Translating Embeddings for Modeling Query Reformulation

Query reformulation understanding is important for Information Retrieval (IR) tasks, such as search results reranking and query recommendation. Conventional works rely on the textual content of queries to understand reformulation behaviors, which suffer from data sparsity problems. To address this issue, We propose a novel method to efficiently represent the behaviors of query reformulation by the translating embedding from the original query to its reformulated query. We utilize two-stage training algorithm to make the learning of multilevel intentions representation more adequate. We construct a new corpus of shopping search query log and create a query reformulation graph based on this dataset. Referring to knowledge graph embedding methods, we use the accuracy of intentions prediction to evaluate experimental results. Our final result, an increase of 20.6% of the average prediction accuracy in 21 intentions, shows significant improvement compared to baselines.