Learning cascaded latent variable models for biomedical text classification

Learning cascaded latent variable models for biomedical text classification Ming Liu Gholamreza Haffari Faculty of Information Technology, Monash University ming.m.liu, gholamreza.haffari, wray.buntine @ monash.edu Wray Buntine Abstract In this paper, we develop a weakly su- pervised version of logistic regression to help to improve biomedical text classi- fication performance when there is lim- ited annotated data. We learn cascaded latent variable models for the classifica- tion tasks. First, with a large number of unlabelled but limited amount of labelled biomedical text, we will bootstrap and semi-automate the annotation task with partially and weakly annotated data. Sec- ond, both coarse-grained (document) and fine-grained (sentence) levels of each in- dividual biomedical report will be taken into consideration. Our experimental work shows this achieves higher classification results. 1 Introduction In recent years, large amounts of biomedical text have become available with the development of electronic medical record (EMR) systems. The type of biomedical text ranges from reports of CT scans to doctoral notes and discharge summaries. Based on these biomedical text, there are medi- cal tasks such as disease identification, diagnostic surveillance and evaluation and other clinical sup- port services. Manual extraction and classification for these medical tasks from biomedical text is a time-consuming and often costly effort. Biomedical text classification systems which consider both manual effort (e.g. annotation) and predictive performance are more appropriate in the medical context than those which only con- sider classification predictive performance. Early biomedical classification methods are rule-based (Tinoco et al., 2011; Matheny et al., 2012), which requires medical experts to develop logical rules to identify reports consistent with some diseases. The main advantages of such rule-based systems is that high precision can be achieved, but the weak- ness lies in the fact that the process is not easily transferable to similar tasks, because medical ex- perts have to carefully develop specific types of rules and formulas for different kinds of diseases. In recent years, machine learning methods have been widely used in disease identification from biomedical text(Ehrentraut et al., 2012; Bejan et al., 2012; Martinez et al., 2015; Hassanpour and Langlotz, 2015), which also ask medical experts to do some annotation work for building training data. Unlabelled free biomedical text in hospitals and other clinical organizations is abundant but manual annotation is very expensive. Exploiting fine-grained sentence level proper- ties for coarse-grained document level classifi- cation has attracted large amounts of attention. Pang(Pang and Lee, 2004) first explored subjec- tivity extraction methods based on a minimum cut formulation, in which they performed subjectiv- ity detection on individual sentences and imple- mented document level polarity classification by leveraging those extracted subjective sentences. McDonald(T¨ackstr¨om and McDonald, 2011) pro- posed a structured model for jointly classifying the sentiment of text at varying levels of granu- larity, they showed that this task can be reduced to sequential classification with constrained in- ference. Yessenalina(Yessenalina et al., 2010) described a joint two-level approach for docu- ment level sentiment classification that simulta- neously extracts useful sentences, and Fang(Fang and Huang, 2012) extended it …

Content truncated to 500 words