Abstract

Although the diversity of results has been studied since the early information retrieval systems, few studies explore diversity and its representation in an educational context. Inherently, approaches that seek to address difficulties in web search are focused on maximizing the relevance of results over the original query. This work presents a method that integrates semantic relationships using Word Embedding for expansion with blind feedback to improve diversity. Using a corpus based on the user's query logs from a realistic setting, three Word2vec models are trained to obtain semantically relevant terms for each naturally elaborated query by students. The proposed architecture is studied in a specific search task, limiting the number of candidate terms in each model according to the allowed frequency of words. Finally, the diversity in two groups of queries is compared, measuring the lexical similarity of the snippets of the results pre-expansion and post-expansion. Results indicate the potential for improving diversity, also showing that lower semantic similarity can lead to better diversity. Therefore, we provide a method to improve learning through web searches.