A Multi-Strategy Approach for Location Mining in Tweets: AUT NLP Group Entry for ALTA-2014 Shared Task

A Multi-Strategy Approach for Location Mining in Tweets: AUT NLP Group Entry for ALTA-2014 Shared Task Parma Nand, Rivindu Perera, Anju Sreekumar Natural Language Processing Group Auckland University of Technology Auckland 1010, New Zealand {pnand,rperera,anjus}@aut.ac.nz He Lingmin Department of Computer Science China Jiliang University Hangzhou, China helm@cjlu.edu.cn Abstract This paper describes the strategy and the results of a location mining system used for the ALTA-2014 shared task competi- tion. The task required the participants to identify the location mentions in 1003 Twitter test messages given a separate an- notated training set of 2000 messages. We present an architecture that uses a ba- sic named entity recognizer in conjunc- tion with various rule-based modules and knowledge infusion to achieve an average F score of 0.747 which won the second place in the competition. We used the pre-trained Stanford NER which gives us an F score of 0.532 and used an ensem- ble of other techniques to reach the 0.747 value. The other major source of loca- tion resolver was the DBpedia location list which was used to identify a large percent- age of locations with an individual F-score of 0.935. 1 Introduction The objective of the ALTA competition was to identify a wide range of location mentions in Twit- ter messages. Among the various types of in- formation that can be mined from Twitter mes- sages, location mining has attracted a lot attention because of its application in identifying the geo- graphical location of a topic in the Tweet, whether it be an accident, a natural disaster, a concert or an open invitation party gone out of hand. Loca- tion mining is a special case of a generic NLP task called Named Entity Recognition (NER), which involves identifying and classifying entities into person, organization and location type entities. This particular competition focussed on only iden- tifying location type entities, however had a wider scope compared to a typical NER task. The ob- jective of the task was to identify the string which specifies the most specific possible locations in the message. A typical NER tool such as the Stanford1 and OpenNLP2 only identifies word based loca- tions which has to be then composed into noun phrases consisting of multiple words identifying the full location. In addition to the locations iden- tified by noun phrases, the task required the identi- fication of the specific locations defined by propo- sitional attachments such as “40km south of ten- terfield2” and “100 mi from nations capital”. Our approach to solving the location mining task at hand was to use a NER system to attain a benchmark performance and then to use various techniques to fine tune the system to account for the noise in Twitter messages. The architecture forms part of a bigger project called the Twitter Miner3. This project is a higher level research project currently underway at Auckland Univer- sity of Technology which is meant to extract infor- mation for various purposes from microblogging type texts such as Twitter. For the NER, we …

Content truncated to 500 words