Abstract:
The generic Question Answering (QA) framework processes questions by querying a knowledge base and extracting answers from retrieved passages using various Natural Language Processing techniques. The problem is validating whether the retrieved passages from the passage retrieval module contain expected answers to asked questions. Besides, extraction based on lexical and syntactic similarities alone is not enough coverage for scoring the correct answers in a QA framework. Therefore, this work aims to infuse validation techniques into the QA framework. Four similarity scores (Word Form (WF), Word Order (WO), Distance (DIST), and Semantic Similarity (SemSim)) were implemented for Answer Extraction. Instant snippets returned by the Google search engine were used as a corpus to generate candidate answer sets. On a dataset of 1370 factoid questions, the proposed method achieved an accuracy of 77.71%, precision of 77.91%, recall of 91.37%, and F1-measure of 91.37%. The results show that the inclusion of the validation techniques helps reduce the time spent by the system in analyzing passages without possible answers. The proposed system could be adapted for automatic QA Systems and grading factoid computer-based tests.
Machine summary:
Keywords: Factoid questions, Question answering, Semantics, Open domain, Textual Entailment Introduction QA system is an effective automatic technique that provides appropriate answers to the queries issued by humans in a natural language pattern (Dwivedi & Singh, 2013).
Open-domain QA systems are not limited to a specific domain; instead, they depend on universal ontology and world knowledge (Allam & Haggag, 2012; Kumar & Sharma, 2014; Sun, Dhingra, Zaheer, Mazaitis, Salakhutdinov & Cohen, 2018).
Related Works Most times, users on the internet are interested in extracting precise and relevant information about factoid questions to retrieve a complete document on the Web (Olvera- Lobo & Gutiérrez-Artacho, 2010).
Cases of open-domain QA systems include the following: Webclopedia (Hovy, Gerber, Hermjakob, Junk & Lin, 2000), Mulder (Kwok, Etzioni & Weld, 2001), QAPD (Abdi & Ahmad, 2018), and the ones also in Lin et al.
Abacha and Demner-Fushman (2019) examined question entailment about the medical domain and the usefulness of the end-to-end RQE-based QA approach by evaluating the relevance of retrieved answers.
The extracted paragraphs and questions are re-formulated as a textual entailment problem adopting the methods of Novotny (2018) and Zhu, Wong and Chao (2014) for cosine similarity and chunking measure computation, respectively.
Then, the maximum function was used to estimate the semantic similarity rate between words as in Equation (10) (Song, Feng, Gu & Wenyin, 2007; Pilehvar & Navigli, 2015): SemSim(w1, w2) = max(c1, c2)esyn(w1)xsyn(w2 Sim(c1, c2) (10) Also, given the question q and the answer a, q contains these keywords wt .