Knowledge integration by probabilistic argumentation

Argumentation is a form of reasoning more general than some other symbolic forms developed in AI such as non-monotonic and defeasible reasoning. The recent integration of the Abstract Argumentation (AA) framework with probability theory results in different Probabilistic Argumentation (PA) frameworks notably PABA which uses assumption-based argumentation to structure arguments. Indeed, PABA subsumes many other reasoning formalisms especially probabilistic or argument-based ones and this provides a basic for this dissertation. In particular, we are interested in combining multiple knowledge sources currently represented different formalisms but their informational contents somehow overlap, for example one source may be a rule base while another source may be a Bayesian network but both concerns diagnosis of liver disorders and biliary tracts. Our first approach is to use PABA as a common language for knowledge representation. That is, we translate each given knowledge source from its original representation to PABA, and then combine the resulted PABA frameworks to arrive a single PABA framework. As the performance of the first approach, we demonstrate the strengths and weakness of the approach and compare it with existing approaches which base on some formalism different from PA. Our second approach is to combine a convolutional neural network model and case-based reasoning model through a structured argumentation framework called ABA for performing sentiment analysis. As the performance of the second approach, we evaluate and compare the results of our approach with other traditional machine learning techniques in term of accuracy.

论证（Argumentation）是一类比人工智能领域中开发的其他符号化推理形式（如非单调推理与可废止推理）更为通用的推理范式。近年来，抽象论证（Abstract Argumentation, AA）框架与概率论的融合催生了多种概率论证（Probabilistic Argumentation, PA）框架，其中尤为典型的是PABA——该框架依托基于假设的论证（Assumption-Based Argumentation, ABA）来构建论证结构。事实上，PABA能够涵盖诸多其他推理形式体系，尤其是概率类或论证类形式体系，这也为本论文的研究奠定了核心基础。具体而言，本研究聚焦于整合多种当前以不同形式体系表示的知识源：尽管这些知识源的信息内容存在一定重叠，例如一个知识源为规则库，另一个为贝叶斯网络，但二者均围绕肝脏疾病与胆道系统的诊断任务展开。我们的第一种研究方法是以PABA作为知识表示的统一语言。具体而言，我们将每个给定的知识源从其原始表示形式转换为PABA框架，随后将生成的多个PABA框架整合为单一的PABA框架。作为第一种方法的性能验证环节，我们将阐明该方法的优势与局限，并与其他基于非概率论证形式体系的现有方法进行对比。我们的第二种研究方法则是依托名为ABA的结构化论证框架，将卷积神经网络（Convolutional Neural Network, CNN）模型与基于案例的推理（Case-Based Reasoning, CBR）模型相结合，以实现情感分析任务。作为第二种方法的性能评估环节，我们将以准确率为评价指标，对本方法的结果与其他传统机器学习技术的结果进行评估与对比。