ACL 2024 论文列表

Multiple-choice questions (MCQs) are important in enhancing concept learning and student engagement for educational purposes. Despite the multimodal nature of educational content, current methods focus mainly on text-based inputs and often neglect the integration of visual information. In this work, we study the problem of multimodal educational question generation, which aims at generating subject-specific educational questions with plausible yet incorrect distractors based on multimodal educational content. To tackle this problem, we introduce a novel framework, named Chain-of-Exemplar (CoE), which utilizes multimodal large language models (MLLMs) with Chain-of-Thought reasoning to improve the generation of challenging distractors. Furthermore, CoE leverages three-stage contextualized exemplar retrieval to retrieve exemplary questions as guides for generating more subject-specific educational questions. Experimental results on the ScienceQA benchmark demonstrate the superiority of CoE in both question generation and distractor generation over existing methods across various subjects and educational levels.

Rule-based reasoning, a fundamental type of legal reasoning, enables us to draw conclusions by accurately applying a rule to a set of facts. We explore causal language models as rule-based reasoners, specifically with respect to compositional rules - rules consisting of multiple elements which form a complex logical expression. Reasoning about compositional rules is challenging because it requires multiple reasoning steps, and attending to the logical relationships between elements. We introduce a new prompting method, Chain of Logic, which elicits rule-based reasoning through decomposition (solving elements as independent threads of logic), and recomposition (recombining these sub-answers to resolve the underlying logical expression). This method was inspired by the IRAC (Issue, Rule, Application, Conclusion) framework, a sequential reasoning approach used by lawyers. We evaluate chain of logic across eight rule-based reasoning tasks involving three distinct compositional rules from the LegalBench benchmark and demonstrate it consistently outperforms other prompting methods, including chain of thought and self-ask, using open-source and commercial language models.

Minimum Bayes risk (MBR) decoding achieved state-of-the-art translation performance by using COMET, a neural metric that has a high correlation with human evaluation.However, MBR decoding requires quadratic time since it computes the expected score between a translation hypothesis and all reference translations.We propose centroid-based MBR (CBMBR) decoding to improve the speed of MBR decoding.Our method clusters the reference translations in the feature space, and then calculates the score using the centroids of each cluster.The experimental results show that our CBMBR not only improved the decoding speed of the expected score calculation 5.7 times, but also outperformed vanilla MBR decoding in translation quality by up to 0.5 COMET in the WMT’22 En↔Ja, En↔De, En↔Zh, and WMT’23 En↔Ja translation tasks.

Large language models (LLMs) show remarkable human-like capability in various domains and languages. To bridge this quality gap, we introduce Cendol, a collection of Indonesian LLMs encompassing both decoder-only and encoder-decoder architectures across a range of model sizes. We highlight Cendol’s effectiveness across a diverse array of tasks, attaining ~20% improvement, and demonstrate its capability to generalize to unseen tasks and indigenous languages of Indonesia. Furthermore, Cendol models showcase improved human favorability despite their limitations in capturing indigenous knowledge and cultural values in Indonesia. In addition, we discuss the shortcomings of parameter-efficient tunings, such as LoRA, for language adaptation. Alternatively, we propose the usage of vocabulary adaptation to enhance efficiency. Lastly, we evaluate the safety of Cendol and showcase that safety in pre-training in one language such as English is transferable to low-resource languages, such as Indonesian, even without RLHF and safety fine-tuning.

Pre-trained Language Models (PLMs), like BERT, with self-supervision objectives exhibit remarkable performance and generalization across various tasks. However, they suffer in inference latency due to their large size. To address this issue, side branches are attached at intermediate layers, enabling early inference of samples without requiring them to pass through all layers. However, the challenge is to decide which layer to infer and exit each sample so that the accuracy and latency are balanced. Moreover, the distribution of the samples to be inferred may differ from that used for training necessitating cross-domain adaptation. We propose an online learning algorithm named Cross-Domain Inference in Early Exit BERT (CeeBERT) that dynamically determines early exits of samples based on the level of confidence at each exit point. CeeBERT learns optimal thresholds from domain-specific confidence observed at intermediate layers on the fly, eliminating the need for labeled data. Experimental results on five distinct datasets with BERT and ALBERT models demonstrate CeeBERT’s ability to improve latency by reducing unnecessary computations with minimal drop in performance. By adapting to the threshold values, CeeBERT can speed up the BERT/ALBERT models by 2× - 3.1× with minimal drop in accuracy. The anonymized source code is available at https://github.com/Div290/CeeBERT.

Language models (LMs) have proven to be powerful tools for psycholinguistic research, but most prior work has focused on purely behavioural measures (e.g., surprisal comparisons). At the same time, research in model interpretability has begun to illuminate the abstract causal mechanisms shaping LM behavior. To help bring these strands of research closer together, we introduce CausalGym. We adapt and expand the SyntaxGym suite of tasks to benchmark the ability of interpretability methods to causally affect model behaviour. To illustrate how CausalGym can be used, we study the pythia models (14M–6.9B) and assess the causal efficacy of a wide range of interpretability methods, including linear probing and distributed alignment search (DAS). We find that DAS outperforms the other methods, and so we use it to study the learning trajectory of two difficult linguistic phenomena in pythia-1b: negative polarity item licensing and filler–gap dependencies. Our analysis shows that the mechanism implementing both of these tasks is learned in discrete stages, not gradually.

Citation count of a paper is a commonly used proxy for evaluating the significance of a paper in the scientific community. Yet citation measures are widely criticized for failing to accurately reflect the true impact of a paper. Thus, we propose CausalCite, a new way to measure the significance of a paper by assessing the causal impact of the paper on its follow-up papers. CausalCite is based on a novel causal inference method, TextMatch, which adapts the traditional matching framework to high-dimensional text embeddings. TextMatch encodes each paper using text embeddings from large language models (LLMs), extracts similar samples by cosine similarity, and synthesizes a counterfactual sample as the weighted average of similar papers according to their similarity values. We demonstrate the effectiveness of CausalCite on various criteria, such as high correlation with paper impact as reported by scientific experts on a previous dataset of 1K papers, (test-of-time) awards for past papers, and its stability across various subfields of AI. We also provide a set of findings that can serve as suggested ways for future researchers to use our metric for a better understanding of the quality of a paper. Our code is available at https://github.com/causalNLP/causal-cite.

Although achieving promising performance, recent analyses show that current generative large language models (LLMs) may still capture dataset biases and utilize them for generation, leading to poor generalizability and harmfulness of LLMs. However, due to the diversity of dataset biases and the over-optimization problem, previous prior-knowledge-based debiasing methods and fine-tuning-based debiasing methods may not be suitable for current LLMs.To address this issue, we explore combining active learning with the causal mechanisms and propose a casual-guided active learning (CAL) framework, which utilizes LLMs itself to automatically and autonomously identify informative biased samples and induce the bias patterns. Then a cost-effective and efficient in-context learning based method is employed to prevent LLMs from utilizing dataset biases during generation.Experimental results show that CAL can effectively recognize typical biased instances and induce various bias patterns for debiasing LLMs.

Understanding memorisation in language models has practical and societal implications, e.g., studying models’ training dynamics or preventing copyright infringements.Prior work defines memorisation as the causal effect of training with an instance on the model’s ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance.Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model’s memorisation profile–its memorisation trends across training–by only observing its behaviour on a small set of instances throughout training.In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.

Conversational systems often rely on embedding models for intent classification and intent clustering tasks. The advent of Large Language Models (LLMs), which enable instructional embeddings allowing one to adjust semantics over the embedding space using prompts, are being viewed as a panacea for these downstream conversational tasks. However, traditional evaluation benchmarks rely solely on task metrics that don’t particularly measure gaps related to semantic understanding. Thus, we propose an intent semantic toolkit that gives a more holistic view of intent embedding models by considering three tasks– (1) intent classification, (2) intent clustering, and (3) a novel triplet task. The triplet task gauges the model’s understanding of two semantic concepts paramount in real-world conversational systems– negation and implicature. We observe that current embedding models fare poorly in semantic understanding of these concepts. To address this, we propose a pre-training approach to improve the embedding model by leveraging augmentation with data generated by an auto-regressive model and a contrastive loss term. Our approach improves the semantic understanding of the intent embedding model on the aforementioned linguistic dimensions while slightly effecting their performance on downstream task metrics.

Do LMs infer the semantics of text from co-occurrence patterns in their training data? Merrill et al. (2022) argue that, in theory, sentence co-occurrence probabilities predicted by an optimal LM should reflect the entailment relationship of the constituent sentences, but it is unclear whether probabilities predicted by neural LMs encode entailment in this way because of strong assumptions made by Merrill et al. (namely, that humans always avoid redundancy). In this work, we investigate whether their theory can be used to decode entailment relations from neural LMs. We find that a test similar to theirs can decode entailment relations between natural sentences, well above random chance, though not perfectly, across many datasets and LMs. This suggests LMs implicitly model aspects of semantics to predict semantic effects on sentence co-occurrence patterns. However, we find the test that predicts entailment in practice works in the opposite direction to the theoretical test. We thus revisit the assumptions underlying the original test, finding its derivation did not adequately account for redundancy in human-written text. We argue that better accounting for redundancy related to *explanations* might derive the observed flipped test and, more generally, improve computational models of speakers in linguistics.

Model editing has become a promising method for precisely and effectively updating knowledge in language models. In this paper, we investigate knowledge attenuation, in which the retention of updated knowledge within the language model decreases as the number of edits increases after sequential editing. Through empirical study, we discovered that existing editing methods generally suffer from knowledge attenuation. We attribute this phenomenon to two aspects: (1) redundant parameters interference and (2) update weight disentanglement. To this end, we propose the AdaPLE method. It not only mitigates the knowledge attenuation issue but also improves the performance on existing benchmarks. To the best of our knowledge, we are the first to investigate the cause and mitigation of knowledge attenuation in sequential LLM editing.

Recently proposed two-pass direct speech-to-speech translation (S2ST) models decompose the task into speech-to-text translation (S2TT) and text-to-speech (TTS) within an end-to-end model, yielding promising results. However, the training of these models still relies on parallel speech data, which is extremely challenging to collect. In contrast, S2TT and TTS have accumulated a large amount of data and pretrained models, which have not been fully utilized in the development of S2ST models. Inspired by this, in this paper, we first introduce a composite S2ST model named ComSpeech, which can seamlessly integrate any pretrained S2TT and TTS models into a direct S2ST model. Furthermore, to eliminate the reliance on parallel speech data, we propose a novel training method ComSpeech-ZS that solely utilizes S2TT and TTS data. It aligns representations in the latent space through contrastive learning, enabling the speech synthesis capability learned from the TTS data to generalize to S2ST in a zero-shot manner. Experimental results on the CVSS dataset show that when the parallel speech data is available, ComSpeech surpasses previous two-pass models like UnitY and Translatotron 2 in both translation quality and decoding speed. When there is no parallel speech data, ComSpeech-ZS lags behind by only 0.7 ASR-BLEU and outperforms the cascaded models.

Text watermarking technology aims to tag and identify content produced by large language models (LLMs) to prevent misuse. In this study, we introduce the concept of cross-lingual consistency in text watermarking, which assesses the ability of text watermarks to maintain their effectiveness after being translated into other languages. Preliminary empirical results from two LLMs and three watermarking methods reveal that current text watermarking technologies lack consistency when texts are translated into various languages. Based on this observation, we propose a Cross-lingual Watermark Removal Attack (CWRA) to bypass watermarking by first obtaining a response from an LLM in a pivot language, which is then translated into the target language. CWRA can effectively remove watermarks, decreasing the AUCs to a random-guessing level without performance loss. Furthermore, we analyze two key factors that contribute to the cross-lingual consistency in text watermarking and propose X-SIR as a defense method against CWRA.

Understanding the deep semantics of images is essential in the era dominated by social media. However, current research works primarily on the superficial description of images, revealing a notable deficiency in the systematic investigation of the inherent deep semantics. In this work, we introduce DEEPEVAL, a comprehensive benchmark to assess Large Multimodal Models’ (LMMs) capacities of visual deep semantics. DEEPEVAL includes human-annotated dataset and three progressive subtasks: fine-grained description selection, in-depth title matching, and deep semantics understanding. Utilizing DEEPEVAL, we evaluate 9 open-source LMMs and GPT-4V(ision). Our evaluation demonstrates a substantial gap between the deep semantic comprehension capabilities of existing LMMs and humans. For example, GPT-4V is 30% behind humans in understanding deep semantics, even though it achieves human-comparable performance in image description. Further analysis reveals that LMM performance on DEEPEVAL varies according to the specific facets of deep semantics explored, indicating the fundamental challenges remaining in developing LMMs.

Emotional Support Conversation (ESC) is a task aimed at alleviating individuals’ emotional distress through daily conversation. Given its inherent complexity and non-intuitive nature, ESConv dataset incorporates support strategies to facilitate the generation of appropriate responses. Recently, despite the remarkable conversational ability of large language models (LLMs), previous studies have suggested that they often struggle with providing useful emotional support. Hence, this work initially analyzes the results of LLMs on ESConv, revealing challenges in selecting the correct strategy and a notable preference for a specific strategy. Motivated by these, we explore the impact of the inherent preference in LLMs on providing emotional support, and consequently, we observe that exhibiting high preference for specific strategies hinders effective emotional support, aggravating its robustness in predicting the appropriate strategy. Moreover, we conduct a methodological study to offer insights into the necessary approaches for LLMs to serve as proficient emotional supporters. Our findings emphasize that (1) low preference for specific strategies hinders the progress of emotional support, (2) external assistance helps reduce preference bias, and (3) existing LLMs alone cannot become good emotional supporters. These insights suggest promising avenues for future research to enhance the emotional intelligence of LLMs.

Finding interpretable factors for stock returns is the most vital issue in the empirical asset pricing domain. As data-driven methods, existing factor mining models can be categorized into symbol-based and neural-based models. Symbol-based models are interpretable but inefficient, while neural-based approaches are efficient but lack interpretability. Hence, mining interpretable factors effectively presents a significant challenge. Inspired by the success of Large Language Models (LLMs) in various tasks, we propose a FActor Mining Agent (FAMA) model that enables LLMs to integrate the strengths of both neural and symbolic models for factor mining. In this paper, FAMA consists of two main components: Cross-Sample Selection (CSS) and Chain-of-Experience (CoE). CSS addresses the homogeneity challenges in LLMs during factor mining by assimilating diverse factors as in-context samples, whereas CoE enables LLMs to leverage past successful mining experiences, expediting the mining of effective factors. Experimental evaluations on real-world stock market data demonstrate the effectiveness of our approach by surpassing the SOTA RankIC by 0.006 and RankICIR by 0.105 in predicting S&P 500 returns. Furthermore, the investment simulation shows that our model can achieve superior performance with an annualized return of 38.4% and a Sharpe ratio of 667.2%.

Noun-noun compounds interpretation is the task where a model is given one of such constructions, and it is asked to provide a paraphrase, making the semantic relation between the nouns explicit, as in carrot cake is “a cake made of carrots.” Such a task requires the ability to understand the implicit structured representation of the compound meaning. In this paper, we test to what extent the recent Large Language Models can interpret the semantic relation between the constituents of lexicalized English compounds and whether they can abstract from such semantic knowledge to predict the semantic relation between the constituents of similar but novel compounds by relying on analogical comparisons (e.g., carrot dessert). We test both Surprisal metrics and prompt-based methods to see whether i.) they can correctly predict the relation between constituents, and ii.) the semantic representation of the relation is robust to paraphrasing. Using a dataset of lexicalized and annotated noun-noun compounds, we find that LLMs can infer some semantic relations better than others (with a preference for compounds involving concrete concepts). When challenged to perform abstractions and transfer their interpretations to semantically similar but novel compounds, LLMs show serious limitations.

Although large language models (LLMs) exhibit remarkable capacity to leverage in-context demonstrations, it is still unclear to what extent they can learn new facts or concept definitions via prompts. To address this question, we examine the capacity of instruction-tuned LLMs to follow in-context concept annotation guidelines for zero-shot sentence labeling tasks. We design guidelines that present different types of factual and counterfactual concept definitions, which are used as prompts for zero-shot sentence classification tasks. Our results show that although concept definitions consistently help in task performance, only the larger models (with 70B parameters or more) have limited ability to work under counterfactual contexts. Importantly, only proprietary models such as GPT-3.5 can recognize nonsensical guidelines, which we hypothesize is due to more sophisticated alignment methods. Finally, we find that Falcon-180B-chat is outperformed by Llama-2-70B-chat is most cases, which indicates that increasing model scale does not guarantee better adherence to guidelines. Altogether, our simple evaluation method reveals significant gaps in concept understanding between the most capable open-source language models and the leading proprietary APIs.