Bert word embeddings'i temel alan sorgu genişletme ile belge sıralamasını geliştirme

In this thesis, we present a query expansion approach based on contextualized word embeddings for improving document ranking performance. We employ Bidirectional Encoder Representations from Transformers(BERT) word embeddings to expand the original query with semantically similar terms. After deciding the best method for extracting word embeddings from BERT, we extend our query with the best candidate terms. As our primary goal, we show how BERT performs over the Word2Vec model, known as the most common procedure for representing terms in the vector space. After that, by leveraging the relevance judgment list, we show positive contributions of integrating tf-idf and term co-occurrence properties of terms to our query expansion system. Our experiments demonstrate that BERT outperforms Word2Vec in well-known evaluation metrics. In addition, we also conduct several experiments that address the most popular issues in information retrieval systems.