A Proposed Open-Domain Factoid Question Answering System for Noisy and Ambiguous Knowledge-Bases

Most advanced web search engines help users to retrieve relevant web pages for their questions in fraction of seconds. But in many case, users have to find an exact and short answer for their questions. In this situation we need an open Question answering (QA) system. QA problem is still considered as a challenging problem. This paper proposes a factoid open domain QA system on noisy knowledge bases which are extracted from the web, automatically. These noisy knowledge bases contain the extractions which are noisy (e.g., the string "Obama", "Barack Obama" and "president Obama" all appear as a distinct entities) and ambiguous (e.g., the relation "born in" contains facts about both date). We use noisy REVERB database contains a large crosssection of world knowledge and is a good testbed for developing an open domain QA system, as well as a new database which is created for any given question based on web, because the web is known as an attractive resource of knowledge for seeking the answer of questions. The proposed system combines PARALEX parser with natural language processing techniques and normalization methods (like removing adverbs and fixing misspelled words in questions) and creates a hybrid system which has better performance than PARALEX- a well-known open question answering system on noisy knowledge base. The proposed approach applies a similarity measure to improve the efficiency of the answer extraction method. Experimental results have shown that the proposed scheme significantly outperforms PARALEX, in terms of recall (14% better), precision (7.5 % better), f-measure (13 % better), and MRR (15% better).