Computation and Language (cs.CL)

Abstract: Exposure bias poses a common challenge in numerous natural language processing tasks, particularly in the dialog generation. In response to this issue, researchers have devised various techniques, among which scheduled sampling has proven to be an effective method for mitigating exposure bias. However, the existing state-of-the-art scheduled sampling methods solely consider the current sampling words' quality for threshold truncation sampling, which overlooks the importance of sentence-level information and the method of threshold truncation warrants further discussion. In this paper, we propose a bilevel scheduled sampling model that takes the sentence-level information into account and incorporates it with word-level quality. To enhance sampling diversity and improve the model's adaptability, we propose a smooth function that maps the combined result of sentence-level and word-level information to an appropriate range, and employ probabilistic sampling based on the mapped values instead of threshold truncation. Experiments conducted on the DailyDialog and PersonaChat datasets demonstrate the effectiveness of our proposed methods, which significantly alleviate the exposure bias problem and outperform state-of-the-art scheduled sampling methods.

Abstract: Large Language Models (LLMs) have made significant strides in both scientific research and practical applications. Existing studies have demonstrated the state-of-the-art (SOTA) performance of LLMs in various natural language processing tasks. However, the question of how to further enhance LLMs' performance in specific task using prompting strategies remains a pivotal concern. This paper explores the enhancement of LLMs' performance in sentiment analysis through the application of prompting strategies. We formulate the process of prompting for sentiment analysis tasks and introduce two novel strategies tailored for sentiment analysis: RolePlaying (RP) prompting and Chain-of-thought (CoT) prompting. Specifically, we also propose the RP-CoT prompting strategy which is a combination of RP prompting and CoT prompting. We conduct comparative experiments on three distinct domain datasets to evaluate the effectiveness of the proposed sentiment analysis strategies. The results demonstrate that the adoption of the proposed prompting strategies leads to a increasing enhancement in sentiment analysis accuracy. Further, the CoT prompting strategy exhibits a notable impact on implicit sentiment analysis, with the RP-CoT prompting strategy delivering the most superior performance among all strategies.

Abstract: Large language models (LLMs) have achieved significant success in interacting with human. However, recent studies have revealed that these models often suffer from hallucinations, leading to overly confident but incorrect judgments. This limits their application in the medical domain, where tasks require the utmost accuracy. This paper introduces an automated evaluation framework that assesses the practical capabilities of LLMs as virtual doctors during multi-turn consultations. Consultation tasks are designed to require LLMs to be aware of what they do not know, to inquire about missing medical information from patients, and to ultimately make diagnoses. To evaluate the performance of LLMs for these tasks, a benchmark is proposed by reformulating medical multiple-choice questions from the United States Medical Licensing Examinations (USMLE), and comprehensive evaluation metrics are developed and evaluated on three constructed test sets. A medical consultation training set is further constructed to improve the consultation ability of LLMs. The results of the experiments show that fine-tuning with the training set can alleviate hallucinations and improve LLMs' performance on the proposed benchmark. Extensive experiments and ablation studies are conducted to validate the effectiveness and robustness of the proposed framework.

Abstract: The term emotion analysis in text subsumes various natural language processing tasks which have in common the goal to enable computers to understand emotions. Most popular is emotion classification in which one or multiple emotions are assigned to a predefined textual unit. While such setting is appropriate to identify the reader's or author's emotion, emotion role labeling adds the perspective of mentioned entities and extracts text spans that correspond to the emotion cause. The underlying emotion theories agree on one important point; that an emotion is caused by some internal or external event and comprises several subcomponents, including the subjective feeling and a cognitive evaluation. We therefore argue that emotions and events are related in two ways. (1) Emotions are events; and this perspective is the fundament in NLP for emotion role labeling. (2) Emotions are caused by events; a perspective that is made explicit with research how to incorporate psychological appraisal theories in NLP models to interpret events. These two research directions, role labeling and (event-focused) emotion classification, have by and large been tackled separately. We contributed to both directions with the projects SEAT (Structured Multi-Domain Emotion Analysis from Text) and CEAT (Computational Event Evaluation based on Appraisal Theories for Emotion Analysis), both funded by the German Research Foundation. In this paper, we consolidate the findings and point out open research questions.

Abstract: Query-Focused Meeting Summarization (QFMS) aims to generate a summary of a given meeting transcript conditioned upon a query. The main challenges for QFMS are the long input text length and sparse query-relevant information in the meeting transcript. In this paper, we propose a knowledge-enhanced two-stage framework called Knowledge-Aware Summarizer (KAS) to tackle the challenges. In the first stage, we introduce knowledge-aware scores to improve the query-relevant segment extraction. In the second stage, we incorporate query-relevant knowledge in the summary generation. Experimental results on the QMSum dataset show that our approach achieves state-of-the-art performance. Further analysis proves the competency of our methods in generating relevant and faithful summaries.

Abstract: Multimodal emotion recognition (MER) aims to detect the emotional status of a given expression by combining the speech and text information. Intuitively, label information should be capable of helping the model locate the salient tokens/frames relevant to the specific emotion, which finally facilitates the MER task. Inspired by this, we propose a novel approach for MER by leveraging label information. Specifically, we first obtain the representative label embeddings for both text and speech modalities, then learn the label-enhanced text/speech representations for each utterance via label-token and label-frame interactions. Finally, we devise a novel label-guided attentive fusion module to fuse the label-aware text and speech representations for emotion classification. Extensive experiments were conducted on the public IEMOCAP dataset, and experimental results demonstrate that our proposed approach outperforms existing baselines and achieves new state-of-the-art performance.

Abstract: Reasoning is a cognitive process of using evidence to reach a sound conclusion. The reasoning capability is essential for large language models (LLMs) to serve as the brain of the artificial general intelligence agent. Recent studies reveal that fine-tuning LLMs on data with the chain of thought (COT) reasoning process can significantly enhance their reasoning capabilities. However, we find that the fine-tuned LLMs suffer from an \textit{Assessment Misalignment} problem, i.e., they frequently assign higher scores to subpar COTs, leading to potential limitations in their reasoning abilities. To address this problem, we introduce an \textit{Alignment Fine-Tuning (AFT)} paradigm, which involves three steps: 1) fine-tuning LLMs with COT training data; 2) generating multiple COT responses for each question, and categorizing them into positive and negative ones based on whether they achieve the correct answer; 3) calibrating the scores of positive and negative responses given by LLMs with a novel constraint alignment loss. Specifically, the constraint alignment loss has two objectives: a) Alignment, which guarantees that positive scores surpass negative scores to encourage answers with high-quality COTs; b) Constraint, which keeps the negative scores confined to a reasonable range to prevent the model degradation. Beyond just the binary positive and negative feedback, the constraint alignment loss can be seamlessly adapted to the ranking situations when ranking feedback is accessible. Furthermore, we also delve deeply into recent ranking-based alignment methods, such as DPO, RRHF, and PRO, and discover that the constraint, which has been overlooked by these approaches, is also crucial for their performance. Extensive experiments on four reasoning benchmarks with both binary and ranking feedback demonstrate the effectiveness of AFT.

Abstract: In recent research, in the domain of speech processing, large End-to-End (E2E) systems for Automatic Speech Recognition (ASR) have reported state-of-the-art performance on various benchmarks. These systems intrinsically learn how to handle and remove noise conditions from speech. Previous research has shown, that it is possible to extract the denoising capabilities of these models into a preprocessor network, which can be used as a frontend for downstream ASR models. However, the proposed methods were limited to specific fully convolutional architectures. In this work, we propose a novel method to extract the denoising capabilities, that can be applied to any encoder-decoder architecture. We propose the Cleancoder preprocessor architecture that extracts hidden activations from the Conformer ASR model and feeds them to a decoder to predict denoised spectrograms. We train our pre-processor on the Noisy Speech Database (NSD) to reconstruct denoised spectrograms from noisy inputs. Then, we evaluate our model as a frontend to a pretrained Conformer ASR model as well as a frontend to train smaller Conformer ASR models from scratch. We show that the Cleancoder is able to filter noise from speech and that it improves the total Word Error Rate (WER) of the downstream model in noisy conditions for both applications.

Abstract: In this paper, we investigate the use of transformers for Neural Machine Translation of text-to-GLOSS for Deaf and Hard-of-Hearing communication. Due to the scarcity of available data and limited resources for text-to-GLOSS translation, we treat the problem as a low-resource language task. We use our novel hyper-parameter exploration technique to explore a variety of architectural parameters and build an optimal transformer-based architecture specifically tailored for text-to-GLOSS translation. The study aims to improve the accuracy and fluency of Neural Machine Translation generated GLOSS. This is achieved by examining various architectural parameters including layer count, attention heads, embedding dimension, dropout, and label smoothing to identify the optimal architecture for improving text-to-GLOSS translation performance. The experiments conducted on the PHOENIX14T dataset reveal that the optimal transformer architecture outperforms previous work on the same dataset. The best model reaches a ROUGE (Recall-Oriented Understudy for Gisting Evaluation) score of 55.18% and a BLEU-1 (BiLingual Evaluation Understudy 1) score of 63.6%, outperforming state-of-the-art results on the BLEU1 and ROUGE score by 8.42 and 0.63 respectively.

Abstract: We investigate the potential benefit of incorporating dictionary information into a neural network architecture for natural language processing. In particular, we make use of this architecture to extract several concepts related to COVID-19 from an on-line medical forum. We use a sample from the forum to manually curate one dictionary for each concept. In addition, we use MetaMap, which is a tool for extracting biomedical concepts, to identify a small number of semantic concepts. For a supervised concept extraction task on the forum data, our best model achieved a macro $F_1$ score of 90\%. A major difficulty in medical concept extraction is obtaining labelled data from which to build supervised models. We investigate the utility of our models to transfer to data derived from a different source in two ways. First for producing labels via weak learning and second to perform concept extraction. The dataset we use in this case comprises COVID-19 related tweets and we achieve an $F_1$ score 81\% for symptom concept extraction trained on weakly labelled data. The utility of our dictionaries is compared with a COVID-19 symptom dictionary that was constructed directly from Twitter. Further experiments that incorporate BERT and a COVID-19 version of BERTweet demonstrate that the dictionaries provide a commensurate result. Our results show that incorporating small domain dictionaries to deep learning models can improve concept extraction tasks. Moreover, models built using dictionaries generalize well and are transferable to different datasets on a similar task.

Abstract: Environmental, Social, and Governance (ESG) has been used as a metric to measure the negative impacts and enhance positive outcomes of companies in areas such as the environment, society, and governance. Recently, investors have increasingly recognized the significance of ESG criteria in their investment choices, leading businesses to integrate ESG principles into their operations and strategies. The Multi-Lingual ESG Issue Identification (ML-ESG) shared task encompasses the classification of news documents into 35 distinct ESG issue labels. In this study, we explored multiple strategies harnessing BERT language models to achieve accurate classification of news documents across these labels. Our analysis revealed that the RoBERTa classifier emerged as one of the most successful approaches, securing the second-place position for the English test dataset, and sharing the fifth-place position for the French test dataset. Furthermore, our SVM-based binary model tailored for the Chinese language exhibited exceptional performance, earning the second-place rank on the test dataset.

Abstract: Large-scale language models (LLMs), such as ChatGPT, are capable of generating human-like responses for various downstream tasks, such as task-oriented dialogues and question answering. However, applying LLMs to medical domains remains challenging due to their inability to leverage domain-specific knowledge. In this study, we present the Large-scale Language Models Augmented with Medical Textbooks (LLM-AMT), which integrates authoritative medical textbooks as the cornerstone of its design, enhancing its proficiency in the specialized domain through plug-and-play modules, comprised of a Hybrid Textbook Retriever, supplemented by the Query Augmenter and the LLM Reader. Experimental evaluation on three open-domain medical question-answering tasks reveals a substantial enhancement in both the professionalism and accuracy of the LLM responses when utilizing LLM-AMT, exhibiting an improvement ranging from 11.4% to 13.2%. Despite being 100 times smaller, we found that medical textbooks as the retrieval corpus serves as a more valuable external knowledge source than Wikipedia in the medical domain. Our experiments show that textbook augmentation results in a performance improvement ranging from 9.7% to 12.2% over Wikipedia augmentation.

Abstract: Recent works usually address Dialog policy learning DPL by training a reinforcement learning (RL) agent to determine the best dialog action. However, existing works on deep RL require a large volume of agent-user interactions to achieve acceptable performance. In this paper, we propose to make full use of the plain text knowledge from the pre-trained language model to accelerate the RL agent's learning speed. Specifically, we design a dialog action-aware transformer encoder (DaTrans), which integrates a new fine-tuning procedure named masked last action task to encourage DaTrans to be dialog-aware and distils action-specific features. Then, DaTrans is further optimized in an RL setting with ongoing interactions and evolves through exploration in the dialog action space toward maximizing long-term accumulated rewards. The effectiveness and efficiency of the proposed model are demonstrated with both simulator evaluation and human evaluation.

Abstract: Recent computational approaches for combating online hate speech involve the automatic generation of counter narratives by adapting Pretrained Transformer-based Language Models (PLMs) with human-curated data. This process, however, can produce in-domain overfitting, resulting in models generating acceptable narratives only for hatred similar to training data, with little portability to other targets or to real-world toxic language. This paper introduces novel attention regularization methodologies to improve the generalization capabilities of PLMs for counter narratives generation. Overfitting to training-specific terms is then discouraged, resulting in more diverse and richer narratives. We experiment with two attention-based regularization techniques on a benchmark English dataset. Regularized models produce better counter narratives than state-of-the-art approaches in most cases, both in terms of automatic metrics and human evaluation, especially when hateful targets are not present in the training data. This work paves the way for better and more flexible counter-speech generation models, a task for which datasets are highly challenging to produce.

Abstract: State-of-the-art language models like T5 have revolutionized the NLP landscape, but their computational demands hinder a large portion of the research community. To address this challenge, we present nanoT5, a specially-optimized PyTorch framework for efficient pre-training and fine-tuning of T5 models. Drawing on insights from optimizer differences and prioritizing efficiency, nanoT5 allows a T5-Base model to be pre-trained on a single GPU in just 16 hours, without any loss in performance. With the introduction of this open-source framework, we hope to widen the accessibility to language modelling research and cater to the community's demand for more user-friendly T5 (Encoder-Decoder) implementations. Our contributions, including configurations, codebase, software/hardware insights, and pre-trained models, are available to the public, aiming to strike a balance between research accessibility and resource constraints in NLP.

Abstract: Measuring semantic change has thus far remained a task where methods using contextual embeddings have struggled to improve upon simpler techniques relying only on static word vectors. Moreover, many of the previously proposed approaches suffer from downsides related to scalability and ease of interpretation. We present a simplified approach to measuring semantic change using contextual embeddings, relying only on the most probable substitutes for masked terms. Not only is this approach directly interpretable, it is also far more efficient in terms of storage, achieves superior average performance across the most frequently cited datasets for this task, and allows for more nuanced investigation of change than is possible with static word vectors.