Purpose: In this paper, we attempt to use query refinements to identify users’ search intents and seek a method for intent clustering based on real world query data. Design/methodology/approach: An experiment has been conducted to analyze selected search sessions from the American Online(AOL) query logs with a two-stage approach. The first stage is to identify underlying intent by combining query co-occurrence information with query expression similarity. The work in the second stage is to cluster identified results by constructing query vectors through performing random walks on a Markov graph. Findings: Average correctness for identifying search intent is 0.74. Precision, recall, F-score values for intent clustering are 0.73, 0.72 and 0.71,respectively. The results indicate that combining session co-occurrence information and query expression similarity can further filter noises and our clustering method is more suitable for sparse data. Research limitations: We use the time-out threshold(15-minutc) method to group queries in one session, but a user may have multiple search goals at the same time and the multi-task behavior of a user is hard to capture in a session defined based on time notions. Practical implications: This study provides insights into the ways of understanding users’ search intents by analyzing their queries and refinements from a new perspective. The results will help search engine developers to identify user intents. Originality/value: We propose a new method to identify users’ search intents by combining session co-occurrence information and query expression similarity, and a new method for clustering sparse data. Purpose: In this paper, we attempt to use query refinements to identify users’ search intents and seek a method for intent clustering based on real world query data. Design / methodology / approach: An experiment has been conducted to analyze selected search sessions from the The first stage is to identify an underlying intent by combining query co-occurrence information with query expression similarity. The work in the second stage is to cluster identified results by constructing a query vector through Findings: Average correctness for identifying search intent is 0.74. Precision, recall, F-score values ​​for intent clustering are 0.73, 0.72 and 0.71, respectively. The results indicate that combining session co-occurrence information and query expression similarity can further filter noises and our clustering method is more suitable for sparse data. Research limitations: We use the time-out thre shold (15-minutc) method to group queries in one session, but a user may have multiple search goals at the same time and the multi-task behavior of a user is hard to capture in a session defined based on time notions. Practical implications : This study provides insights into the ways of understanding users ’search intents by analyzing their queries and refinements from a new perspective. The results will help search engine developers to identify user intents. Originality / value: We propose a new method to identify users’ search intents by combining session co-occurrence information and query expression similarity, and a new method for clustering sparse data.