Automatic Climate Classification of Environmental Science Literature

Automatic Climate Classification of Environmental Science Literature Jared Willett,♠♥ Timothy Baldwin,♠♥ David Martinez♥ and J. Angus Webb♦ ♠ Department of Computing and Information Systems ♥ NICTA Victoria Research Laboratory ♦ Department of Resource Management and Geography The University of Melbourne, VIC 3010, Australia jwillett@student.unimelb.edu.au, tb@ldwin.net, davidm@csse.unimelb.edu.au, angus.webb@unimelb.edu.au Abstract Climate type is one of the potentially most relevant pieces of metadata for identify- ing studies in evidence-based environmen- tal management. In this paper, we pro- pose a method for automatically predict- ing the climate type in environmental sci- ence literature using NLP techniques, rel- ative to a pre-existing set of climate type categories. Our main approaches com- bine toponym detection and resolution us- ing two different resources with support vector machines. The results show great promise, but also further challenges, for using NLP to extract information from the vast and rapidly growing collection of en- vironmental sciences literature. 1 Introduction In this paper, we investigate the task of automatic prediction of climate type (e.g. temperate or arid) in environmental science abstracts. The climate type of an environmental science study is crucial information, which gives context to the research and insight into its wider implications and applica- bility. Availability of climate information as meta- data has clear value to researchers performing a systematic review of the literature or comprehen- sive analysis of the evidence. However, the man- ual annotation of climate type over a large volume of literature is a time-consuming task. In this pa- per, we seek to automate the climate annotation process with natural language processing (NLP) techniques. The task of climate type classification is complex as although the label set is relatively small, the geographic granularity is fine and to- ponym ambiguity becomes a significant problem — toponyms commonly mentioned in the envi- ronmental sciences (e.g. Murray River) are often large and cover multiple climates, which presents difficulties for a point-based representation of to- ponyms. Initially, experiments are run to exam- ine the effectiveness of the direct application of the classifiers developed by Willett et al. (2012) for study region classification. We then investigate methods for adapting these techniques to the cli- mate task through the modification of the toponym resolution component of our classifiers. These ap- proaches include utilizing a K¨oppen-Geiger cli- mate classification world map to resolve toponyms to climate instead of region, in addition to exper- iments with targeting types of toponyms reliable for identifying climate. 2 Related Work The methodology used to extract and disam- biguate toponyms is based on a standard approach to geographic information retrieval, which was presented, e.g., by Stokes et al. (2008) in their study on the performance of individual compo- nents of a geographic IR system. In particu- lar, the named entity recognition and classifica- tion (NERC) and toponym resolution (TR) com- ponents are the basis for the main classifiers in this study. The unique opportunities and challenges spe- cific to retrieving geospatial information have been well documented, particularly in the context of geospatial information retrieval where queries and documents …

Content truncated to 500 words