Toponym Detection in the Bio-Medical Domain: A Hybrid Approach with Deep Learning

Proceedings of Recent Advances in Natural Language Processing, pages 912–921, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_106 912 Toponym Detection in the Bio-Medical Domain: A Hybrid Approach with Deep Learning Alistair Plum, Tharindu Ranasinghe, Constantin Or˘asan Research Group in Computational Linguistics, University of Wolverhampton, UK {a.j.plum, t.d.ranasinghehettiarachchige, c.orasan}@wlv.ac.uk Abstract This paper compares how different ma- chine learning classifiers can be used to- gether with simple string matching and named entity recognition to detect loca- tions in texts. We compare five differ- ent state-of-the-art machine learning clas- sifiers in order to predict whether a sen- tence contains a location or not. Fol- lowing this classification task, we use a string matching algorithm with a gazetteer to identify the exact index of a toponym within the sentence. We evaluate different approaches in terms of machine learning classifiers, text pre-processing and loca- tion extraction on the SemEval-2019 Task 12 dataset, compiled for toponym reso- lution in the bio-medical domain. Fi- nally, we compare the results with our sys- tem that was previously submitted to the SemEval-2019 task evaluation. 1 Introduction The task of toponym resolution (TR) is a topic in natural language processing (NLP) which is aimed at extracting locations from texts. TR includes the sub-tasks of detecting, disambiguating and finally resolving and assigning coordinates to the proper location in the text. The first step is to detect a location in a text, which can be carried out by more simple means of gazetteer lookup or named entity recognition (NER) (Piskorski and Yangar- ber, 2013). Next, each detected location that is ambiguous needs to be distinguished and the cor- rect location chosen (Leidner et al., 2004). The final step involves assigning coordinates or apply- ing some other meta-information to clearly distin- guish each proposed location (Smith and Crane, 2001; Leidner et al., 2004). TR can involve many challenges, caused by cases that are not trivial to resolve. This includes, for example, locations contained in phrases such as London Bus Company, which contains Lon- don but refers to an organisation and should therefore not be marked as a location. Further- more, locations can be contained in other types of phrases, such as genome sequences mentioned in bio-medical texts, as seen in the SemEval-2019 data (Weissenbacher et al., 2019). For exam- ple, the location Henan in the string A/Tree spar- row/Henan/4/04 could be omitted from the results, as it is part of the name of a genome. Similar prob- lems can occur where a location might be detected in a character sequence denoting a chemical. The increasing availability of online resources has been beneficial to TR. On the one hand large-scale geographical databases, such as GeoN- ames1, make information about many different lo- cations easily and freely available. On the other hand, readily available mapping services, such as Google Maps2, allow users to visualise these lo- cations. Although some of these services, such as GeoNames, have been used since the beginnings of TR (Smith and Crane, 2001; Leidner et al., 2004), they are more complete today. …

Content truncated to 500 words