Capsule Network Meets Recurrent Unit to Decode Ambiguous Biomedical Words

A study published in Knowledge and Information Systems on 26 October 2026 proposes a computational method for biomedical word sense disambiguation (WSD) built on a capsule gated recurrent unit. The work, carried out by Chunxiang Zhang, Bingqi Xiang and Xueyao Gao, starts from the observation that words in biomedical texts usually carry ambiguity, so resolving their correct meaning matters for both medical research and clinical practice. The core idea is to pair the semantic compression ability of a capsule neural network with the context modeling ability of a gated recurrent unit (GRU).

The authors describe these two neural architectures as having complementary strengths: one organizes the features that distinguish senses, while the other tracks the linguistic setting in which an ambiguous word appears. The GRU is a recurrent architecture designed to carry information across a sequence. According to the paper, it is used to enhance the network's capacity to understand and remember the surrounding text in which an ambiguous word occurs, which is precisely the information needed to decide which sense applies.

The capsule network, in turn, is adopted to extract spatial hierarchical structure from the encoded representation. The researchers argue that this property helps the model organize the features that separate one meaning of a word from another, supporting the disambiguation task. Before these components go to work, the method relies on BERT to vectorize the input features.

The disambiguation features include the word itself, its part of speech, and its candidate meanings, so each element is converted into dense numerical representations that the downstream layers can process. At the end of the pipeline, a linear transformation computes the probability distribution of the ambiguous biomedical word across each semantic category in the classification layer. The semantic category with the highest probability is then selected as the correct sense of the word.

The team optimized and tested the proposed network on the MSH dataset. According to the paper's abstract, experimental results show that the proposed method can significantly improve the average accuracy of biomedical WSD, although the publicly available preview does not report the specific accuracy figures or the baselines against which the improvement was measured. Readers should note that the claims summarized here come from the article's abstract and preview; the full peer-reviewed text, including detailed experimental comparisons, is available behind the journal's subscription access.

Subject of Research: Biomedical word sense disambiguation based on capsule gated recurrent unit

Article Title: Biomedical word sense disambiguation based on capsule gated recurrent unit

Article References: Zhang, C., Xiang, B., & Gao, X. (2026). Biomedical word sense disambiguation based on capsule gated recurrent unit. Knowledge and Information Systems, 68(1), Article 298. https://doi.org/10.1007/s10115-026-02920-0

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02920-0

Keywords: Biomedical, word, sense, disambiguation, based, capsule, gated, recurrent, unit, scientific research