Weighted Co-Occurrence Bio-Term Graph for Unsupervised Word Sense Disambiguation in the Biomedical Domain

Abstract

Word Sense Disambiguation (WSD) is a significant and challenging task for text understanding and processing. This paper presents an unsupervised approach based on Weighted Co-occurrence bio-Term Graph (WCOTG) for performing WSD in the biomedical domain. The graph is automatically created from biomedical terms that are extracted from a corpus of downloaded scientific abstracts. Two kinds of weights are introduced on the links of the built bio-term graph and are taken as important factors in the process of disambiguation. The modified Personalised PageRank (PPR) algorithm is used for performing WSD. When evaluated on the NLM-WSD and MSH-WSD test datasets, and an acronym test set, the method outperforms the widely used unsupervised ones addressing the same problem, and the average result is almost equal to that of the BlueBERT_LE-based method. In contrast, our method has no additional enhancement or training for BERT-based models. Comparative experiments validate the positive effect of links' weight on disambiguation efficiency. Last, the statistical experiments on the relation among system accuracy, the numbers of medical abstracts in the corpus, and the corresponding extracted terms suggest an excellent minimum corpus scale, when resources are limited.INDEX TERMS Biomedical informatics, biomedical natural language processing, word sense disambiguation, unified medical language system, personalised PageRank algorithm. I. INTRODUCTIONWord Sense Disambiguation (WSD) systems attempt to automatically identify the proper sense of ambiguous words in context [1], [2]. For example, WSD would aim to identify the meaning of the word ''cold'' to be ''cold temperature'' or ''common cold'' depending on the context in which it occurs. WSD is often characterized as an intermediary step in the process of understanding natural language texts [2], [3]. It is beneficial for applications in the biomedical domain, such as information extraction, automated knowledge discovery, question-answering [4], etc.The associate editor coordinating the review of this manuscript and approving it for publication was Giovanni Dimauro .The disambiguation task in the biomedical field is aimed at medical texts and terms. Combining domain knowledge can better improve the system's performance. Currently, the Unified Medical Language System (UMLS) Knowledge Sources are widely used in tasks such as disambiguation and automatic question-answering in the biomedical field. However, existing work such as [5] and [6], etc., typically involves extracting concepts (CUIs) from UMLS to build systems. In disambiguation tasks, this can lead to a mapping bottleneck when converting polysemy in biomedical documents into concepts. Therefore, we build term graphs directly, not concept graphs, the latter requires the use of term-toconcept mapping tools. This method can be considered an improvement of the existing work by Duque et al. [5], whose