| Title | : | Text Mining Approaches for Brain-Connectivity Extraction |
| Speaker | : | Ashika (IITM) |
| Details | : | Fri, 30 Jun, 2023 3:00 PM @ Google Meet |
| Abstract: | : | Understanding the complex connectivity structure of the brain is a significant challenge in neuroscience. Studies on brain connectivity provide important bio-markers for analysing the neurological functions of the brain, which in turn helps to predict brain related disorders. Neuronal connectivity is widely studied by researchers, and literature in the form of published research articles and databases about brain connectivity is vast and ever-expanding. Manually curating and continuously updating such a resource involves significant time and effort. The central objective of this thesis is to mine ‘Brain-connectivity’ information from neuroscience research articles. This aids in building a centralized resource for Brain-Connectivity that neuroscience researchers can use to gain quick access to research findings in published articles and repositories. We propose a CBR framework for Brain-Connectivity extraction using supervised learning algorithms that perform shallow and deep linguistic analysis of the text. Issues pertaining to the representation of, and retrieval over, textual cases are explored. Proposed algorithms are evaluated using benchmark datasets collated from abstracts on PubMed and a newly created dataset of full-text articles annotated by a domain expert. An effort towards automatically generating interpretable patterns of connectivity for extracting connected brain regions from the text is presented. The proposed techniques range from simple word-based features to deep syntactic representation using supervised learning. The proposed algorithms achieve better recall and F2 score compared to the previous state-of-the-art while eliminating the need for any manually defined, domain-specific, linguistic patterns. We describe a mechanism to induce human interpretable patterns from the corpus that can be inspected and potentially improved based on expert feedback. The interpretable syntax-based method provides the ability to explain results generated by the algorithm. We present ConnExt-BioBERT (http://www.braincircuits.org/text-mining), a system for Brain-Connectivity extraction from a large corpus of 53000 full-text neuroscience articles. The system fine-tunes BioBERT, the biomedical variant of the neural method Bidirectional Encoder Representations for Transformers (BERT) to extract brain regions and connections from the repository of full-text articles. We have designed web applications for search over brain region connections extracted by the proposed methods from the repository of 53000 articles. These applications are currently being used by neuroscience researchers to mine brain connectivity information reported by various authors. In order to enrich the manually curated, centralized resources for Brain-Connectivity, like the Brain Architecture Management System (BAMS), an attempt is made to normalize the extracted brain regions across nomenclatures and name variations between ConnExt and BAMS. ConnExt-BioBERT suggests potential augmentation to enrich BAMS using the large-scale text mining of Brain-Connectivity from research articles. Additionally, a bootstrapping approach for Brain-Connectivity extraction using Semi-Supervised Learning (SSL) is explored. This is applicable in scenarios where supervision is limited, as SSL overcomes the need for manual labelling of a large dataset. The algorithm is bootstrapped using a few hand-labelled connected brain region pairs in the form of seeds. Empirical evaluation based on varying quantities and the nature of seeds is explored. We observe that the SSL methods achieve a performance comparable with the supervised learning methods while providing a major saving in limited manual labelling of data required for training. SSL also proves to be an effective technique with negligible performance reduction in supervision-deprived domains like neuroscience, where labelled data is scanty or unavailable. Web Conference Link :https://meet.google.com/fte-dasq-vsr |
