Status : Verified
| Personal Name | Rey, Charles Arthel R. |
|---|---|
| Resource Title | Transformer-based approach to variable typing |
| Date Issued | 06 June 2022 |
| Abstract | The surge of academic and industrial pursuits across all scientific disciplines propelled the phenomenon of Big Data in the scientific domain. This means that data consumers are working on massively overwhelming volumes that require time and resources to probe. Hence, this necessitates an efficient knowledge management system that can structure scattered data so that knowledge can still be drawn from heterogeneous information inherent in various branches of science. A unifying lens of scoping mathematical knowledge through variable typing across domains can be a useful kick-off to managing scientific data. In this work, a first attempt to finetune the Bidirectional Encoder for Representations from Transformers (BERT) model was implemented to conduct an end-to-end Entity Recognition (ER) and Relation Extraction (RE) approach to variable typing. For this task, a new micro-dataset was introduced and used for the fine-tuning process. Our overall results show a Precision of 0.8142 and 0.4919, Recall of 0.7816 and 0.6030, and F1-Score of 0.7975 and 0.5418 for the ER model and RE model, respectively; the model performance shows that notwithstanding the limited number of positive instances, the ER component performed at par with similar studies in literature, inferring that it is effective in identifying nodes in a mathematical document. The results also affirmed that the model is a good circumvention to extend or refine current variable typing works in literature. This work also elucidated the most telling factors affecting both components, which gave rise to suggested interventions in future implementations. |
| Degree Course | MS Chemical Engineering |
| Language | English |
| Keyword | entity recognition; knowledge management; mathematical knowledge management; natural language processing; relation extraction; variable typing |
| Material Type | Thesis/Dissertation |
Preliminary Pages
139.45 Kb