BioProp Version 1.0

Introduction

BioProp Version 1.0 was developed by researchers at Academia Sinica, Taipei, Taiwan. It consists of proposition bank-style annotations for approximately 500 English biomedical journal abstracts. The source abstracts, annotated in accordance with Penn Treebank II guidelines, are contained in the GENIA Treebank (GTB). The GTB was developed at the Tsujii Laboratory at the University of Tokyo.

The purpose of the GENIA Project is to develop tools and resources for automatic information extraction of biomedical information. One result of that work is the GENIA corpus, a collection of 2000 biomedical journal abstracts containing semantic class annotation for biomedical terms, part-of-speech (POS) tags and coreferences. The GTB is a subset of that corpuse. BioProp Version 1.0 adds a proposition bank to the GTB.

Proposition Bank (PropBank) contains annotations of predicate argument structures and semantic roles in a treebank schema in the newswire domain. To construct BioProp Version 1.0, a semantic role labeling (SRL) system trained on PropBank was used to annotate the GTB. SRL, also called shallow semantic parsing, is a popular semantic analysis technique. In SRL, sentences are represented by one or more predicate-argument structures (PAS), also known as propositions. Each PAS is composed of a predicate (e.g., a verb) and several arguments (e.g., noun phrases) that have different semantic roles, including main arguments such as agent and patient, and adjunct arguments, such as time, manner and location. The term "argument" refers to a syntactic constituent of the sentence related to the predicate, and the term "semantic role" refers to the semantic relationship between a sentence's predicate and argument.

To suit the needs in the biomedical domain, the PropBank annotation guidelines were modified to characterize semantic roles as components of biological events. Specifically, thirty verbs were selected according to their frequency of use or importance in biomedical texts. Since targets in information extraction are relations of named entities, only sentences containing protein or gene names were used to count each verb's frequency. Verbs of general usage were filtered out in order to keep the focus on biomedical verbs. Some verbs that do not have a high frequency but play important roles in describing biomedical relations, such as "phosphorylate" and "transactivate," were also selected. The BioProp annotation was based on Levin?s verb classes as defined in the VerbNet lexicon. In VerbNet, the arguments of each verb are represented at the semantic level, and thus have associated semantic roles. However, since some verbs may have different usages in biomedical and newswire texts, it is necessary to customize the framesets of biomedical verbs. After selecting the predicate verbs, a semi-automatic method was used to annotate BioProp. The annotation process consisted of the following steps:

• Identification of predicate candidates
• Automatic annotation of the biomedical semantic roles using newswire SRL system
• Transformation of automatic tagging results into WordFreak format
• Review by human annotators

Data

BioProp Version 1.0 consists of approximately 150,000 words. Each line in the corpus provides a PAS annotation that can be mapped to a sentence in the GTB.

Samples