THU_NGN at SemEval-2019 Task 12: Toponym Detection and Disambiguation on Scientific Papers

Proceedings of the 13th International Workshop on Semantic Evaluation (SemEval-2019), pages 1302–1307 Minneapolis, Minnesota, USA, June 6–7, 2019. ©2019 Association for Computational Linguistics 1302 THU NGN at SemEval-2019 Task 12: Toponym Detection and Disambiguation on Scientific Papers Tao Qi , Suyu Ge , Chuhan Wu , Yubo Chen , Yongfeng Huang Tsinghua National Laboratory for Information Science and Technology, Department of Electronic Engineering, Tsinghua University Beijing 100084, China {qit16,gesy17,wuch15,chen-yb18,yfhuang}@mails.tsinghua.edu.cn Abstract Toponym resolution is an important and chal- lenging task in the neural language processing field, and has wide applications such as emer- gency response and social media geographi- cal event analysis. Toponym resolution can be roughly divided into two independent steps, i.e., toponym detection and toponym disam- biguation. In order to facilitate the study on toponym resolution, the SemEval 2019 task 12 is proposed, which contains three subtasks, i.e., toponym detection, toponym disambigua- tion and toponym resolution. In this paper, we introduce our system that participated in the SemEval 2019 task 12. For toponym detec- tion, in our approach we use TagLM as the ba- sic model, and explore the use of various fea- tures in this task, such as word embeddings extracted from pre-trained language models, POS tags and lexical features extracted from dictionaries. For toponym disambiguation, we propose a heuristics rule-based method using toponym frequency and population. Our sys- tems achieved 83.03% strict macro F1, 74.50 strict micro F1, 85.92 overlap macro F1 and 78.47 overlap micro F1 in toponym detection subtask. 1 Introduction Toponym resolution is an important task in the nat- ural language processing field and has many ap- plications such as emergency response and social media geographical event analysis(Gritta et al., 2018). Toponym resolution is usually modelled as a two-step task. The first step is toponym detec- tion, which is a typical named entity recognition (NER) task. The second step is toponym disam- biguation, which aims to map locations to its co- ordinates in the real world. NER is a widely explored task and most NER methods can be applied to toponym detection. For example, Ratinov and Roth (2009) used n-grams, history predictions as the input features of con- ditional random fields (CRF) for toponym detec- tion. Usually the performance of these methods heavily relies on the quality of hand-crafted fea- tures. However, manually selected features may be sub-optimal. Also, these methods cannot ef- fectively exploit contextual information due to the dependency on bag-of-word features. In recent years, many neural network based methods have been proposed for NER. For example, Ma and Hovy (2016) proposed a CNN-LSTM-CRF model for NER. They use CNN layer to learn character features of each word, LSTM layer to learn the contextual word representations and CRF layer to predict the label jointly. Gregoric et al. (2018) proposed Parallel RNN architecture. They split a single LSTM into multiple equally-size ones with a penalty to promote diversity. However, these methods cannot utilize external knowledge to rec- ognize entities, which is usually important to to- ponym detection. Usually, linguistic knowledge such as part-of-speech …

Content truncated to 500 words