毕业生知识库

问答系统旨在为用户提出的自然语言问题提供精准答案。问答系统不仅是机器智能水平的重要验证手段，也是智能服务的主要形态，适用于智能助理、自动客服和搜索引擎等广泛的应用场景，具有重要研究意义和应用价值。由于文本具有获取容易、表达自然、覆盖广泛等优势，目前较多问答系统依赖文本数据作为知识来源，这类文本问答系统（Textual Question Answering，TQA）是问答系统的重要分支。

近年来，得益于深度学习技术的发展，文本问答系统得以快速发展。文本问答模型已经具备一些简单的理解文本和回答问题的能力。但是，目前的文本问答系统依然面临以下问题：（1）现有文本问答数据集主要关注简单类型的问题，未能包含更多样的复杂问题，同时也缺乏多个证据的详细求解过程，不能很好反映真实场景中的问答需求。（2）现有文本问答方法依赖于对篇章级文本进行整体建模，没有对多个证据句子的细粒度推理过程的建模，在检索证据回答各类复杂问题时缺乏可解释性推理能力。为了解决上述问题，本文从数据集构建、方法研究等方面对文本问答任务中的复杂问题展开系统性探索。整体而言，我们构造了具有细粒度推理过程描述的多证据文本库问答数据集，同时提出了能够获取链式、图式证据链的文本问答方法，以提高文本问答系统的智能水平和可解释性能力。主要的创新点和研究成果包括：
    面向文本事实库的多证据问答数据集构建：针对当前文本问答数据集缺乏多样化复杂问题及清晰推理过程的问题，本文构建了句子级文本事实库为来源的文本问答数据集，不仅覆盖了多样化的复杂问题，同时每个问题还包含句子级的显式证据推理过程，为文本问答提供证据过程支撑。具体地，首先，为了获取文本描述的事实库，本文利用数据文本生成技术（Data-to-Text），将结构化知识库自动转化为文本数据库。文本事实库中的每个证据是包含事实（fact）描述的句子级文本，由若干知识图谱事实三元组（子图）转换而来。其次，为了获取更丰富复杂问题的多样化问题，本文利用知识图谱问答数据集并进行相应转换，因为知识图谱问答包含了形式化的推理过程，能够将推理步骤对齐到知识图谱三元组，基于这样的形式化描述能够得到描述每个问题求解的推理过程。本文构建的多证据问答的文本事实库数据集在广泛性和推理深度上都比之前的数据集有很大的提升，同时能够为复杂问题提供准确的支持证据过程。通过自动评估和人工验证，本文构建数据集质量较高，达到了应用水平。
    
     基于路径链式证据检索的文本问答方法: 针对当前文本问答检索方法没有充分利用证据链的完整信息，进而导致长推理路径的复杂问答任务上表现不佳的问题，本文提出了一个基于链式证据路径检索的文本问答检索方法。具体地，首先，本文从路径角度对证据文本建模，并在训练过程中引入单步损失和整体损失，从整体和部分两个角度对证据表示进行学习优化；其次，在证据链解码阶段，本文使用惩罚解码机制实现路径证据的高效解码。最后，为了综合评价方法的有效性，本文设计了有顺序监督和无顺序监督的训练方式，能够适用于更广泛的应用场景。在数据集上的大规模实验表明，该方法在有顺序监督和无顺序监督的场景下，都能够提高证据链检索准确率和文本问答效果。
  
    基于双向图证据检索的文本问答方法: 针对现有文本问答方法仅限于单链或单跳证据检索方法，不能有效处理复杂问句所需的图式证据结构的问题，本文提出了一种能够检索多链多跳的证据图检索问答方法。我们从前向和后向两个相反方向检索初步证据子图，然后将它们进行融合得到一个完整的证据图。本文提出的方法包含用于证据图检索的双向图跳检索模块和用于进一步解释图生成的子图重建模块，能推广到处理多种推理类问题的场景。一方面，双向检索方法能够缓解模型单向检索随着推理过程深度增加呈现准确率降低的问题；另一方面，该方法能够利用逆向思维提高证据图的检索准确率。在文本问答数据集上的实验表明，方法能够有效提升单链单跳、单链多跳、多链单跳和多链多跳等多种证据结构的的文本问答性能。

Text Question Answering (QA) system is an important research direction in Natural Language Processing, where knowledge retrieval and reasoning are the core capabilities. Real-world question answering involves various complex question types, requiring QA systems to have the ability to support various types of knowledge reasoning and have clear QA processes. However, existing text QA datasets lack in-depth and comprehensive sorting of question types and do not have clear reasoning processes. This paper focuses on complex questions and evidence retrieval in text QA. We construct the first multi-evidence QA dataset with fine-grained multi-structure reasoning processes and propose multiple targeted methods to solve reasoning problems using text knowledge bases and pre-trained language models. This research is not only important for the intelligence and interpretability of QA systems but also for exploring the knowledge reasoning abilities of pre-trained language models. We conduct our research based on a text fact library. The main achievements of this work are as follows:

\textbf{Multi-evidence QA dataset based on Textual Fact Databases}: We construct a database called "Textual Fact Database" where each text contains knowledge fact sentences at the sentence level, converted from knowledge graph triplets. The text facts are mutually independent and non-overlapping, but can be related to each other, forming a structure graph composed of relationships. Complex QA tasks should be able to measure a model's ability to reason based on explicit knowledge and include QA solving processes. We first use structured knowledge to text generation technology to convert a structured knowledge base into a natural language text database. Furthermore, since KBQA datasets contain formal reasoning processes that can align reasoning steps to knowledge graph triplets, we map the triplets to text to provide a reasoning process from the question to the answer for each question. Compared with previous datasets, our dataset has significantly improved in terms of breadth and depth of reasoning and can automatically provide accurate supporting evidence processes for complex questions.

\textbf{Chain evidence retrieval method based on path modeling}: Based on the retriever-reader architecture, existing multi-hop retrieval models solve complex multi-hop problems through iterative computation, mostly using discrete retrieval methods that do not fully utilize the complete information of evidence chains and perform poorly for questions with longer reasoning paths. Some methods use pre-defined aggregation modules to combine sub-answers, which heavily rely on pre-set question patterns and can only handle fixed types of questions. To fully utilize the evidence chain information, we model the overall path information from a global perspective and introduce single-step and overall chain loss functions to calculate similarity loss, learning both the overall and partial perspectives. This method supports both unordered unsupervised training and ordered supervised training, improving both evidence chain retrieval accuracy and QA accuracy.

\textbf{Evidence graph retrieval method based on bidirectional graph}: Although existing text QA retrieval methods can directly retrieve relevant paragraphs complementarily, they cannot retrieve structured evidence graphs, which limits the model's reasoning and interpretability abilities. Previous models typically use single-directional retrieval to obtain supporting evidence and cannot model the structure of evidence (e.g., evidence order) when answering complex questions. In addition, due to the complex structure of multiple chains and hops, single-directional retrieval models may require a large search space. We propose a graph-jump retrieval and reasoning method that includes bidirectional graph-jump retrieval (BGR) and a subgraph reconstruction module for graph-jump evidence retrieval and explanatory graph generation. This greatly reduces the constraints on the evidence graph structure and can be extended to various reasoning problems. Bidirectional retrieval alleviates the problem of the model's accuracy decreasing as the depth increases in single-directional retrieval and the backward search can provide the model with reverse thinking, greatly improving the accuracy of evidence graph retrieval.