When, Where, Who, What or Why? A Hybrid Model to Question Answering Systems

摘要

Question Answering Systems is a field of Information Retrieval and Natural Language Processing that automatically answers questions posed by humans in a natural language. One of the main steps of these systems is the Question Classification, where the system tries to identify the type of question (i.e. if it is related to a person, time or a location) facilitate the generation of a precise answer. Machine learning techniques are commonly employed in tasks where the text is represented as a vector of features, such as bag-of-words, Term Frequency-Inverse Document Frequency (TF-IDF) or word embeddings. However, the quality of results produced by supervised algorithms is dependent on the existence of a large, domain-dependent training dataset which sometimes is unavailable due to labor-intense of manual annotation of datasets. Normally, word embedding presents a related better performance on small training sets, while bag-of-words and TF-IDF presents better results on large training sets. In this work, we propose a hybrid model that combines TF-IDF and word embedding in order to provide the answer type to text questions using small and large training sets. Our experiments using the Portuguese language, using several different sizes of training sets, showed that the proposed hybrid model statistically outperforms bag-of-words, TF-IDF, and word embedding approaches.