Using natural language processing for question answering in closed and open domains

• Autores: Majid Latifi
• Directores de la Tesis: Horacio Rodríguez Hontoria (dir. tes.) , Miquel Sànchez i Marrè (dir. tes.)
• Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2018
• Idioma: español
• Tribunal Calificador de la Tesis: María Antonia Martí Antonín (presid.) , Antonio Moreno Ribas (secret.) , Bernardo Magnini (voc.)
• Enlaces
  • Tesis en acceso abierto en: TDX
• Resumen

  • With regard to the growth in the amount of social, environmental, and biomedical information available digitally, there is a growing need for Question Answering (QA) systems that can empower users to master this new wealth of information. Despite recent progress in QA, the quality of interpretation and extraction of the desired answer is not adequate. We believe that striving for higher accuracy in QA systems is subject to on-going research, i.e., it is better to have no answer is better than wrong answers. However, there are diverse queries, which the state of the art QA systems cannot interpret and answer properly.

    The problem of interpreting a question in a way that could preserve its syntactic-semantic structure is considered as one of the most important challenges in this area. In this work we focus on the problems of semantic-based QA systems and analyzing the effectiveness of NLP techniques, query mapping, and answer inferencing both in closed (first scenario) and open (second scenario) domains. For this purpose, the architecture of Semantic-based closed and open domain Question Answering System (hereafter “ScoQAS”) over ontology resources is presented with two different prototyping: Ontology-based closed domain and an open domain under Linked Open Data (LOD) resource.

    The ScoQAS is based on NLP techniques combining semantic-based structure-feature patterns for question classification and creating a question syntactic-semantic information structure (QSiS). The QSiS provides an actual potential by building constraints to formulate the related terms on syntactic-semantic aspects and generating a question graph (QGraph) which facilitates making inference for getting a precise answer in the closed domain. In addition, our approach provides a convenient method to map the formulated comprehensive information into SPARQL query template to crawl in the LOD resources in the open domain.

    The main contributions of this dissertation are as follows: 1.Developing ScoQAS architecture integrated with common and specific components compatible with closed and open domain ontologies.

    2.Analysing user’s question and building a question syntactic-semantic information structure (QSiS), which is constituted by several processes of the methodology: question classification, Expected Answer Type (EAT) determination, and generated constraints.

    3.Presenting an empirical semantic-based structure-feature pattern for question classification and generalizing heuristic constraints to formulate the relations between the features in the recognized pattern in terms of syntactical and semantical.

    4.Developing a syntactic-semantic QGraph for representing core components of the question.

    5.Presenting an empirical graph-based answer inference in the closed domain.

    In a nutshell, a semantic-based QA system is presented which provides some experimental results over the closed and open domains. The efficiency of the ScoQAS is evaluated using measures such as precision, recall, and F-measure on LOD challenges in the open domain. We focus on quantitative evaluation in the closed domain scenario. Due to the lack of predefined benchmark(s) in the first scenario, we define measures that demonstrate the actual complexity of the problem and the actual efficiency of the solutions. The results of the analysis corroborate the performance and effectiveness of our approach to achieve a reasonable accuracy.