Classification of Study Region in Environmental Science Abstracts

Classification of Study Region in Environmental Science Abstracts Jared Willett,♠ Timothy Baldwin,♠♥ David Martinez♥ and 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 One of the potentially most relevant pieces of metadata for filtering studies in environ- mental science is the geographic region in which the study took place (the “study re- gion”). In this paper, we apply support vector machines to the automatic classifi- cation of study region in a dataset of ti- tles and abstracts from environmental sci- ence literature, using features including fre- quency distributions of resolved toponyms and a bag of word unigrams. We found that we can determine the study region with high accuracy, with the strongest classifier achieving an accuracy of 0.892 combin- ing toponym resolution from DBpedia and GeoNames with the bag-of-toponyms fea- tures. 1 Introduction One of the potentially most relevant pieces of metadata for filtering studies in environmental science is the region in which the study took place, as users making queries are often looking for studies performed in a specific area. However, bibliographic databases do not systematically in- clude information on study location. The Eco Ev- idence database, a compendium of literature cita- tions and linked evidence items that is used for ev- idence synthesis in companion software (Webb et al., 2011), is one such system for which location information is very helpful. However, the manual annotation of such metadata over large quantities of literature is a tedious and time-consuming task. One possible solution to this issue is to have this information automatically extracted with the aid of natural language processing (NLP) tech- niques. The abstracts of studies, which are com- monly available in bibliographic databases, fre- quently contain geographic references of various granularities. If identified and resolved, these to- ponyms provide the potential to make a strong es- timation of the overall location of the study. In these experiments, we evaluate the performance of various NLP techniques for automatic classi- fication of the study region in environmental sci- ence literature abstracts. Beyond the aim of being able to quickly as- semble a collection of literature from a given area, our motivation in applying NLP to automati- cally extract information from environmental sci- ence literature is driven by our interest in mov- ing towards an evidence-based model of decision- making in the environmental sciences (Sutherland et al., 2004). Similar to evidence-based medicine (Sackett et al., 1996), such a model relies heav- ily on systematic literature reviews as a means of synthesizing evidence from the literature. The Eco Evidence database (Webb et al., in press) is a compendium of literature citations and linked evi- dence items that is used for systematic review and evidence synthesis in companion software (Webb et al., 2011). The database is in active use in a number of research projects currently, and evi- dence therein has also formed the basis of several …

Content truncated to 500 words