Experiments with Clustering-based Features for Sentence Classification in Medical Publications: Macquarie Test’s participation in the ALTA 2012 shared task.

Experiments with Clustering-based Features for Sentence Classification in Medical Publications: Macquarie Test’s participation in the ALTA 2012 shared task. Diego Moll´a Department of Computing Macquarie University Sydney, NSW 2109, Australia diego.molla-aliod@mq.edu.au Abstract In our contribution to the ALTA 2012 shared task we experimented with the use of cluster-based features for sentence clas- sification. In a first stage we cluster the documents according to the distribution of sentence labels. We then use this informa- tion as a feature in standard classifiers. We observed that the cluster-based feature im- proved the results for Naive-Bayes classi- fiers but not for better-informed classifiers such as MaxEnt or Logistic Regression. 1 Introduction In this paper we describe the experiments that led to our participation to the ALTA 2012 shared task. The ALTA shared tasks1 are programming com- petitions where all participants attempt to solve a problem based on the same data. The partic- ipants are given annotated sample data that can be used to develop their systems, and unannotated test data that is used to submit the results of their runs. There are no constraints about what tech- niques of information are used to produce the fi- nal results, other than that the process should be fully automatic. The 2012 task was about classifying sentences of medical publications according to the PIBOSO taxonomy. PIBOSO (Kim et al., 2011) is an alter- native to PICO for the specification of the main types of information useful for evidence-based medicine. The taxonomy specifies the follow- ing types: Population, Intervention, Background, Outcome, Study design, and Other. The dataset was provided by NICTA2 and consisted of 1,000 1http://alta.asn.au/events/sharedtask2012/ 2http://www.nicta.com.au/ medical abstracts extracted from PubMed split into an annotated training set of 800 abstracts and an unannotated test set of 200 abstracts. The com- petition was hosted by “Kaggle in Class”3. Each sentence of each abstract can have mul- tiple labels, one per sentence type. The “other” label is special in that it applies only to sentences that cannot be categorised into any of the other categories. The “other” label is therefore disjoint from the other labels. Every sentence has at least one label. 2 Approach The task can be approached as a multi-label se- quence classification problem. As a sequence classification problem, one can attempt to train a sequence classifier such as Conditional Random Fields (CRF), as was done by Kim et al. (2011). As a multi-label classification problem, one can attempt to train multiple binary classifiers, one per target label. We followed the latter approach. It has been observed that the abstracts of differ- ent publication types present different character- istics that can be exploited. This lead Sarker and Moll´a (2010) to the implementation simple but ef- fective rule-based classifiers that determine some of the key publication types for evidence based medicine. In our contribution to the ALTA shared task, we want to use information about different publication types to determine the actual sentence labels of the abstract. To recover the publication types one can at- tempt to use …

Content truncated to 500 words