ACL 2024 Papers

Citing pertinent literature is pivotal to writing and reviewing a scientific document. Existing techniques mainly focus on the local context or the global context for recommending citations but fail to consider the actual human citation behaviour. We propose SymTax, a three-stage recommendation architecture that considers both the local and the global context, and additionally the taxonomical representations of query-candidate tuples and the Symbiosis prevailing amongst them. SymTax learns to embed the infused taxonomies in the hyperbolic space and uses hyperbolic separation as a latent feature to compute query-candidate similarity. We build a novel and large dataset ArSyTa containing 8.27 million citation contexts and describe the creation process in detail. We conduct extensive experiments and ablation studies to demonstrate the effectiveness and design choice of each module in our framework. Also, combinatorial analysis from our experiments shed light on the choice of language models (LMs) and fusion embedding, and the inclusion of section heading as a signal. Our proposed module that captures the symbiotic relationship solely leads to performance gains of 26.66% and 39.25% in Recall@5 w.r.t. SOTA on ACL-200 and RefSeer datasets, respectively. The complete framework yields a gain of 22.56% in Recall@5 wrt SOTA on our proposed dataset. The code and dataset are available at https://github.com/goyalkaraniit/SymTax.

Semantic Parsing of natural language questions into their executable logical form (LF) has shown state-of-the-art (SOTA) performance for Knowledge Graph Question Answering (KGQA). However, these methods are not applicable for real-world applications, due to lack of KG-specific training data. Recent advances in the capabilities of Large Language Models (LLMs) has led towards generating low-level LFs such as SPARQL and S-Expression in a few-shot setting. Unfortunately, these methods: (1) are limited to the knowledge of underlying LLM about the LF, (2) performs inferior for the harder complex benchmarks such as KQA Pro, (3) suffers while grounding the generated LF to a specific Knowledge Graph. Recently, a new LF called KoPL has been introduced that explicitly models complex reasoning process step-by-step in a symbolic manner and has shown SOTA on KQA Pro in fully-supervised setting. Inspired by this, we propose SymKGQA framework that generates step-by-step Symbolic LF i.e., KoPL in a few-shot in-context learning setting using LLM. Our framework is not dependent on pre-trained information of LLM about KoPL. We further build a Retrieval-Augmented Generation based Question-Aware Contextual KoPL (QUACK) resolver to ground the generated LF. Our experiments with different LLMs and few-shot settings demonstrate that SymKGQA outperforms all other few-shot and even many of the fully-supervised KGQA approaches.

Automated teaching assistants and chatbots have significant potential to reduce the workload of human instructors, especially for logistics-related question answering, which is important to students yet repetitive for instructors. However, due to privacy concerns, there is a lack of publicly available datasets. We introduce SyllabusQA, an open-source dataset with 63 real course syllabi covering 36 majors, containing 5,078 open-ended course logistics-related question-answer pairs that are diverse in both question types and answer formats. Since many logistics-related questions contain critical information like the date of an exam, it is important to evaluate the factuality of answers. We benchmark several strong baselines on this task, from large language model prompting to retrieval-augmented generation. We introduce Fact-QA, an LLM-based (GPT-4) evaluation metric to evaluate the factuality of predicted answers. We find that despite performing close to humans on traditional metrics of textual similarity, there remains a significant gap between automated approaches and humans in terms of fact precision.

Mixture of experts (MoE) is a popular technique to improve capacity of Large Language Models (LLMs) with conditionally-activated parallel experts. However, serving MoE models on memory-constrained devices is challenging due to the large parameter size. Typical solutions such as memory swapping or expert pruning may lead to significantly higher latency or severe accuracy loss.In this paper, we introduce SwapMoE, a framework for efficient serving of MoE-based large language models with tunable memory budgets. The main idea of SwapMoE is to keep a small dynamic set of important experts, namely Virtual Experts, in the main memory for inference, while seamlessly maintaining how the Virtual Experts map to the actual experts. Experiments have shown that SwapMoE can reduce the memory footprint while maintaining reasonable accuracy. For example, on text summarization tasks with Switch Transformer, SwapMoE can reduce the memory consumption from 14.2 GiB to 4.7 GiB, together with 50% latency reduction and a slight Rouge-2 score drop of 0.041.

Low-rank approximations, of the weight and feature space can enhance the performance of deep learning models, whether in terms of improving generalization or reducing the latency of inference. However, there is no clear consensus yet on how, when and why these approximations are helpful for large language models (LLMs). In this work, we empirically study the efficacy of weight and feature space decomposition in transformer-based LLMs. We demonstrate that surgical decomposition not only provides critical insights into the trade-off between compression and language modelling performance, but also sometimes enhances commonsense reasoning performance of LLMs. Our empirical analysis identifies specific network segments that intrinsically exhibit a low-rank structure. Furthermore, we extend our investigation to the implications of low-rank approximations on model bias. Overall, our findings offer a novel perspective on optimizing LLMs, presenting the low-rank approximation not only as a tool for performance enhancements, but also as a means to potentially rectify biases within these models.

Instruction tuning is critical to improve LLMs but usually suffers from low-quality and redundant data. Data filtering for instruction tuning has proved important in improving both the efficiency and performance of the tuning process. But it also leads to extra cost and computation due to the involvement of LLMs in this process. To reduce the filtering cost, we study Superfiltering: Can we use a smaller and weaker model to select data for finetuning a larger and stronger model? Despite the performance gap between weak and strong language models, we find their highly consistent capability to perceive instruction difficulty and data selection results. This enables us to use a much smaller and more efficient model to filter the instruction data used to train a larger language model. Not only does it largely speed up the data filtering, but the filtered-data-finetuned LLM achieves even better performance on standard benchmarks. Extensive experiments validate the efficacy and efficiency of our approach.

With the popularity of large language models (LLMs) and their ability to handle longer input documents, there is a growing need for high-quality long document summarization datasets. Although many models already support 16k input, current lengths of summarization datasets are inadequate, and salient information is not evenly distributed. To bridge these gaps, we collect a new summarization dataset called SumSurvey, consisting of more than 18k scientific survey papers. With an average document length exceeding 12k and a quarter exceeding 16k, as well as the uniformity metric outperforming current mainstream long document summarization datasets, SumSurvey brings new challenges and expectations to both fine-tuned models and LLMs. The informativeness of summaries and the models supporting the evaluation of long document summarization warrant further attention. Automatic and human evaluation results on this abstractive dataset confirm this view. Our dataset and code are available at https://github.com/Oswald1997/SumSurvey.

The widespread adoption of Large Language Models (LLMs) has led to an increase in AI-generated text on the Internet, presenting a crucial challenge to differentiate AI-created content from human-written text. This challenge is critical to prevent issues of authenticity, trust, and potential copyright violations. Current research focuses on watermarking LLM-generated text, but traditional techniques struggle to balance robustness with text quality. We introduce a novel watermarking approach, Robust and Imperceptible Watermarking (RIW) for LLMs, which leverages token prior probabilities to improve detectability and maintain watermark imperceptibility. RIW methodically embeds watermarks by partitioning selected tokens into two distinct groups based on their prior probabilities and employing tailored strategies for each group. In the detection stage, the RIW method employs the ‘voted z-test’ to provide a statistically robust framework to identify the presence of a watermark accurately. The effectiveness of RIW is evaluated across three key dimensions: success rate, text quality, and robustness against removal attacks. Our experimental results on various LLMs, including GPT2-XL, OPT-1.3B, and LLaMA2-7B, indicate that RIW surpasses existing models, and also exhibits increased robustness against various attacks and good imperceptibility, thus promoting the responsible use of LLMs.

Research on Large Language Models (LLMs) has often neglected subtle biases that, although less apparent, can significantly influence the models’ outputs toward particular social narratives. This study addresses two such biases within LLMs: representative bias, which denotes a tendency of LLMs to generate outputs that mirror the experiences of certain identity groups, and affinity bias, reflecting the models’ evaluative preferences for specific narratives or viewpoints. We introduce two novel metrics to measure these biases: the Representative Bias Score (RBS) and the Affinity Bias Score (ABS), and present the Creativity-Oriented Generation Suite (CoGS), a collection of open-ended tasks such as short story writing and poetry composition, designed with customized rubrics to detect these subtle biases. Our analysis uncovers marked representative biases in prominent LLMs, with a preference for identities associated with being white, straight, and men. Furthermore, our investigation of affinity bias reveals distinctive evaluative patterns within each model, akin to ‘bias fingerprints’. This trend is also seen in human evaluators, highlighting a complex interplay between human and machine bias perceptions.

Given a script, the challenge in Movie Dubbing (Visual Voice Cloning, V2C) is to generate speech that aligns well with the video in both time and emotion, based on the tone of a reference audio track. Existing state-of-the-art V2C models break the phonemes in the script according to the divisions between video frames, which solves the temporal alignment problem but leads to incomplete phoneme pronunciation and poor identity stability. To address this problem, we propose StyleDubber, which switches dubbing learning from the frame level to phoneme level. It contains three main components: (1) A multimodal style adaptor operating at the phoneme level to learn pronunciation style from the reference audio, and generate intermediate representations informed by the facial emotion presented in the video; (2) An utterance-level style learning module, which guides both the mel-spectrogram decoding and the refining processes from the intermediate embeddings to improve the overall style expression; And (3) a phoneme-guided lip aligner to maintain lip sync. Extensive experiments on two of the primary benchmarks, V2C and Grid, demonstrate the favorable performance of the proposed method as compared to the current state-of-the-art. The code will be made available at https://github.com/GalaxyCong/StyleDubber.

The widespread use of large language models (LLMs) is increasing the demand for methods that detect machine-generated text to prevent misuse. The goal of our study is to stress test the detectors’ robustness to malicious attacks under realistic scenarios. We comprehensively study the robustness of popular machine-generated text detectors under attacks from diverse categories: editing, paraphrasing, co-generating, and prompting. Our attacks assume limited access to the generator LLMs, and we compare the performance of detectors on different attacks under different budget levels. Our experiments reveal that almost none of the existing detectors remain robust under all the attacks, and all detectors exhibit different loopholes. Averaging all detectors, the performance drops by 35% across all attacks. Further, we investigate the reasons behind these defects and propose initial out-of-the-box patches.

Code LLMs have the potential to make it easier for non-experts to understand and write code. However, current CodeLLM benchmarks rely on a single expert-written prompt per problem, making it hard to generalize their success to non-expert users. In this paper, we present a new natural-language-to-code benchmark of prompts written by a key population of non-experts: beginning programmers. StudentEval contains 1,749 prompts written by 80 students who have only completed one introductory Python course. StudentEval contains numerous non-expert prompts describing the same problem, enabling exploration of key factors in prompt success. We use StudentEval to evaluate 12 Code LLMs and find that StudentEval is a better discriminator of model performance than existing benchmarks. Our analysis of student prompting strategies reveals that nondeterministic LLM sampling can mislead students about the quality of their descriptions, a finding with key implications for Code LLMs in education.

We present the structured average intersection-over-union ratio (STRUCT-IOU), an evaluation metric that compares a constituency parse tree over automatically recognized spoken word boundaries with the ground-truth parse tree over written words. To compute the metric, we (1) project the ground-truth parse tree to the speech domain by forced alignment, (2) align the projected ground-truth constituents with the predicted ones under certain structured constraints, and (3) calculate the average IOU score across all aligned constituent pairs. STRUCT-IOU takes word boundaries into account and overcomes the challenge that the predicted words and ground truth may not have perfect one-to-one correspondence. Extending to the evaluation of text constituency parsing, we demonstrate that STRUCT-IOU shows higher tolerance to syntactically plausible parses than PARSEVAL (Black et al., 1991).

Neural parameterization has significantly advanced unsupervised grammar induction. However, training these models with a traditional likelihood loss for all possible parses exacerbates two issues: 1) *structural optimization ambiguity* that arbitrarily selects one among structurally ambiguous optimal grammars despite the specific preference of gold parses, and 2) *structural simplicity bias* that leads a model to underutilize rules to compose parse trees. These challenges subject unsupervised neural grammar induction (UNGI) to inevitable prediction errors, high variance, and the necessity for extensive grammars to achieve accurate predictions. This paper tackles these issues, offering a comprehensive analysis of their origins. As a solution, we introduce *sentence-wise parse-focusing* to reduce the parse pool per sentence for loss evaluation, using the structural bias from pre-trained parsers on the same dataset.In unsupervised parsing benchmark tests, our method significantly improves performance while effectively reducing variance and bias toward overly simplistic parses. Our research promotes learning more compact, accurate, and consistent explicit grammars, facilitating better interpretability.

Evaluation is the baton for the development of large language models. Current evaluations typically employ a single-item assessment paradigm for each atomic test objective, which struggle to discern whether a model genuinely possesses the required capabilities or merely memorizes/guesses the answers to specific questions. To this end, this paper proposes a novel evaluation framework referred to as StructEval. Starting from an atomic test objective, StructEval deepens and broadens the evaluation by conducting a structured assessment across multiple cognitive levels and critical concepts, and therefore offers a comprehensive, robust and consistent evaluations for large language models. Experiments on three widely-used benchmarks demonstrate that StructEval serves as a reliable tool for resisting the risk of data contamination, and reducing the interference of potential biases, thereby providing a more reliable and consistent conclusion regarding model capabilities. Our framework also sheds light on the design of future principled and trustworthy LLM evaluation protocols.

Attribute value extraction involves identifying the value spans of predetermined attributes in product texts. This area of research has traditionally operated under a closed-world assumption, focusing on products from a static set of categories and their associated attributes. However, products in e-commerce stores are ever-increasing and evolving, calling for life-long learning. If continuously trained on the fast-increasing products and attributes, most existing solutions not only struggle for parameter efficiency but also endure foreseeable defects due to data contamination, catastrophic forgetting, etc. As a remedy, we propose and study a new task, which aims to effectively maintain a strong single model for many domains in a life-long learning fashion, without jeopardizing the model performance and parameter efficiency. We introduce factorization into the model and make it domain-aware by decoupling the modeling of product type and attribute, as a way to promote de-contamination and parameter efficiency while scaling up. Tuning the model with distillation prevents forgetting historical knowledge and enables continuous learning from emerging domains. Experiments on hundreds of domains showed that our model attains the near state-of-the-art performance with affordable parameter size, the least historical knowledge forgetting, and the greatest robustness against noises, whilst adding only a few parameters per domain when compared with competitive baselines.

Despite great performance on many tasks, language models (LMs) still struggle with reasoning, sometimes providing responses that cannot possibly be true because they stem from logical incoherence. We call such responses strong hallucinations and prove that they follow from an LM’s computation of its internal representations for logical operators and outputs from those representations. Focusing on negation, we provide a novel solution in which negation is treated not as another element of a latent representation, but as an operation over an LM’s latent representations that constrains how they may evolve. We show that our approach improves model performance in cloze prompting and natural language inference tasks with negation without requiring training on sparse negative data.

In response to the limitations of manual ad creation, significant research has been conducted in the field of automatic ad text generation (ATG). However, the lack of comprehensive benchmarks and well-defined problem sets has made comparing different methods challenging. To tackle these challenges, we standardize the task of ATG and propose a first benchmark dataset, CAMERA, carefully designed and enabling the utilization of multi-modal information and facilitating industry-wise evaluations. Our extensive experiments with a variety of nine baselines, from classical methods to state-of-the-art models including large language models (LLMs), show the current state and the remaining challenges. We also explore how existing metrics in ATG and an LLM-based evaluator align with human evaluations.

Multiple-Choice Questions (MCQs) constitute a critical area of research in the study of Large Language Models (LLMs). Previous works have investigated the selection bias problem in MCQs within few-shot scenarios, in which the LLM’s performance may be influenced by the presentation of answer choices, leaving the selection bias during Supervised Fine-Tuning (SFT) unexplored. In this paper, we reveal that selection bias persists in the SFT phase , primarily due to the LLM’s inadequate Multiple Choice Symbol Binding (MCSB) ability. This limitation implies that the model struggles to associate the answer options with their corresponding symbols (e.g., A/B/C/D) effectively. To enhance the model’s MCSB capability, we first incorporate option contents into the loss function and subsequently adjust the weights of the option symbols and contents, guiding the model to understand the option content of the current symbol. Based on this, we introduce an efficient SFT algorithm for MCQs, termed Point-wise Intelligent Feedback (PIF). PIF constructs negative instances by randomly combin- ing the incorrect option contents with all candidate symbols, and proposes a point-wise loss to provide feedback on these negative samples into LLMs. Our experimental results demonstrate that PIF significantly reduces the model’s selection bias by improving its MCSB capability. Remarkably, PIF exhibits a substantial enhancement in the accuracy for MCQs.