on external tool feedback and fail to simultaneously lessen toxicity and bias. Motivated by social psychology principles, we propose a novel strategy named perspective-taking prompting (PeT) that inspires LLMs to integrate diverse human perspectives and self-regulate their responses. This self-correction mechanism can significantly diminish toxicity (up to 89%) and bias (up to 73%) in LLMs’ responses. Rigorous evaluations and ablation studies are conducted on two commercial LLMs (ChatGPT and GLM) and three open-source LLMs, revealing PeT’s superiority in producing less harmful responses, outperforming five strong baselines.

Large language models (LLMs) have received a lot of attention in natural language processing (NLP) research because of their exceptional performance in understanding and generating human languages. However, low-resource languages are left behind due to the unavailability of resources. In this work, we focus on enhancing the LLaMA-2-Amharic model by integrating task-specific and generative datasets to improve language model performance for Amharic. We compile an Amharic instruction fine-tuning dataset and fine-tuned LLaMA-2-Amharic model. The fine-tuned model shows promising results in different NLP tasks. We also explore the effectiveness of translated instruction datasets compared to the dataset we created. Our dataset creation pipeline, along with instruction datasets, trained models, and evaluation outputs, is made publicly available to encourage research in language-specific models.

Reinforcement learning from human feedback (RLHF) is a promising solution to align large language models (LLMs) more closely with human values. Off-policy preference optimization, where the preference data is obtained from other models, is widely adopted due to its cost efficiency and scalability. However, off-policy preference optimization often suffers from a distributional gap between the policy used for data collection and the target policy, leading to suboptimal optimization. In this paper, we propose a novel strategy to mitigate this problem by simulating on-policy learning with off-policy preference data. Our Weighted Preference Optimization (WPO) method adapts off-policy data to resemble on-policy data more closely by reweighting preference pairs according to their probability under the current policy. This method not only addresses the distributional gap problem but also enhances the optimization process without incurring additional costs. We validate our method on instruction following benchmarks including Alpaca Eval 2 and MT-bench. WPO not only outperforms Direct Preference Optimization (DPO) by up to 5.6% on Alpaca Eval 2 but also establishes a remarkable length-controlled winning rate against GPT-4-turbo of 76.7% based on Gemma-2-9b-it. We release the code and models at https://github.com/wzhouad/WPO.

Language models have been shown to reproduce underlying biases existing in their training data, which is the majority perspective by default. Proposed solutions aim to capture minority perspectives by either modelling annotator disagreements or grouping annotators based on shared metadata, both of which face significant challenges. We propose a framework that trains models without encoding annotator metadata, extracts latent embeddings informed by annotator behaviour, and creates clusters of similar opinions, that we refer to as voices. Resulting clusters are validated post-hoc via internal and external quantitative metrics, as well a qualitative analysis to identify the type of voice that each cluster represents. Our results demonstrate the strong generalisation capability of our framework, indicated by resulting clusters being adequately robust, while also capturing minority perspectives based on different demographic factors throughout two distinct datasets.

Yoruba—an African language with roughly 47 million speakers—encompasses a continuum with several dialects. Recent efforts to develop NLP technologies for African languages have focused on their standard dialects, resulting in disparities for dialects and varieties for which there are little to no resources or tools. We take steps towards bridging this gap by introducing a new high-quality parallel text and speech corpus; YORULECT across three domains and four regional yoruba dialects. To develop this corpus, we engaged native speakers, traveling to communities where these dialects are spoken, to collect text and speech data. Using our newly created corpus, we conducted extensive experiments on (text) machine translation, automatic speech recognition, and speech-to-text translation. Our results reveal substantial performance disparities between standard yoruba and the other dialects across all tasks. However, we also show that with dialect-adaptive finetuning, we are able to narrow this gap. We believe our dataset and experimental analysis will contribute greatly to developing NLP tools for Yoruba and its dialects, and potentially for other African languages, by improving our understanding of existing challenges and offering a high-quality dataset for further development. We will release YORULECT dataset and models publicly under an open license.