EMNLP 2024 论文列表

Modern natural language generation (NLG) systems have led to the development of synthetic human-like open-ended texts, posing concerns as to who the original author of a text is. To address such concerns, we introduce DeB-Ang: the utilisation of a custom DeBERTa model with angular loss and contrastive loss functions for effective class separation in neural text classification tasks. We expand the application of this model on binary machine-generated text detection and multi-class neural authorship attribution. We demonstrate improved performance on many benchmark datasets whereby the accuracy for machine-generated text detection was increased by as much as 38.04% across all datasets.

As Large Language Models (LLMs) continue to exhibit remarkable performance in natural language understanding tasks, there is a crucial need to measure their ability for human-like multi-step logical reasoning. Existing logical reasoning evaluation benchmarks often focus primarily on simplistic single-step or multi-step reasoning with a limited set of inference rules. Furthermore, the lack of datasets for evaluating non-monotonic reasoning represents a crucial gap since it aligns more closely with human-like reasoning. To address these limitations, we propose Multi-LogiEval, a comprehensive evaluation dataset encompassing multi-step logical reasoning with various inference rules and depths. Multi-LogiEval covers three logic types — propositional, first-order, and non-monotonic consisting of more than 30 inference rules and more than 60 of their combinations with various depths. Leveraging this dataset, we conduct evaluations on a range of LLMs such as GPT-4, ChatGPT, Gemini-Pro, Orca, and Mistral, employing a zero-shot chain-of-thought. Experimental results show that there is a significant drop in the performance of LLMs as the reasoning steps/depth increases (average accuracy of ~68% at depth-1 to ~43% at depth-5). We further conduct a thorough investigation of reasoning chains generated by LLMs which reveals several important findings. We believe that Multi-LogiEval facilitates future research for evaluating and enhancing the logical reasoning ability of LLMs.

The Knowledge-Aware Visual Question Answering about Entity task aims to disambiguate entities using textual and visual information, as well as knowledge. It usually relies on two independent steps, information retrieval then reading comprehension, that do not benefit each other. Retrieval Augmented Generation (RAG) offers a solution by using generated answers as feedback for retrieval training. RAG usually relies solely on pseudo-relevant passages retrieved from external knowledge bases which can lead to ineffective answer generation. In this work, we propose a multi-level information RAG approach that enhances answer generation through entity retrieval and query expansion. We formulate a joint-training RAG loss such that answer generation is conditioned on both entity and passage retrievals. We show through experiments new state-of-the-art performance on the VIQuAE KB-VQA benchmark and demonstrate that our approach can help retrieve more actual relevant knowledge to generate accurate answers.

Speech Relation Extraction (SpeechRE) aims to extract relation triplets from speech data. However, existing studies usually use synthetic speech to train and evaluate SpeechRE models, hindering the further development of SpeechRE due to the disparity between synthetic and real speech. Meanwhile, the modality gap issue, unexplored in SpeechRE, limits the performance of existing models. In this paper, we construct two real SpeechRE datasets to facilitate subsequent researches and propose a Multi-level Cross-modal Alignment Model (MCAM) for SpeechRE. Our model consists of three components: 1) a speech encoder, extracting speech features from the input speech; 2) an alignment adapter, mapping these speech features into a suitable semantic space for the text decoder; and 3) a text decoder, autoregressively generating relation triplets based on the speech features. During training, we first additionally introduce a text encoder to serve as a semantic bridge between the speech encoder and the text decoder, and then train the alignment adapter to align the output features of speech and text encoders at multiple levels. In this way, we can effectively train the alignment adapter to bridge the modality gap between the speech encoder and the text decoder. Experimental results and in-depth analysis on our datasets strongly demonstrate the efficacy of our method.

Vietnamese, a low-resource language, is typically categorized into three primary dialect groups that belong to Northern, Central, and Southern Vietnam. However, each province within these regions exhibits its own distinct pronunciation variations. Despite the existence of various speech recognition datasets, none of them has provided a fine-grained classification of the 63 dialects specific to individual provinces of Vietnam. To address this gap, we introduce Vietnamese Multi-Dialect (ViMD) dataset, a novel comprehensive dataset capturing the rich diversity of 63 provincial dialects spoken across Vietnam. Our dataset comprises 102.56 hours of audio, consisting of approximately 19,000 utterances, and the associated transcripts contain over 1.2 million words. To provide benchmarks and simultaneously demonstrate the challenges of our dataset, we fine-tune state-of-the-art pre-trained models for two downstream tasks: (1) Dialect identification and (2) Speech recognition. The empirical results suggest two implications including the influence of geographical factors on dialects, and the constraints of current approaches in speech recognition tasks involving multi-dialect speech data. Our dataset is available for research purposes.

The tool-use Large Language Models (LLMs) that integrate with external Python interpreters have significantly enhanced mathematical reasoning capabilities for open-source LLMs, while tool-free methods chose another track: augmenting math reasoning data. However, a great method to integrate the above two research paths and combine their advantages remains to be explored. In this work, we firstly include new math questions via **mu**lti-perspective data augmenting methods and then synthesize **code**-nested solutions to them. The open LLMs (e.g., Llama-2) are finetuned on the augmented dataset to get the resulting models, **MuMath-Code** (𝜇-Math-Code). During the inference phase, our MuMath-Code generates code and interacts with the external python interpreter to get the execution results. Therefore, MuMath-Code leverages the advantages of both the external tool and data augmentation. To fully leverage the advantages of our augmented data, we propose a two-stage training strategy: In Stage-1, we finetune Llama-2 on pure CoT data to get an intermediate model, which then is trained on the code-nested data in Stage-2 to get the resulting MuMath-Code.Our MuMath-Code-7B achieves 83.8% on GSM8K and 52.4% on MATH, while MuMath-Code-70B model achieves new state-of-the-art performance among open methods—achieving 90.7% on GSM8K and 55.1% on MATH. Extensive experiments validate the combination of tool use and data augmentation, as well as our two-stage training strategy.We release the proposed dataset along with the associated code for public use: https://github.com/youweihao-tal/MuMath-Code.

The performance on general tasks decreases after Large Language Models (LLMs) are fine-tuned on domain-specific tasks, the phenomenon is known as Catastrophic Forgetting (CF). However, this paper presents a further challenge for real application of domain-specific LLMs beyond CF, called General Capabilities Integration (GCI), which necessitates the integration of both the general capabilities and domain knowledge within a single instance. The objective of GCI is not merely to retain previously acquired general capabilities alongside new domain knowledge, but to harmonize and utilize both sets of skills in a cohesive manner to enhance performance on domain-specific tasks. Taking legal domain as an example, we carefully design three groups of training and testing tasks without lacking practicability, and construct the corresponding datasets. To better incorporate general capabilities across domain-specific scenarios, we introduce ALoRA, which utilizes a multi-head attention module upon LoRA, facilitating direct information transfer from preceding tokens to the current one. This enhancement permits the representation to dynamically switch between domain-specific knowledge and general competencies according to the attention. Extensive experiments are conducted on the proposed tasks. The results exhibit the significance of our setting, and the effectiveness of our method.

Incorrect student answers can become valuable learning opportunities, provided that the student understands where they went wrong and why. To this end, rather than being given the correct answer, students should receive elaborated feedback on how to correct a mistake on their own. Highlighting the complex demands that the generation of such feedback places on a model’s input utilization abilities, we propose two extensions to the training pipeline. Firstly, we employ a KL regularization term between a standard and enriched input format to achieve more targeted input representations. Secondly, we add a preference optimization step to encourage student answer-adaptive feedback generation. The effectiveness of those extensions is underlined by a significant increase in model performance of 3.3 METEOR points. We go beyond traditional surface form-based metrics to assess two important dimensions of feedback quality, i.e., faithfulness and informativeness. Hereby, we are the first to propose an automatic metric measuring the degree to which feedback divulges the correct answer, that we call Informativeness Index I2. We verify in how far each metric captures feedback quality.

Word Usage Graphs (WUGs) represent human semantic proximity judgments for pairs of word uses in a weighted graph, which can be clustered to infer word sense clusters from simple pairwise word use judgments, avoiding the need for word sense definitions. SemEval-2020 Task 1 provided the first and to date largest manually annotated, diachronic WUG dataset. In this paper, we check the robustness and correctness of the annotations by continuing the SemEval annotation algorithm for two more rounds and comparing against an established annotation paradigm. Further, we test the reproducibility by resampling a new, smaller set of word uses from the SemEval source corpora and annotating them. Our work contributes to a better understanding of the problems and opportunities of the WUG annotation paradigm and points to future improvements.

We deal with the problem of Question Answering (QA) over a long document, which poses a challenge for modern Large Language Models (LLMs). Although LLMs can handle increasingly longer context windows, they struggle to effectively utilize the long content. To address this issue, we introduce the concept of a virtual document (VDoc). A VDoc is created by selecting chunks from the original document that are most likely to contain the information needed to answer the user’s question, while ensuring they fit within the LLM’s context window. We hypothesize that providing a short and focused VDoc to the LLM is more effective than filling the entire context window with less relevant information. Our experiments confirm this hypothesis and demonstrate that using VDocs improves results on the QA task.

Moral foundations theory (MFT) is a psychological assessment tool that decomposes human moral reasoning into five factors, including care/harm, liberty/oppression, and sanctity/degradation (Graham et al., 2009). People vary in the weight they place on these dimensions when making moral decisions, in part due to their cultural upbringing and political ideology. As large language models (LLMs) are trained on datasets collected from the internet, they may reflect the biases that are present in such corpora. This paper uses MFT as a lens to analyze whether popular LLMs have acquired a bias towards a particular set of moral values. We analyze known LLMs and find they exhibit particular moral foundations, and show how these relate to human moral foundations and political affiliations. We also measure the consistency of these biases, or whether they vary strongly depending on the context of how the model is prompted. Finally, we show that we can adversarially select prompts that encourage the moral to exhibit a particular set of moral foundations, and that this can affect the model’s behavior on downstream tasks. These findings help illustrate the potential risks and unintended consequences of LLMs assuming a particular moral stance.

Performance of large language models (LLMs) may vary with different prompts or instructions of even the same task. One commonly recognized factor for this phenomenon is the model’s familiarity with the given prompt or instruction, which is typically estimated by its perplexity. However, finding the prompt with the lowest perplexity is challenging, given the enormous space of possible prompting phrases. In this paper, we propose monotonic paraphrasing (MonoPara), an end-to-end decoding strategy that paraphrases given prompts or instructions into their lower perplexity counterparts based on an ensemble of a paraphrase LM for prompt (or instruction) rewriting, and a target LM (i.e. the prompt or instruction executor) that constrains the generation for lower perplexity. The ensemble decoding process can efficiently paraphrase the original prompt without altering its semantic meaning, while monotonically decrease the perplexity of each generation as calculated by the target LM. We explore in detail both greedy and search-based decoding as two alternative decoding schemes of MonoPara. Notably, MonoPara does not require any training and can monotonically lower the perplexity of the paraphrased prompt or instruction, leading to improved performance of zero-shot LM prompting as evaluated on a wide selection of tasks. In addition, MonoPara is also shown to effectively improve LMs’ generalization on perturbed and unseen task instructions.

Online hate speech propagation is a complex issue, deeply influenced by both the perpetrator and the target’s cultural, historical, and societal contexts. Consequently, developing a universally robust hate speech classifier for diverse social media texts remains a challenging and unsolved task. The lack of mechanisms to track the spread and severity of hate speech further complicates the formulation of effective solutions. In response to this, to monitor hate speech in Indonesia during the recent 2024 presidential election, we have employed advanced Natural Language Processing (NLP) technologies to create an improved hate speech classifier tailored for a narrower subset of texts; specifically, texts that target vulnerable groups that have historically been the targets of hate speech in Indonesia. Our focus is on texts that mention these six vulnerable minority groups in Indonesia: Shia, Ahmadiyyah, Christians, LGBTQ+, Indonesian Chinese, and people with disabilities, as well as one additional group of interest: Jews. The insights gained from our dashboard have assisted stakeholders in devising more effective strategies to counteract hate speech. Notably, our dashboard has persuaded the General Election Supervisory Body in Indonesia (BAWASLU) to collaborate with our institution and the Alliance of Independent Journalists (AJI) to monitor social media hate speech in vulnerable areas in the country known for hate speech dissemination or hate-related violence in the upcoming Indonesian regional elections. This dashboard is available online at https://aji.or.id/hate-speech-monitoring.

Large language models are playing an increasingly significant role in molecular research, yet existing models often generate erroneous information. Traditional evaluations fail to assess a model’s factual correctness. To rectify this absence, we present MoleculeQA, a novel question answering (QA) dataset which possesses 62K QA pairs over 23K molecules. Each QA pair, composed of a manual question, a positive option and three negative options, has consistent semantics with a molecular description from authoritative corpus. MoleculeQA is not only the first benchmark to evaluate molecular factual correctness but also the largest molecular QA dataset. A comprehensive evaluation on MoleculeQA for existing molecular LLMs exposes their deficiencies in specific aspects and pinpoints crucial factors for molecular modeling. Furthermore, we employ MoleculeQA in reinforcement learning to mitigate model hallucinations, thereby enhancing the factual correctness of generated information.

Automatic factuality verification of large language model (LLM) generations is becoming more and more widely used to combat hallucinations. A major point of tension in the literature is the granularity of this fact-checking: larger chunks of text are hard to fact-check, but more atomic facts like propositions may lack context to interpret correctly. In this work, we assess the role of context in these atomic facts. We argue that fully atomic facts are not the right representation, and define two criteria for molecular facts: decontextuality, or how well they can stand alone, and minimality, or how little extra information is added to achieve decontexuality. We quantify the impact of decontextualization on minimality, then present a baseline methodology for generating molecular facts automatically, aiming to add the right amount of information. We compare against various methods of decontextualization and find that molecular facts balance minimality with fact verification accuracy in ambiguous settings.

Pre-trained chemical language models (CLMs) excel in the field of molecular property prediction, utilizing string-based molecular descriptors such as SMILES for learning universal representations. However, such string-based descriptors implicitly contain limited structural information, which is closely associated with molecular property prediction. In this work, we introduce Moleco, a novel contrastive learning framework to enhance the understanding of molecular structures within CLMs. Based on the similarity of fingerprint vectors among different molecules, we train CLMs to distinguish structurally similar and dissimilar molecules in a contrastive manner. Experimental results demonstrate that Moleco significantly improves the molecular property prediction performance of CLMs, outperforming state-of-the-art models. Moreover, our in-depth analysis with diverse Moleco variants verifies that fingerprint vectors are highly effective features in improving CLMs’ understanding of the structural information of molecules.

Chemical representation learning has gained increasing interest due to the limited availability of supervised data in fields such as drug and materials design. This interest particularly extends to chemical language representation learning, which involves pre-training Transformers on SMILES sequences - textual descriptors of molecules. Despite its success in molecular property prediction, current practices often lead to overfitting and limited scalability due to early convergence. In this paper, we introduce a novel chemical language representation learning framework, called MolTRES, to address these issues. MolTRES incorporates generator-discriminator training, allowing the model to learn from more challenging examples that require structural understanding. In addition, we enrich molecular representations by transferring knowledge from scientific literature by integrating external materials embedding. Experimental results show that our model outperforms existing state-of-the-art models on popular molecular property prediction tasks.

While existing alignment paradigms have been integral in developing large language models (LLMs), LLMs often learn an averaged human preference and struggle to model diverse preferences across cultures, demographics, and communities. We propose Modular Pluralism, a modular framework based on multi-LLM collaboration for pluralistic alignment: it “plugs into” a base LLM a pool of smaller but specialized community LMs, where models collaborate in distinct modes to flexibility support three modes of pluralism: Overton, steerable, and distributional. Modular Pluralism is uniquely compatible with black-box LLMs and offers the modular control of adding new community LMs for previously underrepresented communities. We evaluate Modular Pluralism with six tasks and four datasets featuring questions/instructions with value-laden and perspective-informed responses. Extensive experiments demonstrate that Modular Pluralism advances the three pluralism objectives across six black-box and open-source LLMs. Further analysis reveals that LLMs are generally faithful to the inputs from smaller community LLMs, allowing seamless patching by adding a new community LM to better cover previously underrepresented communities.

Alignment with human preferences is an important step in developing accurate and safe large language models. This is no exception in machine translation (MT), where better handling of language nuances and context-specific variations leads to improved quality. However, preference data based on human feedback can be very expensive to obtain and curate at a large scale. Automatic metrics, on the other hand, can induce preferences, but they might not match human expectations perfectly. In this paper, we propose an approach that leverages the best of both worlds. We first collect sentence-level quality assessments from professional linguists on translations generated by multiple high-quality MT systems and evaluate the ability of current automatic metrics to recover these preferences. We then use this analysis to curate a new dataset, MT-Pref (metric induced translation preference) dataset, which comprises 18k instances covering 18 language directions, using texts sourced from multiple domains post-2022. We show that aligning TOWER models on MT-Pref significantly improves translation quality on WMT23 and FLORES benchmarks.