پاسخ‌دهي به سوال‌هاي فارسي با استفاده از داده‌هاي پيوندي

The challenge of Persian knowledge based question answering systems is the insufficiency of structured data for the Persian language. In this thesis, for the problem of answering Persian questions using an English knowledge base, the sub-problem of relation detection for single relation questions has been examined in detail. In this kind of questions, the answer is obtained from a single KB fact in the form of subject-predicate-object. Conventional fact extraction methods have two steps: entity linking and relation detection. In the first step, the entity identified in the question is mapped to a subject or object of a triple, and in the second step, a predicate that shows the semantic relation in the question is identified. In most of the previous methods, after the entity is identified, all the relations that exist for it in the knowledge base are considered as candidate relations and the best one is selected. This study explored relation detection task without entity-linking step utilizing the hierarchical structure of relations. The relation of a question is detected in three steps by three BiGRU neural networks. For the input of these networks, one-hot encoding is used at the character level. Besides, to increase the accuracy of the proposed model and to avoid the need for large amounts of training data, after each word of the question, its POS tag is inserted before feeding the network. The experimental results show that the accuracy of the proposed solution for the direct relation detection is 89.5%. In addition, the proposed solution can be used for the indirect relation detection whose accuracy is 96.3%, which is higher than state-of-the-art relation detection techniques. For Persian language, direct and indirect detection accuracies are 72% and 90%, respectively. In addition, the positive effects of using POS labels have been investigated. In this research, the SimpleQuestions dataset has been used, which includes English language questions. To eva‎luate the proposed method for Persian language, this dataset was effectively translated into Persian language using Google translator. Finally, a system for answering Persian questions with Wikidata knowledge base is proposed, which uses the combination of three independent methods for entity linking and a combination of two direct and indirect methods for relation detection.

5