Crossroads, Buildings and Neighborhoods: A Dataset for Fine-grained Location Recognition

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3329 - 3339 July 10-15, 2022 ©2022 Association for Computational Linguistics 1 Crossroads, Buildings and Neighborhoods: A Dataset for Fine-grained Location Recognition Pei Chen1 Haotian Xu1 Cheng Zhang2 Ruihong Huang1 1 Department of Computer Science and Engineering, Texas A&M University 2 Purdue University Northwest {chenpei, hx105, huangrh}@tamu.edu zhan4168@pnw.edu Abstract General domain Named Entity Recognition (NER) datasets like CoNLL-2003 mostly an- notate coarse-grained location entities such as a country or a city. But many applications re- quire identifying fne-grained locations from texts and mapping them precisely to geographic sites, e.g., a crossroad, an apartment building, or a grocery store. In this paper, we introduce a new dataset HarveyNER with fne-grained loca- tions annotated in tweets. This dataset presents unique challenges and characterizes many com- plex and long location mentions in informal descriptions. We built strong baseline models using Curriculum Learning and experimented with different heuristic curricula to better recog- nize diffcult location mentions. Experimental results show that the simple curricula can im- prove the system’s performance on hard cases and its overall performance, and outperform several other baseline systems. The dataset and the baseline models can be found at https: //github.com/brickee/HarveyNER. Introduction The Named Entity Recognition (NER) task aims to locate and classify textual phrases as entity men- tions that belong to predefned entity categories. Location is one of the general entity categories and has been annotated in many NER datasets, in- cluding CoNLL-2003 (Tjong Kim Sang, 2002) and OntoNotes 5.0 (Pradhan et al., 2013). However, these datasets contain mostly coarse-grained enti- ties such as a continent (e.g., Europe), a country (e.g., the U.S.), or a city (e.g., London). Many downstream applications require identify- ing fne-grained location entities from texts, such as an apartment building (e.g., Bayou Oaks ) or a specifc store (e.g., the HEB on Montrose), in order to locate the geographic places on a map, which is vital to identify actionable information Figure 1: An example of a disaster response system. (Khanal and Caragea, 2021). For example, in Fig- ure 1, a food disaster happened in the Houston area and then someone tweeted the shortage of ne- cessities in two locations. If a disaster response system can detect the disaster-related tweets, iden- tify the two location mentions from the text, and link them to location entities on the map, necessary help can be directly delivered to the people living in disaster-affected places. Accurately identifying the fne-grained location mentions plays a critical role in such a system. Several previous works have attempted to create crisis-related datasets with fne-grained location mentions, either automatically (Middleton et al., 2013) or by manual (Khanal et al., 2021) annota- tions. However, these prior datasets contain many incomplete descriptions of locations that other- wise can be precisely projected to a map with cer- tain geo-coordinates. For example, "the corner of Richey St and W Harris Ave in Pasadena" is an intersection of two roads and …

Content truncated to 500 words