This chapter discusses two different types of text data mining focusing on the biomedical literature. One deals with explicit information or facts written in articles, and the other targets implicit information or hypotheses inferred from explicit information. A major difference between the two is that the former is bound to the contents within the literature, whereas the latter goes beyond existing knowledge and generates potential scientific hypotheses. As concrete examples applied to real-world problems, this chapter looks at two applications of text data mining: gene functional annotation and genetic association discovery, both considered to have significant practical importance.
TopBackground
There have been numerous efforts in biomedical TDM dealing with explicit information (Ananiadou et al., 2006; Cohen and Hersh, 2005; Shatkay, 2005). One of the earliest and most successful attempts in this type of TDM is named-entity (NE) recognition, the first step to IE, mainly targeting genes and proteins (Fukuda et al., 1998; Seki and Mostafa, 2005b; Hsu et al., 2008). NE recognition in biomedicine is largely different from other domains tackled earlier, such as newspaper articles, in a sense that biomedical NEs have surprisingly many synonyms and writing variants. This issue is essential in dealing with biomedical text and heavily affects the performance of TDM systems as we will see in the next section.
For biomedical IR, the Genomics Track (Hersh and Bhuptiraju, 2003; Hersh et al., 2004, 2005, 2006, 2007) at the Text REtrieval Conference (TREC) was undoubtedly the most significant strides made in the history. The track was a five-year project held between 2003 and 2007 and tackled various types of IR tasks, including Ad Hoc retrieval and passage retrieval, as well as other IE oriented tasks. While the track was successful, having attracted the largest number of research groups world-wide among the TREC tracks, there is still much room for improvement, especially for the passage retrieval challenged in 2007 which, given a user query, required to return passages containing relevant named entities of a certain type within the context of supporting text.
Another task from the Genomics Track that is pertinent to TDM (in broader sense) is Gene Ontology (GO) annotation via automatic text analysis. The following provides some background of the task since this will be one of the main focuses in this chapter.
After the completion of the Human Genome Project, the major activities in molecular biology have shifted to understanding the precise functions of individual genes. The consequence in part is the increasing, large number of publications that one cannot digest. To provide direct access to the information regarding gene functions buried in natural language text, three model organism databases created controlled vocabularies, namely, the Gene Ontology (GO), to annotate genes with their functions. GO is structured as directed acyclic graph (DAG) under three top level nodes, Molecular Function (MF), Cellular Component (CC), and Biological Process (BP), as shown in Figure 1.