Nominalization and Alternations in Biomedical Language

Kevin Bretonnel Cohen, Martha Palmer, Lawrence HunterPLoS 2008 · DOI

Abstract

We found that even in this semantically restricted domain, alternations are quite common, and alternations involving nominalizations are exceptionally diverse. Nonetheless, the sublanguage model applies to biomedical language. We also report on a previously undescribed alternation involving an adjectival present participle.

Introduction

In the field of Natural Language Processing (NLP), biomedical text is often categorized as a "sublanguage"—a domain with specialized vocabulary but an assumedly simpler or more rigid syntax than general English. However, Information Extraction (IE) systems frequently struggle with the diverse ways biological events are articulated. For instance, the semantic equivalence between "FSH stimulates follicular development" and "follicular development is stimulated by FSH" is clear to a human reader, but may be opaque to a system not designed to handle such syntactic alternations. This complexity is compounded by nominalizations, where actions are expressed as nouns (e.g., "the stimulation of follicular development"). This study investigates the frequency and nature of these syntactic variations within biomedical literature to improve the robustness of text-mining systems.

Methodology

The researchers conducted a deep linguistic analysis of the PennBioIE corpus, specifically focusing on sections related to CYP450 enzymes and Oncology. The GENIA corpus was utilized as a secondary dataset to compare the frequency of specific verb lemmas. The study targeted the ten most frequent domain-specific verbs, such as inhibit, induce, and activate.

The methodology consisted of:

  1. Verb Annotation: 1,872 verb tokens were labeled for alternations including passive vs. active, transitive vs. intransitive, and verbal vs. adjectival forms.
  2. Nominalization Annotation: 746 noun tokens derived from these verbs (e.g., inhibition) were annotated to track the realization of arguments (the actors and the acted-upon).
  3. Linguistic Framework: The analysis utilized Levin classes to compare biomedical verb behaviors against general English standards.
  4. Inter-Annotator Agreement (IAA): Consistency was measured using the Kappa statistic and F-measure to ensure high data quality.

Results

The analysis revealed that biomedical language is syntactically "exceptionally diverse" despite its semantic restrictions. Key findings include:

  • Prevalence of Passives: 29.3% of verb tokens were passive, a rate significantly higher than that found in news (15%) or conversational English (2%).
  • Adjectival Forms: 18.2% of non-noun tokens were adjectival forms of verbs, a category frequently overlooked by standard IE systems.
  • Nominalization Diversity: Nominalizations exhibited extreme variation; for the noun activation, 14 out of 16 possible argument realization patterns were identified within just 91 tokens.
  • Annotation Consistency: While annotators showed high agreement (95.8%) on the physical location of arguments (pre- or post-nominal), identifying arguments external to the immediate noun phrase (NP-external) proved more challenging.

Discussion

A critical insight from this study is the role of nominalizations as "argument-dropping" mechanisms. In these constructions, the "agent" (the entity performing the action) is absent 66% of the time. This absence significantly complicates the task of automated systems attempting to determine the causal factors of biological processes compared to when a standard verb is used.

Furthermore, the study identifies a previously undescribed alternation involving the adjectival present participle (e.g., "increasing concentrations of BH4"). This construction appears to parallel the well-known adjectival passive, suggesting that the relationship between underlying semantics and surface syntax in this domain is more fluid than previously theorized. The findings support the sublanguage model but emphasize that "restricted" does not equate to "simple."

Conclusion

This research highlights the necessity for IE systems to account for the high frequency of alternations and the complex argument structures of nominalizations in biomedical prose. While the study was limited to the ten most frequent verbs and did not include agentive nouns (e.g., inhibitor), it provides a foundational roadmap for developing more sophisticated, rule-based IE systems. Future work should aim to expand this analysis to a broader set of verbs and explore the point of "closure" where no new syntactic patterns emerge, ultimately creating a comprehensive lexicon of biomedical verb semantics.

Summary updated

Abstract

We found that even in this semantically restricted domain, alternations are quite common, and alternations involving nominalizations are exceptionally diverse. Nonetheless, the sublanguage model applies to biomedical language. We also report on a previously undescribed alternation involving an adjectival present participle.
BibTeX
@inproceedings{cohen2008nominalization,
  author = {Kevin Bretonnel Cohen and Martha Palmer and Lawrence Hunter},
  title = {{Nominalization and Alternations in Biomedical Language}},
  booktitle = {PLoS},
  year = {2008},
  doi = {10.1371/journal.pone.0003158},
}
Figure