Statistical and Semantic Question Analysis for Situated Question Answering

Abstract

Natural Language Processing (NLP) aims to study techniques and systems for processing and eventually understanding natural language speech and text. Typical NLP tasks include speech recognition, natural language generation, machine translation, question answering (QA) and so on. As natural language is still the most natural and indispensable way to convey knowledge and exchange ideas in e-learning. It is very beneficial to use the state-of-the art NLP technology to support efficient and adaptive teaching and learning.

This project concentrates on question answering, which aims to automatically extract answers from relevant resources to natural language questions. First of all, we apply deep analysis on questions, as question analysis is the first and key step in QA and can significantly affect the performance of QA systems. Both statistical and semantic techniques are explored in question analysis for supporting QA situated in specific educational environments. The questions are parsed and semantic roles in the questions are recognized by the help of VerbNet, a lexical semantics resource that incorporates both semantic and syntactic information about verbs and their semantic roles. The semantic frames of both questions and its possible answers are constructed and compared to filter out the most possible answers to the questions. We also defined a set of features that we think have strong influence on the semantic similarity of two sentences, and incorporate machine learning algorithm to learn the structure and weights of the features. The learned classifier is then used to decide the best answers to a question. A prototype experimental question answering system based on the proposed structure and methods has been built for demonstration and experiments. The experimental results have been presented in our paper and report.

An online presentation/demo is also available for a short introduction of this research work. The first part shows the use of an NLP server for a deeper question analysis in a search engine like question answering application designed for facilitating students' access and retrieval of knowledge and information from the learning materials in the form of natural language text. The second part shows how NLP parsing tools analyze natural language sentences, and thus provide information for further understanding of the natural language questions and answers. In the third part, an interface of our NLP based QA system shows the main underlying processes of the system step by step, from a user query, syntactic parsing, to semantic analysis, searching and matching, and finally lead to the possible answer sets of the query. The last two parts are experiments on semantic analysis and feature learning for automated question answering respectively. They evaluate the performance of our methods for a set of test sentences against the test corpora, and serve as platforms for us to develop effective QA system of the research program.