COLING 2024 论文列表

In recent years, Large Language Models (LLMs) have demonstrated exceptional performance in code-generation tasks. However, under enterprise scenarios where private APIs are pre-built, general LLMs often fail to meet expectations. Existing approaches are confronted with drawbacks of high resource consumption and inadequate handling of multi-API tasks. To address these challenges, we propose EpiGEN, an Efficient multi-Api code GENeration framework under enterprise scenario. It consists of three core modules: Task Decomposition Module (TDM), API Retrieval Module (ARM), and Code Generation Module (CGM), in which Langchain played an important role. Through a series of experiments, EpiGEN shows good acceptability and readability, compared to fully fine-tuned LLM with a larger number of parameters. Particularly, in medium and hard level tasks, the performance of EpiGEN on a single-GPU machine even surpasses that of a fully fine-tuned LLM that requires multi-GPU configuration. Generally, EpiGEN is model-size agnostic, facilitating a balance between the performance of code generation and computational requirements.

Existing models for diverse generative reasoning still struggle to generate multiple unique and plausible results. Through an in-depth examination, we argue that it is critical to leverage a mixture of experts as prefixes to enhance the diversity of generated results and make task-oriented adaptation in the latent space of the generation models to improve the quality of the responses. At this point, we propose EpLSA, an innovative model based on the synergy of expert-prefix mixtures and task-oriented latent space adaptation for diverse generative reasoning. Specifically, we use expert-prefixes mixtures to encourage the model to create multiple responses with different semantics and design a loss function to address the problem that the semantics is interfered by the expert-prefixes. Meanwhile, we design a task-oriented adaptation block to make the pre-trained encoder within the generation model more effectively adapted to the pre-trained decoder in the latent space, thus further improving the quality of the generated text. Extensive experiments on three different types of generative reasoning tasks demonstrate that EpLSA outperforms existing baseline models in terms of both the quality and diversity of the generated outputs. Our code is publicly available at https://github.com/IMU-MachineLearningSXD/EpLSA.

We propose a method to combine the hybrid Kaldi-based Automatic Speech Recognition (ASR) system with the end-to-end wav2vec 2.0 XLS-R ASR using confidence measures. Our research is focused on the low-resource Irish language. Given the limited available open-source resources, neither the standalone hybrid ASR nor the end-to-end ASR system can achieve optimal performance. By applying the Recognizer Output Voting Error Reduction (ROVER) technique, we illustrate how ensemble learning could facilitate mutual error correction between both ASR systems. This paper outlines the strategies for merging the hybrid Kaldi ASR model and the end-to-end XLS-R model with the help of confidence scores. Although contemporary state-of-the-art end-to-end ASR models face challenges related to prediction overconfidence, we utilize Renyi’s entropy-based confidence approach, tuned with temperature scaling, to align it with the Kaldi ASR confidence. Although there was no significant difference in the Word Error Rate (WER) between the hybrid and end-to-end ASR, we could achieve a notable reduction in WER after ensembling through ROVER. This resulted in an almost 14% Word Error Rate Reduction (WERR) on our primary test set and an approximately 20% WERR on other noisy and imbalanced test data.

Interest in Astrophysical Natural Language Processing (NLP) has increased recently, fueled by the development of specialized language models for information extraction. However, the scarcity of annotated resources for this domain is still a significant challenge. Most existing corpora are limited to Named Entity Recognition (NER) tasks, leaving a gap in resource diversity. To address this gap and facilitate a broader spectrum of NLP research in astrophysics, we introduce astroECR, an extension of our previously built Time-Domain Astrophysics Corpus (TDAC). Our contributions involve expanding it to cover named entities, coreferences, annotations related to astrophysical relationships, and normalizing celestial object names. We showcase practical utility through baseline models for four NLP tasks and provide the research community access to our corpus, code, and models.

We present a dataset of word usage graphs (WUGs), where the existing WUGs for multiple languages are enriched with cluster labels functioning as sense definitions. They are generated from scratch by fine-tuned encoder-decoder language models. The conducted human evaluation has shown that these definitions match the existing clusters in WUGs better than the definitions chosen from WordNet by two baseline systems. At the same time, the method is straightforward to use and easy to extend to new languages. The resulting enriched datasets can be extremely helpful for moving on to explainable semantic change modeling.

When creating a new dataset for evaluation, one of the first considerations is the size of the dataset. If our evaluation data is too small, we risk making unsupported claims based on the results on such data. If, on the other hand, the data is too large, we waste valuable annotation time and costs that could have been used to widen the scope of our evaluation (i.e. annotate for more domains/languages). Hence, we investigate the effect of the size and a variety of sampling strategies of evaluation data to optimize annotation efforts, using dependency parsing as a test case. We show that for in-language in-domain datasets, 5,000 tokens is enough to obtain a reliable ranking of different parsers; especially if the data is distant enough from the training split (otherwise, we recommend 10,000). In cross-domain setups, the same amounts are required, but in cross-lingual setups much less (2,000 tokens) is enough.

Recent advancements in large language models have showcased their remarkable generalizability across various domains. However, their reasoning abilities still have significant room for improvement, especially when confronted with scenarios requiring multi-step reasoning. Although large language models possess extensive knowledge, their reasoning often fails to effectively utilize this knowledge to establish a coherent thinking paradigm. These models sometimes show hallucinations as their reasoning procedures are unconstrained by logical principles. Aiming at improving the zero-shot chain-of-thought reasoning ability of large language models, we propose LoT (Logical Thoughts), a self-improvement prompting framework that leverages principles rooted in symbolic logic, particularly Reductio ad Absurdum, to systematically verify and rectify the reasoning processes step by step. Experimental evaluations conducted on language tasks in diverse domains, including arithmetic, commonsense, symbolic, causal inference, and social problems, demonstrate the efficacy of enhanced reasoning by logic. The implementation code for LoT can be accessed at: https://github.com/xf-zhao/LoT.

Developmental Education (DevEd) courses align students’ college-readiness skills with higher education literacy demands. These courses often use automated assessment tools like Accuplacer for student placement. Existing literature raises concerns about these exams’ accuracy and placement precision due to their narrow representation of the writing process. These concerns warrant further attention within the domain of automatic placement systems, particularly in the establishment of a reference corpus of annotated essays for these systems’ machine/deep learning. This study aims at an enhanced annotation procedure to assess college students’ writing patterns more accurately. It examines the efficacy of machine-learning-based DevEd placement, contrasting Accuplacer’s classification of 100 college-intending students’ essays into two levels (Level 1 and 2) against that of 6 human raters. The classification task encompassed the assessment of the 6 textual criteria currently used by Accuplacer: mechanical conventions, sentence variety & style, idea development & support, organization & structure, purpose & focus, and critical thinking. Results revealed low inter-rater agreement, both on the individual criteria and the overall classification, suggesting human assessment of writing proficiency can be inconsistent in this context. To achieve a more accurate determination of writing proficiency and improve DevEd placement, more robust classification methods are thus required.

Event Coreference Resolution remains a challenging discourse-oriented task within the domain of Natural Language Processing. In this paper we propose a methodology where we combine traditional mention-pair coreference models with a lightweight and modular graph reconstruction algorithm. We show that building graph models on top of existing mention-pair models leads to improved performance for both a wide range of baseline mention-pair algorithms as well as a recently developed state-of-the-art model and this at virtually no added computational cost. Moreover, additional experiments seem to indicate that our method is highly robust in low-data settings and that its performance scales with increases in performance for the underlying mention-pair models.

Fine-tuning Large Language Models (LLMs) for machine translation is effective but costly. It also increases the risk of overfitting and catastrophic forgetting, especially when training data is limited. To tackle these challenges, we propose a novel method that involves adjusting task-related layers in large models to better harness their machine translation capabilities. This method aims to retain the model’s knowledge on other tasks while optimizing performance on translation tasks. By revealing the structure and characteristics of attention weights through singular value decomposition (SVD), we can make fine adjustments to specific layers, leveraging the model’s potential for more accurate and efficient translations. Our method not only addresses computational resource consumption and catastrophic forgetting but also offers a new perspective on utilizing the capabilities of large models effectively. Experimental validation shows that adjusting task-related layers significantly improves performance on translation tasks while maintaining stability and accuracy on other tasks. This finding provides valuable insights for fine-tuning and applying large models, advancing the field of machine translation.

Large language models (LLMs) with prompting have achieved encouraging results on many natural language processing (NLP) tasks based on task-tailored promptings. Text-to-SQL is a critical task that generates SQL queries from natural language questions. However, prompting on LLMs haven’t show superior performance on Text-to-SQL task due to the absence of tailored promptings. In this work, we propose three promptings specifically designed for Text-to-SQL: SL-prompt, CC-prompt, and SL+CC prompt. SL-prompt is designed to guide LLMs to identify relevant tables; CC-prompt directs LLMs to generate SQL clause by clause; and SL+CC prompt is proposed to combine the strengths of these above promptings. The three prompting strategies makes three solutions for Text-to-SQL. Then, another prompting strategy, the RS-prompt is proposed to direct LLMs to select the best answer from the results of the solutions. We conducted extensive experiments, and experimental results show that our method achieved an execution accuracy of 86.2% and a test-suite accuracy of 76.9%, which is 1.1%, and 2.7% higher than the current state-of-the-art Text-to-SQL methods, respectively. The results confirmed that the proposed promptings enhanced the capabilities of LLMs on Text-to-SQL. Experimental results also show that the granularity of schema linking and the order of clause generation have great impact on the performance, which are considered little in previous research.

Machine translation focuses mainly on high-resource languages (HRLs), while low-resource languages (LRLs) like Taiwanese Hokkien are relatively under-explored. The study aims to address this gap by developing a dual translation model between Taiwanese Hokkien and both Traditional Mandarin Chinese and English. We employ a pre-trained LLaMA 2-7B model specialized in Traditional Mandarin Chinese to leverage the orthographic similarities between Taiwanese Hokkien Han and Traditional Mandarin Chinese. Our comprehensive experiments involve translation tasks across various writing systems of Taiwanese Hokkien as well as between Taiwanese Hokkien and other HRLs. We find that the use of a limited monolingual corpus still further improves the model’s Taiwanese Hokkien capabilities. We then utilize our translation model to standardize all Taiwanese Hokkien writing systems into Hokkien Han, resulting in further performance improvements. Additionally, we introduce an evaluation method incorporating back-translation and GPT-4 to ensure reliable translation quality assessment even for LRLs. The study contributes to narrowing the resource gap for Taiwanese Hokkien and empirically investigates the advantages and limitations of pre-training and fine-tuning based on LLaMA 2.

Chain of thought (CoT) has proven useful for problems requiring complex reasoning. Many of these problems are both textual and multimodal. Given the inputs in different modalities, a model generates a rationale and then uses it to answer a question. Because of the hallucination issue, the generated soft negative rationales with high textual quality but illogical semantics do not always help improve answer accuracy. This study proposes a rationale generation method using soft negative sampling (SNSE-CoT) to mitigate hallucinations in multimodal CoT. Five methods were applied to generate soft negative samples that shared highly similar text but had different semantics from the original. Bidirectional margin loss (BML) was applied to introduce them into the traditional contrastive learning framework that involves only positive and negative samples. Extensive experiments on the ScienceQA dataset demonstrated the effectiveness of the proposed method. Code and data are released at https://github.com/zgMin/SNSE-CoT.

Scientific paper summarization has been the focus of much recent research. Unlike previous research which summarizes only the paper in question, or which summarizes the paper and the papers that it references, or which summarizes the paper and the citing sentences from the papers that cite it, this work puts all three of these summarization techniques together. To accomplish this, we have, by utilizing the citation network, introduced a corpus for scientific document summarization that provides information about the document being summarized, the papers referenced by it, as well as the papers that have cited it. The proposed summarizer model utilizes the referenced articles as background information and citing articles to capture the impact of the scientific document on the research community. Another aspect of the proposed model is its ability to generate both the extractive and abstractive summaries in parallel. The parallel training helps the counterparts to improve their individual performance. Results have shown that the summaries are of high quality when considering the standard metrics.

Keyphrase extraction (KPE) is an important task in Natural Language Processing for many scenarios, which aims to extract keyphrases that are present in a given document. Many existing supervised methods treat KPE as sequential labeling, span-level classification, or generative tasks. However, these methods lack the ability to utilize keyphrase information, which may result in biased results. In this study, we propose Diff-KPE, which leverages the supervised Variational Information Bottleneck (VIB) to guide the text diffusion process for generating enhanced keyphrase representations. Diff-KPE first generates the desired keyphrase embeddings conditioned on the entire document and then injects the generated keyphrase embeddings into each phrase representation. A ranking network and VIB are then optimized together with rank loss and classification loss, respectively. This design of Diff-KPE allows us to rank each candidate phrase by utilizing both the information of keyphrases and the document. Experiments show that Diff-KPE outperforms existing KPE methods on a large open domain keyphrase extraction benchmark, OpenKP, and a scientific domain dataset, KP20K.

Lightweight fine-tuning is widely used as an important technique for efficiently adapting pre-trained language models (PLM) to downstream tasks. Despite the reduction in trainable parameters, existing lightweight fine-tuning methods are found to be effective in low-resource settings but often fail in high-resource settings, leading to unreliable outcomes. This limitation can be attributed to inflexible strategies: they identify the parameters of the model to be trained before fine-tuning and remain unchanged without taking into account the inherent variance of generalization ability in model components (i.e., feed-forward, attention layers) and potential changes during the fine-tuning process. In this paper, we introduce a simple but effective calibration for lightweight fine-tuning PLMs based on the matrix’s stable rank according to both model components and the training process. We proposed both theoretical analyses and experimental verification for the proposed calibration strategy. Considering efficiency, we further propose time-aware and structure-aware strategies to determine the most crucial time to commence the fine-tuning procedure and selectively apply parameter matrices for lightweight fine-tuning, respectively. Extensive experiments demonstrate the superiority of our proposed fine-tuning approach (average improvement 3.1 for GLUE score compared to lightweight fine-tuning method).

Large Language Models (LLMs) operating in 0-shot or few-shot settings achieve competitive results in Text Classification tasks. In-Context Learning (ICL) typically achieves better accuracy than the 0-shot setting, but it pays in terms of efficiency, due to the longer input prompt. In this paper, we propose a strategy to make LLMs as efficient as 0-shot text classifiers, while getting comparable or better accuracy than ICL. Our solution targets the low resource setting, i.e., when only 4 examples per class are available. Using a single LLM and few-shot real data we perform a sequence of generation, filtering and Parameter-Efficient Fine-Tuning steps to create a robust and efficient classifier. Experimental results show that our approach leads to competitive results on multiple text classification datasets.

In this paper, we introduce a novel approach for enhancing the reasoning capabilities of large language models (LLMs) for constraint satisfaction problems (CSPs), by converting reasoning problems into classification tasks. Our method leverages the LLM’s ability to decide when to call a function from a set of logical-linguistic primitives, each of which can interact with a local “scratchpad” memory and logical inference engine. Invocation of these primitives in the correct order writes the constraints to the scratchpad memory and enables the logical engine to verifiably solve the problem. We additionally propose a formal framework for exploring the “linguistic” hardness of CSP reasoning-problems for LLMs. Our experimental results demonstrate that under our proposed method, tasks with significant computational hardness can be converted to a form that is easier for LLMs to solve and yields a 40% improvement over baselines. This opens up new avenues for future research into hybrid cognitive models that integrate symbolic and neural approaches.

Knowledge selection is a crucial sub-task of Document Grounded Dialogue System. Existing methods view knowledge selection as a sentence matching or classification. However, those methods can’t capture the semantic relationships within complex document. We propose a flexible method that can construct multi-level document semantic graph from the grounding document automatically and store semantic relationships within the documents effectively. Besides, we also devise an auxiliary task to leverage the graph more efficiently and can help the optimization of knowledge selection task. We conduct extensive experiments on public datasets: WoW(CITATION) and Holl-E(CITATION). And we achieves state-of-the-art result on WoW. Our code has been released at https://github.com/ddf62/multi-level-semantic-document-graph.