Multi-level Persian Dataset for Information Retrieval

Abstract

An information retrieval system tries to retrieve documents related to a question/query. The retrieval is done from a large<br />Information retrieval systems are an essential part of many smart systems. The applications of this research field include search engines such as Google and Bing, question-answering systems, modern databases, etc. An information retrieval system tries to retrieve documents related to a question/query. The retrieval is done from a large collection of documents, and the size of this collection can be from a few thousand documents to millions of documents. In recent years, a lot of research has been done to develop information retrieval systems using language models. However, in this research field, no research has been done for the Persian language. One of its main reasons is the lack of a suitable Persian dataset for training language models. In this research, first, a Persian dataset for information retrieval is presented. After that, methods for enriching this data set are investigated. This enrichment is done by defining multi-level relationships between a document and a question. In this regard, the new dataset can show the relationship between question and document in four levels (unrelated, related, highly related, completely related) instead of two levels (completely unrelated, completely related). The name of the generated dataset is PersianMLIR. Experiments show that by using multi-level relationships, the performance of the system improves for both Persian and English languages, where the improvement is 1.87% for the Persian language. The results conclude that enriching information retrieval datasets by increasing the number of relations between query and document lead to improving the performance of information retrieval systems.<br />