Search Query Language Identification Using Weak Labeling

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3520–3527 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3520 Search Query Language Identification Using Weak Labeling Ritiz Tambi, Ajinkya Kale, Tracy Holloway King Adobe Inc. San Jose, California, USA {tambi, akale, tking}@adobe.com Abstract Language identification is a well-known task for natural language documents. In this paper we explore search query language identification which is usually the first task before any other query understanding. Without loss of generalization, we run our experiments on the Adobe Stock search engine. Even though the domain is relatively generic because Adobe Stock queries cover a broad range of objects and concepts, out-of-the-box language identifiers do not perform well due to the extremely short text found in queries. Unlike other well- studied supervised approaches for this task, we examine a practical approach for the cold start problem for automatically getting large-scale query-language pairs for training. We describe the process of creating weak-labeled training data and then human-annotated evaluation data for the search query language identification task. The effectiveness of this technique is demonstrated by training a gradient boosting model for language classification given a query. We out-perform the open domain text model baselines by a large margin. Keywords: language detection, language identification, search queries, short text, weak labeling 1. Introduction Language identification is key to accurately applying deeper natural language processing in contexts where the language of the text is not known. Relatedly, it can be used to clas- sify texts by language and hence ensure that a set of texts are all in the same language. In this paper we focus on search queries1 for Adobe Stock. Adobe Stock allows users to search over a set of images for purchase. The searches can use keywords, images, or a combination thereof; for language identification we are only concerned with the key- words. These keywords reflect vocabulary used to describe items in the world (e.g. dog, woman, coffee cup), abstract concepts (e.g. happiness, success), and some domain spe- cific vocabulary (e.g. icon, vector, background). We want to know the language of the query in order to apply processes like named entity recognition and dependency parsing to improve query understanding and hence search result qual- ity. In addition, we want to provide auto-complete sugges- tions and related searches in the correct language, i.e. the language of the user’s issued queries. Language identification systems are generally trained on well-edited text. They use a variety of signals includ- ing: known words and especially known closed class words (e.g. determiners, auxiliaries); character n-grams (often tri- grams); punctuation and spaces, which can be viewed as a subset of characters; upper vs. lower casing, which can be viewed as a subset of characters. These general purpose language identification models work well for longer texts which have similar editing style. However, their perfor- mance drops significantly on shorter texts such as search queries and shorter tweets (Cavnar and Trenkle, 1994; Si- bun and Reynar, 1996; …

Content truncated to 500 words