A Simple Text-based Relevant Location Prediction Method using Knowledge Base

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 116–121 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 116 A Simple Text-based Relevant Location Prediction Method using Knowledge Base Mei Sasaki Shumpei Okura {mesasaki, sokura, shiono}@yahoo-corp.jp Yahoo Japan Corporation Tokyo, Japan Shingo Ono Abstract In this paper, we propose a simple method to predict salient locations from news article text using a knowledge base (KB). The proposed method uses a dictionary of locations created from the KB to identify occurrences of loca- tions in the text and uses the hierarchical in- formation between entities in the KB for as- signing appropriate saliency scores to regions. It allows prediction at arbitrary region units and has only a few hyperparameters that need to be tuned. We show using manually anno- tated news articles that the proposed method improves the f-measure by > 0.12 compared to multiple baselines. 1 Introduction Predicting relevant locations from news articles can result in numerous useful applications. For ex- ample, it enables the delivery of news related to a specific city that is of user interest, or facilitates the prediction of a disease outbreak in a specific region when used with event detection techniques. In this paper, we focus on predicting relevant locations from news articles. The goal of this task is to identify locations that are salient to the ar- ticle, not those that simply appeared in the article. For example, consider the following excerpt: “The Aoi Festival is one of the three major festivals in Kyoto. It originated as a series of rites to calm down angry gods. A visitor from Australia said...” In this example, the phrase “Kyoto” is highly rel- evant to the article, but “Australia” is not. Traditional methods to predict locations require specific data, such as a training dataset or phrase distribution, that match the application domain and the granularity of the prediction. However, it is costly to prepare such data for individual appli- cations. We propose a knowledge base (KB)-based method that only requires a general-purpose KB instead of a labeled dataset for training. It prop- agates phrase-level importance to region entities following their relationship in the KB. It can theo- retically be applied to predictions at an arbitrary level of granularity (e.g. countries, prefectures, cities) without dedicated training data. In this study, we focus on Japanese news articles and report the performance of predictions at the Japanese prefecture-level. We provide practical tips to tackle un-tokenized language like Japanese. 2 Related Works Depending on the objective, geolocation predic- tion tasks from texts are roughly divided into two types. One type is for detecting and identify- ing mentions of points-of-interest (POIs) in the text, well known by entity linking (Nadeau and Sekine, 2007; Shen et al., 2015). This task fo- cuses on extracting all mentions regardless of their saliency. The other type is for estimating the au- thor’s current location or home town …

Content truncated to 500 words