ACL 2024 Papers

Task embedding, a meta-learning technique that captures task-specific information, has gained popularity, especially in areas such as multi-task learning, model editing, and interpretability. However, it faces challenges with the emergence of prompt-guided Large Language Models (LLMs) operating in a gradient-free manner. Existing task embedding methods rely on fine-tuned, task-specific language models, which hinders the adaptability of task embeddings across diverse models, especially prompt-based LLMs. To hardness the potential of task embeddings in the era of LLMs, we propose a framework for unified task embeddings (FUTE), harmonizing task embeddings from various models, including smaller language models and LLMs with varied prompts, within a single vector space. Such uniformity enables comparison and analysis of similarities amongst different models, broadening the scope and utility of existing task embedding methods in multi-model scenarios, while maintaining their performance comparable to architecture-specific methods.

While task-agnostic debiasing provides notable generalizability and reduced reliance on downstream data, its impact on language modeling ability and the risk of relearning social biases from downstream task-specific data remain as the two most significant challenges when debiasing Pretrained Language Models (PLMs). The impact on language modeling ability can be alleviated given a high-quality and long-contextualized debiasing corpus, but there remains a deficiency in understanding the specifics of relearning biases. We empirically ascertain that the effectiveness of task-agnostic debiasing hinges on the quantitative bias level of both the task-specific data used for downstream applications and the debiased model. We empirically show that the lower bound of the bias level of the downstream fine-tuned model can be approximated by the bias level of the debiased model, in most practical cases. To gain more in-depth understanding about how the parameters of PLMs change during fine-tuning due to the forgetting issue of PLMs, we propose a novel framework which can Propagate Socially-fair Debiasing to Downstream Fine-tuning, ProSocialTuning. Our proposed framework can push the fine-tuned model to approach the bias lower bound during downstream fine-tuning, indicating that the ineffectiveness of debiasing can be alleviated by overcoming the forgetting issue through regularizing successfully debiased attention heads based on the PLMs’ bias levels from stages of pretraining and debiasing.

While Language Agents have achieved promising success by placing Large Language Models at the core of a more versatile design that dynamically interacts with the external world, the existing approaches neglect the notion of uncertainty during these interactions. We present the Uncertainty-Aware Language Agent (UALA), a framework that orchestrates the interaction between the agent and the external world using uncertainty quantification. Compared with other well-known counterparts like ReAct, our extensive experiments across 3 representative tasks (HotpotQA, StrategyQA, MMLU) and various LLM sizes demonstrate that UALA brings a significant improvement of performance, while having a substantially lower reliance on the external world (i.e., reduced number of tool calls and tokens). Our analyses provide various insights including the great potential of UALA compared with agent fine-tuning, and underscore the unreliability of verbalised confidence of LLMs as a proxy for uncertainty.

Ensuring the trustworthiness of large language models (LLMs) is crucial. Most studies concentrate on fully pre-trained LLMs to better understand and improve LLMs’ trustworthiness. In this paper, to reveal the untapped potential of pre-training, we pioneer the exploration of LLMs’ trustworthiness during this period, focusing on five key dimensions: reliability, privacy, toxicity, fairness, and robustness. To begin with, we apply linear probing to LLMs. The high probing accuracy suggests that LLMs in early pre-training can already distinguish concepts in each trustworthiness dimension. Therefore, to further uncover the hidden possibilities of pre-training, we extract steering vectors from a LLM’s pre-training checkpoints to enhance the LLM’s trustworthiness. Finally, inspired by the theoretical result that mutual information estimation is bounded by linear probing accuracy, we also probe LLMs with mutual information to investigate the dynamics of trustworthiness during pre-training. We are the first to observe a similar two-phase phenomenon: fitting and compression. This research provides an initial exploration of trustworthiness modeling during LLM pre-training, seeking to unveil new insights and spur further developments in the field.

The rapid advancement of Large Language Models (LLMs) has demonstrated their vast potential across various domains, attributed to their extensive pretraining knowledge and exceptional generalizability. However, LLMs often encounter challenges in generating harmful content when faced with problematic prompts. To address this problem, existing work attempted to implement a gradient ascent based approach to prevent LLMs from producing harmful output. While these methods can be effective, they frequently impact the model utility in responding to normal prompts. To address this gap, we introduce Selective Knowledge negation Unlearning (SKU), a novel unlearning framework for LLMs, designed to eliminate harmful knowledge while preserving utility on normal prompts. Specifically, SKU is consisted of two stages: harmful knowledge acquisition stage and knowledge negation stage. The first stage aims to identify and acquire harmful knowledge within the model, whereas the second is dedicated to remove this knowledge. SKU selectively isolates and removes harmful knowledge in model parameters, ensuring the model’s performance remains robust on normal prompts. Our experiments conducted across various LLM architectures demonstrate that SKU identifies a good balance point between removing harmful information and preserving utility.

Parameter-efficient fine-tuning (PEFT) has enabled the efficient optimization of cumbersome language models in real-world settings. However, as datasets in such environments often contain noisy labels that adversely affect performance, PEFT methods are inevitably exposed to noisy labels. Despite this challenge, the adaptability of PEFT to noisy environments remains underexplored. To bridge this gap, we investigate various PEFT methods under noisy labels. Interestingly, our findings reveal that PEFT has difficulty in memorizing noisy labels due to its inherently limited capacity, resulting in robustness. However, we also find that such limited capacity simultaneously makes PEFT more vulnerable to interference of noisy labels, impeding the learning of clean samples. To address this issue, we propose Clean Routing (CleaR), a novel routing-based PEFT approach that adaptively activates PEFT modules. In CleaR, PEFT modules are preferentially exposed to clean data while bypassing the noisy ones, thereby minimizing the noisy influence. To verify the efficacy of CleaR, we perform extensive experiments on diverse configurations of noisy labels. The results convincingly demonstrate that CleaR leads to substantially improved performance in noisy environments

Knowledge in the real world is being updated constantly. However, it is costly to frequently update large language models (LLMs). Therefore, it is crucial for LLMs to understand the concept of temporal knowledge. However, prior works on temporal question answering (TQA) did not emphasize multi-answer and multi-hop types of temporal reasoning. In this paper, we propose a complex temporal question-answering dataset Complex-TR that focuses on multi-answer and multi-hop temporal reasoning. Besides, we also propose a novel data augmentation strategy to improve the complex temporal reasoning capability and robustness of LLMs. We conducted experiments on multiple temporal QA datasets. Experimental results show that our method is able to improve LLMs’ performance on temporal QA benchmarks by significant margins.

New Intent Discovery (NID) aims at detecting known and previously undefined categories of user intent by utilizing limited labeled and massive unlabeled data. Most prior works often operate under the unrealistic assumption that the distribution of both familiar and new intent classes is uniform, overlooking the skewed and long-tailed distributions frequently encountered in real-world scenarios. To bridge the gap, our work introduces the imbalanced new intent discovery i-NID task, which seeks to identify familiar and novel intent categories within long-tailed distributions. A new benchmark baNID-Bench comprised of three datasets is created to simulate the real-world long-tail distributions. ImbaNID-Bench ranges from broad cross-domain to specific single-domain intent categories, providing a thorough representation of practical use cases. Besides, a robust baseline model ImbaNID is proposed to achieve cluster-friendly intent representations. It includes three stages: model pre-training, generation of reliable pseudo-labels, and robust representation learning that strengthens the model performance to handle the intricacies of real-world data distributions. Our extensive experiments on previous benchmarks and the newly established benchmark demonstrate the superior performance of ImbaNID in addressing the i-NID task, highlighting its potential as a powerful baseline for uncovering and categorizing user intents in imbalanced and long-tailed distributions.

Writing assistance aims to improve the correctness and quality of input texts, with character checking being crucial in detecting and correcting wrong characters. In the real world where handwriting occupies the vast majority, characters that humans get wrong include faked characters (i.e., untrue characters created due to writing errors) and misspelled characters (i.e., true characters used incorrectly due to spelling errors). However, existing datasets and related studies only focus on misspelled characters that can be represented by computer text encoding systems, thereby ignoring faked characters which are more common and difficult. To break through this dilemma, we present Visual-C3, a human-annotated Visual Chinese Character Checking dataset with faked and misspelled Chinese characters. To the best of our knowledge, Visual-C3 is the first real-world visual and the largest human-crafted dataset for the Chinese character checking scenario. Additionally, we also propose and evaluate novel baseline methods on Visual-C3. Extensive empirical results and analyses show that Visual-C3 is high-quality yet challenging. As the first study focusing on Chinese faked characters, the dataset and the baseline methods are publicly available at https://github.com/THUKElab/Visual-C3.

A major impediment to the advancement of sign language translation (SLT) is data scarcity. Much of the sign language data currently available on the web cannot be used for training supervised models due to the lack of aligned captions. Furthermore, scaling SLT using large-scale web-scraped datasets bears privacy risks due to the presence of biometric information, which the responsible development of SLT technologies should account for. In this work, we propose a two-stage framework for privacy-aware SLT at scale that addresses both of these issues. We introduce SSVP-SLT, which leverages self-supervised video pretraining on anonymized and unannotated videos, followed by supervised SLT finetuning on a curated parallel dataset. SSVP-SLT achieves state-of-the-art finetuned and zero-shot gloss-free SLT performance on the How2Sign dataset, outperforming the strongest respective baselines by over 3 BLEU-4. Based on controlled experiments, we further discuss the advantages and limitations of self-supervised pretraining and anonymization via facial obfuscation for SLT.

The advent of large language models has experienced a remarkable improvement in the field of machine translation. However, machine translation is still vulnerable to critical meaning deviations, which may incur catastrophic issues in social or ethical contexts. In particular, existing critical error detection primarily focuses on identifying sentence-level errors, leaving the precise localization of such errors within the sentence unaddressed. In this paper, we introduce a new task, word-level critical error detection (WCED), to detect critical errors at a fine-grained level in machine translation sentences. The task aims to identify the parts of a machine translation that contain catastrophic meaning distortions. We hypothesize that the ability to determine errors at the sentence level will positively influence the detection of more granular errors. We propose a sentence-level error detection module to predict which words in a sentence have critical errors. Experimental results demonstrate that our method outperforms existing methodologies and LLM in En-De, Zh-En, En-Ru, and En-Ko. Our method is helpful for determining the fine-grained location of errors. We hope that such studies will improve the capacity to address critical errors adeptly.

Prominent large language models have exhibited human-level performance in many domains, even enabling the derived agents to simulate human and social interactions. While practical works have substantiated the practicability of grounding language agents in sandbox simulation or embodied simulators, current social intelligence benchmarks either stay at the language level or use subjective metrics. In pursuit of a more realistic and objective evaluation, we introduce the Social Tasks in Sandbox Simulation (STSS) benchmark, which assesses language agents objectively at the action level by scrutinizing the goal achievements within the multi-agent simulation.Additionally, we sample conversation scenarios to build a language-level benchmark to provide an economically prudent preliminary evaluation and align with prevailing benchmarks. To gauge the significance of agent architecture, we implement a target-driven planning (TDP) module as an adjunct to the existing agent. Our evaluative findings highlight that the STSS benchmark is challenging for state-of-the-art language agents. Furthermore, it effectively discriminates between distinct language agents, suggesting its usefulness as a benchmark for evaluating both language models and agent architectures. Our code is available at https://github.com/wcx21/Social-Tasks-in-Sandbox-Simulation.

Recent research has shown a weak correlation between n-gram-based metrics and human evaluations in machine translation task, particularly when evaluating large language models (LLMs). Additionally, the data leakage risk in LLMs may cause an overestimation problem when evaluating LLMs on downstream tasks. In this work, we identify the limited diversity of references as the primary cause for the inferior performance of n-gram-based metrics and the overestimation problem. To address this issue, we propose to utilize multiple references generated by LLMs, coupled with an effective selection strategy focused on accuracy and diversity, to improve the alignment between automatic metrics and human evaluations. We validate our approach on the WMT22 Metrics benchmark with 4 languages and observe a maximum accuracy gain of 9.5% in F200spBLEU, which makes it on par with computationally expensive neural-based metrics. We also show that using multi-reference with n-gram-based metrics significantly alleviates the overestimation problem when evaluating LLMs with data leakage. Further analysis explores the factors that affect the quality of generated references, offering insights into data synthesis by LLMs.

Temporal knowledge graph reasoning has emerged as a crucial task for answering time-dependent questions within a knowledge graph (KG).Despite tremendous progress, the present research is impeded by the sparsity of a temporal KG and an over-reliance on simple single-relational reasoning patterns. To overcome these challenges, we introduce MulQuestions, a new temporal KG reasoning benchmark featuring over 200k entities and 960k questions designed to facilitate complex, multi-relational and multi-hop reasoning. Additionally, we propose a new model adept at conducting pattern-aware and time-sensitive reasoning across temporal KGs. The model’s efficacy is confirmed through rigorous evaluations, showcasing its effectiveness in sparse data conditions and adeptness at handling questions with long reasoning chains. We have made our benchmark and model publicly accessible at [https://anonymous].

Advances towards more faithful and traceable answers of Large Language Models (LLMs) are crucial for various research and practical endeavors. One avenue in reaching this goal is basing the answers on reliable sources. However, this Evidence-Based QA has proven to work insufficiently with LLMs in terms of citing the correct sources (source quality) and truthfully representing the information within sources (answer attributability). In this work, we systematically investigate how to robustly fine-tune LLMs for better source quality and answer attributability. Specifically, we introduce a data generation pipeline with automated data quality filters, which can synthesize diversified high-quality training and testing data at scale. We further introduce four test sets to benchmark the robustness of fine-tuned specialist models. Extensive evaluation shows that fine-tuning on synthetic data improves performance on both in- and out-of-distribution. Furthermore, we show that data quality, which can be drastically improved by proposed quality filters, matters more than quantity in improving Evidence-Based QA.

Tuning-based large language models for machine translation (aka large translation model, LTM) have demonstrated significant performance in the field of machine translation. Despite their success, these models often face difficulties in leveraging demonstrations to further improve their performance. To tackle this challenge, we introduce a novel approach that integrates demonstration-aware training and inference strategies within the framework of tuning-based LTMs, hereby referred to as demonstration-aware LTMs. During training, we enrich the model’s learning process by incorporating both sentence- and document-level demonstrations derived from its original training dataset. During inference, the model synergizes its own contextual translations with retrieved high-quality demonstrations, leading to more precise and contextually appropriate outputs. Empirical results reveal that our demonstration-aware LTM not only mitigates the negative impacts traditionally associated with demonstrations but also secures substantial improvements in translation accuracy, particularly in domain-specific and document-level translation tasks. Source code and scripts are freely available at https://github.com/ChenLi0620/Demo-Aware-LLM-MT.

For the grammatical error correction (GEC) task, there usually exist multiple correction ways for an erroneous input sentence, leading to multiple references. Observing the high proportion of multi-reference instances in Chinese GEC training data, we target a systematic study on how to better utilize multi-reference training data. We propose two new approaches and a simple two-stage training strategy. We compare them against previously proposed approaches, on two Chinese training datasets, i.e., Lang-8 for second language learner texts and FCGEC-Train for native speaker texts, and three test datasets. The experiments and analyses demonstrate the effectiveness of our proposed approaches and reveal interesting insights. Our code is available at https://github.com/ymliucs/MrGEC.

Sentence Representation Learning (SRL) is a crucial task in Natural Language Processing (NLP), where contrastive Self-Supervised Learning (SSL) is currently a mainstream approach. However, the reasons behind its remarkable effectiveness remain unclear. Specifically, many studies have investigated the similarities between contrastive and non-contrastive SSL from a theoretical perspective. Such similarities can be verified in classification tasks, where the two approaches achieve comparable performance. But in ranking tasks (i.e., Semantic Textual Similarity (STS) in SRL), contrastive SSL significantly outperforms non-contrastive SSL. Therefore, two questions arise: First, *what commonalities enable various contrastive losses to achieve superior performance in STS?* Second, *how can we make non-contrastive SSL also effective in STS?* To address these questions, we start from the perspective of gradients and discover that four effective contrastive losses can be integrated into a unified paradigm, which depends on three components: the **Gradient Dissipation**, the **Weight**, and the **Ratio**. Then, we conduct an in-depth analysis of the roles these components play in optimization and experimentally demonstrate their significance for model performance. Finally, by adjusting these components, we enable non-contrastive SSL to achieve outstanding performance in STS.

Event Extraction (EE) is an essential information extraction task that aims to extract event-related information from unstructured texts.The paradigm of this task has shifted from conventional classification-based methods to more contemporary question-answering-based (QA-based) approaches. However, in QA-based EE, the quality of the questions dramatically affects the extraction accuracy, and how to generate high-quality questions for QA-based EE remains a challenge. In this work, to tackle this challenge, we suggest four criteria to evaluate the quality of a question and propose a reinforcement learning method, RLQG, for QA-based EE that can generate generalizable, high-quality, and context-dependent questions and provides clear guidance to QA models. The extensive experiments conducted on ACE and RAMS datasets have strongly validated our approach’s effectiveness, which also demonstrates its robustness in scenarios with limited training data. The corresponding code of RLQG is released for further research.