TR-SEQ: Named Entity Recognition Dataset for Turkish Search Engine Queries

Proceedings of Recent Advances in Natural Language Processing, pages 1417–1422 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_158 1417 TR-SEQ: Named Entity Recognition Dataset for Turkish Search Engine Queries Berkay Topcu AI Industry Solutions Turkcell Technology ˙Istanbul, Turkey berkay.topcu@turkcell.com.tr ˙Ilknur Durgar El-Kahlout NLP Technologies Turkcell Technology ˙Istanbul, Turkey ilknur.durgar@turkcell.com.tr Abstract Recognizing named entities in short search en- gine queries is a difficult task due to their weaker contextual information compared to long sentences. Standard named entity recog- nition (NER) systems that are trained on gram- matically correct and long sentences fail to per- form well on such queries. In this study, we share our efforts towards creating a cleaned and labeled dataset of real Turkish search en- gine queries (TR-SEQ) and introduce an ex- tended label set to satisfy the search engine needs. A NER system is trained by applying the state-of-the-art deep learning model BERT to the collected data and its high performance on search engine queries is reported. More- over, we compare our results with the state-of- the-art Turkish NER systems. 1 Introduction Named entity recognition (NER) aims to identify the predetermined entity categories (person, orga- nization, location, etc.) accurately in a given text. NER studies which is one of the most widely stud- ied subjects in the field of natural language pro- cessing (NLP), has been started for English and spread over a range of languages including Turk- ish (Yeniterzi, 2011). Initial NER studies for Turk- ish used small labeled datasets that contain gram- matically correct and typo-free sentences. Since manual labeling is a time-consuming and a costly process, gathering larger datasets and manually labeling them has not been feasible. With the in- crease of internet use, correctly classifying named entities has become a crucial need especially for social media entries and search engine (SE) queries. In search engines, identifying entities such as fa- mous persons, places or dates in user queries is crucial in order to accurately determine the infor- mation to be displayed to the user and response with the right information. For example, if the user is searching for a famous person, SE should return the Wikipedia information about the person. More- over if this famous person is an artist, the identified named entity assists in conveying the details of his/her movie/album/artwork details. Similarly, it is important to determine the location in the query for which we want to know the weather. The state-of-the-art Turkish NER tools fail to perform well on search engine queries due to two main reasons. The first reason is that most of the NER tools that have been developed for Turkish so far generally carry out categorization studies with three classes (Person, Organization and Lo- cation). Naturally, a three-class model cannot de- tect all the named entities in a SE query that is required to create a successful SE response. For this reason, it has become a necessity to determine the entity types specific to the SEs and to make a special development for these types. The sec- ond reason is that NER …

Content truncated to 500 words