CASIA OpenIR  > 毕业生  > 博士学位论文
医疗领域任务型对话系统研究
胡泽发
2024-05-15
Pages152
Subtype博士
Abstract

       对话交流是信息传递和获取的重要方式。通过对话,大众能够以迅速且准确的方式传达信息,确保信息的完整性和准确性。对话系统,作为人工智能和自然语言处理领域的核心研究课题,其在实际应用中的价值不容忽视。无论是智能手机、智能家居,无人驾驶还是智慧医疗等领域,对话系统都展现出了广阔的应用前景。特别是在医疗领域,增强对话交流对于提升医疗服务质量、增强患者满意度以及提高医疗效率具有深远的影响。在这些实际应用场景中,对话系统不仅要能够与用户进行自然的交互,更需要能够准确识别、理解和处理用户的任务需求,通过精心设计的对话流程,高效地完成各项任务。

       近期,随着大模型技术的不断发展和进步,对话系统也迎来了前所未有的发展机遇。这些系统通过深入挖掘大量语料数据,积累了丰富的语言知识和语义信息,进而显著提升了对用户意图和需求的精准理解能力。然而,大语言模型生成的流畅文本中仍然可能出现一些并不真实存在的特征或关系,这种现象被为“幻觉”。在医疗领域,这种幻觉现象可能产生严重的后果,因为该领域对对话系统的可靠性和稳定性要求极高。在这一背景下,如何巧妙地利用大语言模型构建更加精准、稳定的医疗领域任务型对话系统,从而有效缓解幻觉现象,成为了一个极具前景的研究课题。

       本论文专注于医疗领域任务型对话系统。在医疗领域中,医疗术语是医疗对话交流中重要的任务信息载体,各种医疗任务的完成都离不开基于医疗术语相关信息的对话交流。而医疗领域任务型对话中出现幻觉现象的重要原因是医疗领域知识不足,缺乏对医疗任务的理解和推理能力。本文以医疗术语为核心,侧重于“基础对话模型” +“术语任务模型”的研究范式,增强系统的任务理解和推理能力,缓解医疗领域任务型对话系统中的幻觉现象。本文首先通过以对话模型结合术语任务模型的方式构建了整体的对话系统框架。然后,针对医疗对话中理解和推理中的难点,本文提出了相应的术语任务模型,对话模型通过调用这些术语任务模型,以缓解基础大模型因医疗对话理解和推理能力不足而产生的幻觉。

       本文的主要研究内容和创新点如下:

       1. 基于术语建模的医疗对话系统:当前的医疗对话系统主要依赖于在医疗对话语料库上直接微调基础大模型。然而,这种方法限制了模型在医疗领域复杂对话中的推理学习能力,导致生成的对话回复缺乏针对性,并可能出现幻觉现象。为了增强医疗对话系统在对话过程中对医疗任务的推理能力并缓解幻觉现象,本文提出了“对话模型” +“术语任务模型”的医疗对话系统框架,并开发两个术语任务模型,即术语抽取工具和术语预测模型。对话模型通过调用这些术语任务模型,能够精确地建模对话中的术语信息流动过程,从而有效执行任务推理。在两个中文多轮医疗对话数据集的实验上,本文提出的方法显著提升了系统性能,特别是在对话中术语信息的准确性方面。

        2. 基于匹配语义预训练的口语化术语抽取方法:在真实的医疗对话场景中,患者往往使用口语化的方式描述病情,其中涉及到的医疗术语并不规范。这种非标准的术语表达方式难以被简单的术语抽取工具准确捕捉,而且容易被基础大模型忽略。这种医疗对话理解的不准确性可能进一步加剧对话回复中的幻觉现象。为了应对这一问题,本工作提出了一种匹配术语语义预训练方法,打造口语化术语理解任务模型。该方法通过精心设计两个与术语抽取紧密相关的自监督任务,即对比术语判别和基于匹配的掩码术语建模,旨在利用大量无标注的医疗对话数据来增强模型的术语语义理解能力。实验结果显示,无论在全数据、低资源还是零资源条件下,本方法均能有效提升模型性能。

       3. 基于知识增强生成式框架的术语状态对抽取方法:孤立的术语词汇信息往往难以精确描述具体的医疗状况,因为术语通常还具有相应的状态信息,如症状类术语的阳性、阴性等。在对话中,状态的共指性以及随着对话进展状态发生变化的特点,都为术语状态的理解带来了挑战。为了解决这些问题,本工作提出了一种基于知识增强的两阶段生成框架,通过打造术语状态对抽取任务模型为对话系统提供更丰富的术语相关信息。该框架通过构建不同的任务提示,以生成的方式分别完成术语和状态的抽取。同时,它还利用术语类别和候选状态等背景知识来增强状态生成阶段的任务提示。这种两阶段生成的特点使得框架能够仅利用包含术语标注的额外数据,进一步提升在低资源环境下的性能。

       4. 基于状态增强掩码预测框架的医疗对话诊断方法:在医疗对话中,疾病诊断是一项至关重要的任务,但通用的对话模型在处理这种复杂的推理时更容易出现幻觉现象。为了缓解这一问题,本工作致力于打造一个专用于对话疾病诊断的术语推理任务模型,以提升医疗对话系统在进行疾病诊断推理时的准确性和可靠性。为此,本文提出了一个统一的掩码预测框架,将症状问询和疾病预测两类术语推理子任务整合为一个整体任务。此外,本文还设计了相应的辅助任务,以增强症状和疾病状态建模的能力。在多个数据集上的实验验证了所提出方法的有效性,表明它能够显著提高医疗对话系统在疾病诊断任务中的性能。

Other Abstract

        Dialogue communication is an essential means of transmitting and acquiring information. Through dialogue, we can convey information quickly and accurately, ensuring its integrity and accuracy. Dialogue systems, as a core research topic in artificial intelligence and natural language processing, have significant value in practical applications. Whether in areas such as smartphones, smart homes, autonomous driving, or smart healthcare, dialogue systems demonstrate broad prospects for application. Especially in the medical field, enhancing dialogue communication has profound effects on improving the quality of medical services, enhancing patient satisfaction, and increasing medical efficiency. In these practical application scenarios, dialogue systems not only need to interact naturally with users but also accurately identify, understand, and process users’ task requirements, efficiently completing various tasks through carefully designed dialogue flows.

        Recently, with the continuous development and advancement of large-scale model technology, dialogue systems have encountered unprecedented opportunities for development. These systems, by deeply exploring a large amount of corpus data, have accumulated rich linguistic knowledge and semantic information, significantly enhancing their ability to accurately understand user intentions and demands. However, fluent text generated by large language models may still contain some features or relationships that do not truly exist, a phenomenon known as "hallucination". In the medical field, this hallucination phenomenon could have serious consequences because this domain requires high reliability and stability from dialogue systems. In this context, how to cleverly utilize large language models to construct more accurate and stable task-oriented dialogue systems in the medical domain, effectively alleviating hallucination phenomena, becomes a highly promising research topic. This thesis focuses on task-oriented dialogue systems in the medical domain. In the medical field, medical terminology serves as an important task information carrier in medical dialogue communication, and the completion of various medical tasks relies on dialogue communication based on relevant medical terminology information. One significant reason for the occurrence of hallucination phenomena in task-oriented dialogues in the medical domain is insufficient medical domain knowledge and a lack of understanding and reasoning ability for medical tasks.

        This thesis, centered on medical terminology, focuses on the research paradigm of "basic dialogue model" + "terminology task model", aiming to alleviate hallucination phenomena in task-oriented dialogue systems in the medical domain from multiple perspectives. This thesis first constructs an overall dialogue system framework by combining dialogue models with terminology task models. Then, addressing the difficulties in understanding and reasoning in medical dialogue, this thesis proposes corresponding terminology task models. The dialogue model alleviates hallucinations generated by basic large models due to insufficient understanding and reasoning capabilities in medical dialogue by invoking these terminology task models.

        The main research contents and innovations of this thesis are as follows:

        1. Medical Dialogue System Based on Term Modeling: Current medical dialogue systems mainly rely on directly fine-tuning base large models on medical dialogue corpora. However, this method limits the model's reasoning ability in complex medical dialogues, leading to non-targeted dialogue responses and potential hallucination phenomena. To enhance the reasoning ability of medical dialogue systems in medical tasks during dialogue and alleviate hallucination phenomena, we propose a framework of "dialogue model" + "term task model" and develop two term task models, namely term extraction tool and term prediction model. By invoking these term task models, the dialogue model can accurately model the flow of term information in dialogue, thus effectively performing task reasoning. In experiments on two Chinese multi-turn medical dialogue datasets, our proposed method significantly improves system performance, especially in terms of the accuracy of term information in dialogue.

        2. Spoken Term Extraction Method by Matching-based Semantic Pretraining: In real medical dialogue scenarios, patients often describe their condition in a colloquial manner, involving non-standard medical terminology. This non-standard expression of terminology is difficult to capture accurately with simple term extraction tools and may be overlooked by base large models. This inaccuracy in medical dialogue understanding may further exacerbate hallucination phenomena in dialogue responses. To address this issue, this work proposes a matching-based term semantic pre-training method to build a colloquial term understanding task model. This method designs two self-supervised tasks closely related to term extraction, namely contrastive term discrimination and matching-based masked term modeling, aiming to enhance the
model’s term semantic understanding ability using a large amount of unlabeled medical dialogue data. Experimental results show that this method effectively improves model performance under full data, low-resource, and zero-resource settings.

        3. Term-Status Pair Extraction Method Based on Knowledge-Enhanced Generative Framework: Using term information alone often cannot accurately describe specific medical conditions because terminology also carry corresponding status information, such as the positivity or negativity of symptom-type terms. In dialogue, the
co-reference nature of statuses and the characteristics of status changes as the dialogue progresses pose challenges for term status understanding. To address these issues, this work proposes a knowledge-enhanced two-stage generative framework, providing the dialogue system with richer term-related information by creating a term status pair extraction task model. This framework completes term and status extraction separately through the generation method by constructing different task prompts. Additionally, it utilizes background knowledge such as term categories and candidate statuses to enhance the task prompts in the status generation stage. This two-stage generation characteristic enables the framework to further improve performance in low-resource environments using only additional data containing term annotations.

        4. Medical Dialogue Diagnosis Method Based on Status-Enhanced Mask Prediction Framework: In medical dialogues, disease diagnosis is a crucial task, but general dialogue models are more prone to hallucination phenomena when dealing with such complex reasoning. To mitigate this issue, this work focuses on creating a term reasoning task model dedicated to dialogue disease diagnosis to improve the accuracy
and reliability of medical dialogue systems in disease diagnosis reasoning. To this end, we propose a unified mask prediction framework that integrates symptom inquiry and disease prediction as two types of term reasoning sub-tasks into a holistic task. Additionally, we design corresponding auxiliary tasks to enhance the modeling ability of
symptom and disease statuses. Experiments on multiple datasets validate the effectiveness of the proposed methods, demonstrating significant improvements in the performance of medical dialogue systems in disease diagnosis tasks.

Keyword医疗对话系统 任务型对话系统 对话理解 对话推理 幻觉现象
Subject Area自然语言处理
MOST Discipline Catalogue工学::计算机科学与技术(可授工学、理学学位)
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/56659
Collection毕业生_博士学位论文
Recommended Citation
GB/T 7714
胡泽发. 医疗领域任务型对话系统研究[D],2024.
Files in This Item:
File Name/Size DocType Version Access License
202406博士明版论文签字版_胡泽发.(3935KB)学位论文 限制开放CC BY-NC-SA
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[胡泽发]'s Articles
Baidu academic
Similar articles in Baidu academic
[胡泽发]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[胡泽发]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.