Spelling Correction of User Search Queries through Statistical Machine Translation

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 451–460, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Spelling Correction of User Search Queries through Statistical Machine Translation Saša Hasan∗ sasa.hasan@gmail.com Carmen Heger† heger.carmen@gmail.com Saab Mansour saamansour@ebay.com eBay Inc. 2065 Hamilton Ave San Jose, CA 95125, USA Abstract We use character-based statistical machine translation in order to correct user search queries in the e-commerce domain. The training data is automatically extracted from event logs where users re-issue their search queries with potentially corrected spelling within the same session. We show results on a test set which was annotated by humans and compare against online autocorrection capabilities of three addi- tional web sites. Overall, the methods presented in this paper outperform fully productized spellchecking and autocorrec- tion services in terms of accuracy and F1 score. We also propose novel evaluation steps based on retrieved search results of the corrected queries in terms of quantity and relevance. 1 Introduction Spelling correction is an important feature for any interactive service that takes input generated by users, e.g. an e-commerce web site that allows searching for goods and products. Misspellings are very common with user-generated input and the reason why many web sites offer spelling cor- rection in the form of “Did you mean?” sug- gestions or automatic corrections. Autocorrec- tion increases user satisfaction by correcting ob- vious errors, whereas suggestions make it con- venient for users to accept a proposed correction without retyping or correcting the query manually. Spelling correction is not a trivial task, as search ∗ The author is now affiliated with Lilt Inc., Stanford, CA, USA. † The author is now affiliated with Stylight GmbH, Mu- nich, Germany. queries are often short and lack context. Mis- spelled queries might be considered correct by a statistical spelling correction system as there is ev- idence in the data through frequent occurrences. While common successful methods (cf. Sec- tion 1.1) rely on either human-annotated data or the entire web, we wanted to use easily accessi- ble in-domain data and on top of that technology that is already available. In this work, we use user event logs from an e-commerce web site to fetch similar search query pairs within an active session. The main idea is that users issue a search query but alter it into something similar within a given time window which might be the correction of a poten- tial typo. To the best of our knowledge, the idea of collecting query corrections using user session and time information is novel. Previous work sug- gested collecting queries using information that a user clicked on a proposed correction. Our pro- posed method for collecting training data has sev- eral advantages. First, we do not rely on a pre- vious spelling correction system, but on user for- mulations. Second, for many search queries, es- pecially from the tail where search recall is gen- erally low, these misspellings yield few results, and thus, users looking for certain products are …

Content truncated to 500 words