Structure of the Web Searcher’s Query Modifications: Sequential, Branching, Merging, Re-Merging and Non-Sequential Execution

This A non-linear model of query modifications during a single session of user searching on the Web is evaluated. In addition to a sequential query modification, in which any query may be reformulation of its direct predecessor, the model allows non-linear interpretation of query reformulations: branching, merging, re-merging (merging of earlier branched chains) and non-sequential execution of logically linear chains of queries. To detect short-term (i.e. during a single time session of interaction with the Web search engine) dependencies between user’s queries we use different query-to-query similarity metrics and consider query-on-query(ies) dependency as similarity of current query to queries submitted earlier in a time session. Usage of different metrics leads to different but complimentary results. The method is applied to query logs of the Yandex, the major Russian search engine. Regardless of difference of results of different metrics, the findings are coincide: about 20% of time sessions containing 3 and more distinct queries contain such combinations of queries which can be interpreted as non-linearities: branchings are occurred in 12% of sessions; mergings are occurred in 6%, and re-mergings in 3%. Number of branches are a little bigger than two; the first branch is smaller then the last. A merging search is frequently executed in the re-merging manner. The last merging chain is smaller then the first one. Keywords— query modification, branching, merging, re-merging