UNH at SemEval-2019 Task 12: Toponym Resolution in Scientific Papers

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1308–1312 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1308 UNH at SemEval-2019 Task 12: Toponym Resolution in Scientific Papers Matthew Magnusson Department of Computer Science University of New Hampshire mfm2@cs.unh.edu Laura Dietz Department of Computer Science University of New Hampshire dietz@cs.unh.edu Abstract The SemEval-2019 Task 12 is toponym resolu- tion in scientific papers. We focus on Subtask 1: Toponym Detection which is the identifi- cation of spans of text for place names men- tioned in a document. We propose two meth- ods: 1) sliding window convolutional neu- ral network using ELMo embeddings (CNN- ELMo), and 2) sliding window multi-Layer perceptron using ELMo embeddings (MLP- ELMo). We also submit a bi-directional LSTM with Conditional Random Fields (bi- LSTM) as a strong baseline given its state- of-art performance in Named Entity Recogni- tion (NER) task. Our best performing model is CNN-ELMo with a F1 of 0.844 which was be- low bi-LSTM F1 of 0.862 when evaluated on overlap macro detection. Eight teams partici- pated in this subtask with a total of 21 submis- sions. 1 Introduction Toponyms are textual spans of text identifying geospatial locations. This can range from the canonical name of populated places, such as “Lon- don” to direct or indirect mentions of geographic entities. The parsing of geographic locations from unstructured text is considered an open challenge due to domain diversity, place name ambiguity, metonymic language and often limited leverag- ing of context (Gritta et al., 2018). Many scien- tific publications contain toponyms which can be challenging to extract automatically. Specifically, names of institutions and viruses often contain ge- ographic references which may confuse the ex- tractor. Often, the extractor needs to handle noisy text parsed from PDF versions of scientific articles which can introduce artifacts. In Task 12, a toponym is defined to include proper names and geographic entities but to ex- clude indirect mentions of places and metonyms. Additional discussion of the motivation and task description is available at the task web site.1 2 Related Work There is significant work in the area of toponym detection (Matsuda et al., 2015; D. Lieberman et al., 2010) and the closely related fields of named entity recognition (NER) (Li et al., 2018) and en- tity mention detection (EMD) (Shen et al., 2015) with many different approaches. State-of-the-art named entity detection models have historically employed a combination of hand-crafted features, rules, natural language processing (NLP), string- pattern matching, and domain knowledge using supervised learning on manually annotated cor- pora (Piskorski and Yangarber, 2018). A common approach to toponym detection has been to utilize place name gazetteers which are directories of ge- ographic names and their corresponding geoloca- tions to perform string matching of place names in text (D. Lieberman et al., 2010). Contemporary approaches in entity detection have utilized neural-based architectures. (Col- lobert et al., 2011) propose a window-based, multi-layer, dense feed-forward neural architec- ture using word embeddings concatenated with orthographic features and a gazetteer as …

Content truncated to 500 words