论文检索

Michiel van der Meer, Neele Falk, Pradeep K. Murukannaiah, Enrico Liscio

Active Learning (AL) addresses the high costs of collecting human annotations by strategically annotating the most informative samples. However, for subjective NLP tasks, incorporating a wide range of perspectives in the annotation process is crucial to capture the variability in human judgments. We introduce Annotator-Centric Active Learning (ACAL), which incorporates an annotator selection strategy following data sampling. Our objective is two-fold: (1) to efficiently approximate the full diversity of human judgments, and (2) to assess model performance using annotator-centric metrics, which value minority and majority perspectives equally. We experiment with multiple annotator selection strategies across seven subjective NLP tasks, employing both traditional and novel, human-centered evaluation metrics. Our findings indicate that ACAL improves data efficiency and excels in annotator-centric performance evaluations. However, its success depends on the availability of a sufficiently large and diverse pool of annotators to sample from.

Asif Hanif, Maha Tufail Agro, Mohammad Areeb Qazi, Hanan Aldarmaki

Audio-Language Models (ALMs) have recently achieved remarkable success in zero-shot audio recognition tasks, which match features of audio waveforms with class-specific text prompt features, inspired by advancements in Vision-Language Models (VLMs). Given the sensitivity of zero-shot performance to the choice of hand-crafted text prompts, many prompt learning techniques have been developed for VLMs. We explore the efficacy of these approaches in ALMs and propose a novel method, Prompt Learning in Audio Language Models (PALM), which optimizes the feature space of the text encoder branch. Unlike existing methods that work in the input space, our approach results in greater training efficiency. We demonstrate the effectiveness of our approach on 11 audio recognition datasets, encompassing a variety of speech-processing tasks, and compare the results with three baselines in a few-shot learning setup. Our method is either on par with or outperforms other approaches while being computationally less demanding. Our code is publicly available at https://asif-hanif.github.io/palm/.

Aida Davani, Mark Díaz, Dylan Baker, Vinodkumar Prabhakaran

While human annotations play a crucial role in language technologies, annotator subjectivity has long been overlooked in data collection. Recent studies that critically examine this issue are often focused on Western contexts, and solely document differences across age, gender, or racial groups. Consequently, NLP research on subjectivity have failed to consider that individuals within demographic groups may hold diverse values, which influence their perceptions beyond group norms. To effectively incorporate these considerations into NLP pipelines, we need datasets with extensive parallel annotations from a variety of social and cultural groups.In this paper we introduce the D3CODE dataset: a large-scale cross-cultural dataset of parallel annotations for offensive language in over 4.5K English sentences annotated by a pool of more than 4k annotators, balanced across gender and age, from across 21 countries, representing eight geo-cultural regions. The dataset captures annotators’ moral values along six moral foundations: care, equality, proportionality, authority, loyalty, and purity. Our analyses reveal substantial regional variations in annotators’ perceptions that are shaped by individual moral values, providing crucial insights for developing pluralistic, culturally sensitive NLP models.

Linhao Zhang, Jintao Liu, Li Jin, Hao Wang, Kaiwen Wei, Guangluan Xu

The illustration or visualization of figurative language, such as linguistic metaphors, is an emerging challenge for existing Large Language Models (LLMs) and multimodal models. Due to their comparison of seemingly unrelated concepts in metaphors, existing LLMs have a tendency of over-literalization, which illustrates figurative language solely based on literal objects, ignoring the underlying groundings and associations across disparate metaphorical domains. Furthermore, prior approaches have ignored the binding process between visual objects and metaphorical attributes, which further intensifies the infidelity of visual metaphors. To address the issues above, we propose GOME (Grounding-based Metaphor Binding), which illustrates linguistic metaphors from the grounding perspective elaborated through LLMs. GOME consists of two steps for metaphor illustration, including grounding-based elaboration and scenario visualization. In the elaboration step, metaphorical knowledge is integrated into systematic instructions for LLMs, which employs a CoT prompting method rooted in rhetoric. This approach specifies metaphorical devices such as vehicles and groundings, to ensure accurate and faithful descriptions consumed by text-to-image models. In the visualization step, an inference-time metaphor binding method is realized based on elaboration outputs, which register attentional control during the diffusion process, and captures the underlying attributes from the abstract metaphorical domain. Comprehensive evaluations using multiple downstream tasks confirm that, GOME is superior to isolated LLMs, diffusion models, or their direct collaboration.

Alessio Devoto, Yu Zhao, Simone Scardapane, Pasquale Minervini

The deployment of large language models (LLMs) is often hindered by the extensive memory requirements of the Key-Value (KV) cache, especially as context lengths increase. Existing approaches to reduce the KV cache size involve either fine-tuning the model to learn a compression strategy or leveraging attention scores to reduce the sequence length. We analyse the attention distributions in decoder-only Transformers-based models and observe that attention allocation patterns stay consistent across most layers. Surprisingly, we find a clear correlation between the L_2 norm and the attention scores over cached KV pairs, where a low L_2 norm of a key embedding usually leads to a high attention score during decoding. This finding indicates that the influence of a KV pair is potentially determined by the key embedding itself before being queried. Based on this observation, we compress the KV cache based on the L_2 norm of key embeddings. Our experimental results show that this simple strategy can reduce the KV cache size by 50% on language modelling and needle-in-a-haystack tasks and 90% on passkey retrieval tasks without losing accuracy. Moreover, without relying on the attention scores, this approach remains compatible with FlashAttention, enabling broader applicability.

Ekaterina Sviridova, Anar Yeginbergen, Ainara Estarrona, Elena Cabrio, Serena Villata, Rodrigo Agerri

Explaining Artificial Intelligence (AI) decisions is a major challenge nowadays in AI, in particular when applied to sensitive scenarios like medicine and law. However, the need to explain the rationale behind decisions is a main issues also for human-based deliberation as it is important to justify why a certain decision has been taken. Resident medical doctors for instance are required not only to provide a (possibly correct) diagnosis, but also to explain how they reached a certain conclusion. Developing new tools to aid residents to train their explanation skills is therefore a central objective of AI in education. In this paper, we follow this direction, and we present, to the best of our knowledge, the first multilingual dataset for Medical Question Answering where correct and incorrect diagnoses for a clinical case are enriched with a natural language explanation written by doctors. These explanations have been manually annotated with argument components (i.e., premise, claim) and argument relations (i.e., attack, support). The Multilingual CasiMedicos-arg dataset consists of 558 clinical cases (English, Spanish, French, Italian) with explanations, where we annotated 5021 claims, 2313 premises, 2431 support relations, and 1106 attack relations. We conclude by showing how competitive baselines perform over this challenging dataset for the argument mining task.

Zijin Feng, Luyang Lin, Lingzhi Wang, Hong Cheng, Kam-Fai Wong

Text clustering is a fundamental task in natural language processing with numerous applications. However, traditional clustering methods often struggle with domain-specific fine-tuning and the presence of outliers. To address these challenges, we introduce LLMEdgeRefine, an iterative clustering method enhanced by large language models (LLMs), focusing on edge points refinement. LLMEdgeRefine enhances current clustering methods by creating super-points to mitigate outliers and iteratively refining clusters using LLMs for improved semantic coherence. Our method demonstrates superior performance across multiple datasets, outperforming state-of-the-art techniques, and offering robustness, adaptability, and cost-efficiency for diverse text clustering applications.

Gabriele Merlin, Mariya Toneva

Pretrained language models have been shown to significantly predict brain recordings of people comprehending language. Recent work suggests that the prediction of the next word is a key mechanism that contributes to this alignment. What is not yet understood is whether prediction of the next word is necessary for this observed alignment or simply sufficient, and whether there are other shared mechanisms or information that are similarly important. In this work, we take a step towards understanding the reasons for brain alignment via two simple perturbations in popular pretrained language models. These perturbations help us design contrasts that can control for different types of information. By contrasting the brain alignment of these differently perturbed models, we show that improvements in alignment with brain recordings are due to more than improvements in next-word prediction and word-level information.

Yao Xu, Shizhu He, Jiabei Chen, Zihao Wang, Yangqiu Song, Hanghang Tong, Guang Liu, Jun Zhao, Kang Liu

To address the issues of insufficient knowledge and hallucination in Large Language Models (LLMs), numerous studies have explored integrating LLMs with Knowledge Graphs (KGs). However, these methods are typically evaluated on conventional Knowledge Graph Question Answering (KGQA) with complete KGs, where all factual triples required for each question are entirely covered by the given KG. In such cases, LLMs primarily act as an agent to find answer entities within the KG, rather than effectively integrating the internal knowledge of LLMs and external knowledge sources such as KGs. In fact, KGs are often incomplete to cover all the knowledge required to answer questions. To simulate these real-world scenarios and evaluate the ability of LLMs to integrate internal and external knowledge, we propose leveraging LLMs for QA under Incomplete Knowledge Graph (IKGQA), where the provided KG lacks some of the factual triples for each question, and construct corresponding datasets. To handle IKGQA, we propose a training-free method called Generate-on-Graph (GoG), which can generate new factual triples while exploring KGs. Specifically, GoG performs reasoning through a Thinking-Searching-Generating framework, which treats LLM as both Agent and KG in IKGQA. Experimental results on two datasets demonstrate that our GoG outperforms all previous methods.

Elias Bassani, Ignacio Sanchez

Generative AI systems powered by Large Language Models have become increasingly popular in recent years. Lately, due to the risk of providing users with unsafe information, the adoption of those systems in safety-critical domains has raised significant concerns. To respond to this situation, input-output filters, commonly called guardrail models, have been proposed to complement other measures, such as model alignment. Unfortunately, the lack of a standard benchmark for guardrail models poses significant evaluation issues and makes it hard to compare results across scientific publications. To fill this gap, we introduce GuardBench, a large-scale benchmark for guardrail models comprising 40 safety evaluation datasets. To facilitate the adoption of GuardBench, we release a Python library providing an automated evaluation pipeline built on top of it. With our benchmark, we also share the first large-scale prompt moderation datasets in German, French, Italian, and Spanish. To assess the current state-of-the-art, we conduct an extensive comparison of recent guardrail models and show that a general-purpose instruction-following model of comparable size achieves competitive results without the need for specific fine-tuning.

Elie Antoine, Frederic Bechet, Géraldine Damnati, Philippe Langlais

We introduce an evaluation methodology for reading comprehension tasks based on the intuition that certain examples, by the virtue of their linguistic complexity, consistently yield lower scores regardless of model size or architecture. We capitalize on semantic frame annotation for characterizing this complexity, and study seven complexity factors that may account for model’s difficulty. We first deploy this methodology on a carefully annotated French reading comprehension benchmark showing that two of those complexity factors are indeed good predictors of models’ failure, while others are less so. We further deploy our methodology on a well studied English benchmark by using chatGPT as a proxy for semantic annotation.Our study reveals that fine-grained linguistically-motivated automatic evaluation of a reading comprehension task is not only possible, but helps understand models’ abilities to handle specific linguistic characteristics of input examples. It also shows that current state-of-the-art models fail with some for those characteristics which suggests that adequately handling them requires more than merely increasing model size.

Tiago Pimentel, Clara Meister

Language models (LMs) estimate a probability distribution over strings in a natural language; these distributions are crucial for computing perplexity and surprisal in linguistics research. While we are usually concerned with measuring these values for words, most LMs operate over subwords. Despite seemingly straightforward, accurately computing probabilities over one unit given probabilities over the other requires care. Indeed, we show here that many recent linguistic studies have been incorrectly computing these values. This paper derives the correct methods for computing word probabilities, highlighting issues when relying on language models that use beginning-of-word (bow)-marking tokenisers, e.g., the GPT family. Empirically, we show that correcting the widespread bug in probability computations affects measured outcomes in sentence comprehension and lexical optimisation analyses.

Esra Dönmez, Thang Vu, Agnieszka Falenska

Offensive speech is highly prevalent on online platforms. Being trained on online data, Large Language Models (LLMs) display undesirable behaviors, such as generating harmful text or failing to recognize it. Despite these shortcomings, the models are becoming a part of our everyday lives by being used as tools for information search, content creation, writing assistance, and many more. Furthermore, the research explores using LLMs in applications with immense social risk, such as late-life companions and online content moderators. Despite the potential harms from LLMs in such applications, whether LLMs can reliably identify offensive speech and how they behave when they fail are open questions. This work addresses these questions by probing sixteen widely used LLMs and showing that most fail to identify (non-)offensive online language. Our experiments reveal undesirable behavior patterns in the context of offensive speech detection, such as erroneous response generation, over-reliance on profanity, and failure to recognize stereotypes. Our work highlights the need for extensive documentation of model reliability, particularly in terms of the ability to detect offensive language.

Qinzhuo Wu, Wei Liu, Jian Luan, Bin Wang

Recently, tool-augmented LLMs have gained increasing attention. Given an instruction, tool-augmented LLMs can interact with various external tools in multiple rounds and provide a final answer. However, previous LLMs were trained on overly detailed instructions, which included API names or parameters, while real users would not explicitly mention these API details. This leads to a gap between trained LLMs and real-world scenarios. In addition, most works ignore whether the interaction process follows the instruction. To address these issues, we constructed a training dataset called MGToolBench, which contains statement and category-level instructions to better reflect real-world scenarios. In addition, we propose ToolPlanner, a two-stage reinforcement learning framework that utilizes path planning and two feedback mechanisms to enhance the LLM’s task completion and instruction-following capabilities. Experimental results show that ToolPlanner significantly improves the Match Rate, Pass Rate and Win Rate by 26.8%, 20.2%, and 5.6% compared to the SOTA model. Human evaluation verifies that the multi-granularity instructions can better align with users’ usage habits. Our data and code will be released upon acceptance.

Matthew Raffel, Victor Agostinelli, Lizhong Chen

Large language models (LLMs) have achieved state-of-the-art performance in various language processing tasks, motivating their adoption in simultaneous translation. Current fine-tuning methods to adapt LLMs for simultaneous translation focus on prompting optimization strategies using either data augmentation or prompt structure modifications. However, these methods suffer from several issues, such as unnecessarily expanded training sets, computational inefficiency from dumping the key and value cache, increased prompt sizes, or restriction to a single decision policy. To eliminate these issues, in this work, we propose SimulMask, a new paradigm for fine-tuning LLMs for simultaneous translation. It utilizes a novel attention mask approach that models simultaneous translation during fine-tuning by masking attention for a desired decision policy. Applying the proposed SimulMask on a Falcon LLM for the IWSLT 2017 dataset, we have observed a significant translation quality improvement compared to state-of-the-art prompting optimization strategies on five language pairs while reducing the computational cost.

Yufang Liu, Tao Ji, Changzhi Sun, Yuanbin Wu, Aimin Zhou

Large Vision-Language Models (LVLMs) have achieved impressive performance, yet research has pointed out a serious issue with object hallucinations within these models. However, there is no clear conclusion as to which part of the model these hallucinations originate from. In this paper, we present an in-depth investigation into the object hallucination problem specifically within the CLIP model, which serves as the backbone for many state-of-the-art vision-language systems. We unveil that even in isolation, the CLIP model is prone to object hallucinations, suggesting that the hallucination problem is not solely due to the interaction between vision and language modalities. To address this, we propose a counterfactual data augmentation method by creating negative samples with a variety of hallucination issues. We demonstrate that our method can effectively mitigate object hallucinations for CLIP model, and we show the the enhanced model can be employed as a visual encoder, effectively alleviating the object hallucination issue in LVLMs.

YongKang Liu, Yiqun Zhang, Qian Li, Tong Liu, Shi Feng, Daling Wang, Yifei Zhang, Hinrich Schuetze

Full-parameter fine-tuning (FPFT) has become the go-to choice for adapting language models (LMs) to downstream tasks due to its excellent performance. As LMs grow in size, fine-tuning the full parameters of LMs requires a prohibitively large amount of GPU memory. Existing approaches utilize zeroth-order optimizer to conserve GPU memory, which potentially compromises the performance of LMs as non-zero order optimizers tend to converge more readily on most downstream tasks. We propose a novel, memory-efficient, optimizer-independent, end-to-end hierarchical fine-tuning strategy, HiFT, which only updates a subset of parameters at each training step. HiFT significantly reduces the amount of gradients and optimizer state parameters residing in GPU memory at the same time, thereby reducing GPU memory usage. Our results demonstrate that: (1) HiFT achieves comparable performance with parameter-efficient fine-tuning and standard FPFT. (2) Results on six models show that HiFT reduces the number of trainable parameters by about 89.18% on average compared to FPFT. (3) HiFT supports FPFT of 7B models for 24G GPU memory devices under mixed precision without using any memory saving techniques. (4) HiFT supports various optimizers including AdamW, AdaGrad, SGD, etc. The source code link is https://github.com/misonsky/HiFT.

Nikhil Kandpal, Krishna Pillutla, Alina Oprea, Peter Kairouz, Christopher A. Choquette-Choo, Zheng Xu

Text written by humans makes up the vast majority of the data used to pre-train and fine-tune large language models (LLMs). Many sources of this data—like code, forum posts, personal websites, and books—are easily attributed to one or a few “users”. In this paper, we ask if it is possible to infer if any of a _user’s_ data was used to train an LLM. Not only would this constitute a breach of privacy, but it would also enable users to detect when their data was used for training. We develop the first effective attacks for _user inference_—at times, with near-perfect success—against LLMs. Our attacks are easy to employ, requiring only black-box access to an LLM and a few samples from the user, which _need not be the ones that were trained on_. We find, both theoretically and empirically, that certain properties make users more susceptible to user inference: being an outlier, having highly correlated examples, and contributing a larger fraction of data. Based on these findings, we identify several methods for mitigating user inference including training with example-level differential privacy, removing within-user duplicate examples, and reducing a user’s contribution to the training data. Though these provide partial mitigation, our work highlights the need to develop methods to fully protect LLMs from user inference.

Chao-Wei Huang, Yun-Nung Chen

Effective information retrieval (IR) from vast datasets relies on advanced techniques to extract relevant information in response to queries. Recent advancements in dense retrieval have showcased remarkable efficacy compared to traditional sparse retrieval methods. To further enhance retrieval performance, knowledge distillation techniques, often leveraging robust cross-encoder rerankers, have been extensively explored. However, existing approaches primarily distill knowledge from pointwise rerankers, which assign absolute relevance scores to documents, thus facing challenges related to inconsistent comparisons. This paper introduces Pairwise Relevance Distillation (PairDistill) to leverage pairwise reranking, offering fine-grained distinctions between similarly relevant documents to enrich the training of dense retrieval models. Our experiments demonstrate that PairDistill outperforms existing methods, achieving new state-of-the-art results across multiple benchmarks. This highlights the potential of PairDistill in advancing dense retrieval techniques effectively. Our source code and trained models are released at https://github.com/MiuLab/PairDistill

Junteng Liu, Shiqi Chen, Yu Cheng, Junxian He

While large language models (LLMs) have demonstrated remarkable abilities across various fields, hallucination remains a significant challenge. Recent studies have explored hallucinations through the lens of internal representations, proposing mechanisms to decipher LLMs’ adherence to facts. However, these approaches often fail to generalize to out-of-distribution data, leading to concerns about whether internal representation patterns reflect fundamental factual awareness, or only overfit spurious correlations on the specific datasets. In this work, we investigate whether a universal truthfulness hyperplane that distinguishes the model’s factually correct and incorrect outputs exists within the model. To this end, we scale up the number of training datasets and conduct an extensive evaluation – we train the truthfulness hyperplane on a diverse collection of over 40 datasets and examine its cross-task, cross-domain, and in-domain generalization. Our results indicate that increasing the diversity of the training datasets significantly enhances the performance in all scenarios, while the volume of data samples plays a less critical role. This finding supports the optimistic hypothesis that a universal truthfulness hyperplane may indeed exist within the model, offering promising directions for future research.