Publication
Development of a recommendation system for scientific literature based on deep learning
| Summary: | The previous few decades have seen an enormous volume of articles from the scientific commu nity on the most diverse biomedical topics, making it extremely challenging for researchers to find relevant information. Methods like Machine Learning (ML) and Deep Learning (DL) have been used to create tools that can speed up this process. In that context, this work focuses on examining the performance of different ML and DL techniques when classifying biomedical documents, mainly regarding their relevance to given topics. To evaluate the different techniques, the dataset from the BioCreative VI Track 4 challenge was used. The objective of the challenge was to identify documents related to protein-protein interactions altered by mutations, a topic extremely important in precision medicine. Protein-protein interactions play a crucial role in the cellular mechanisms of all living organisms, and mutations in these interaction sites could be indicative of diseases. To handle the data to be used in training, some text processing methods were implemented in the Omnia package from OmniumAI, the host company of this work. Several preprocessing and feature extraction methods were implemented, such as removing stopwords and TF-IDF, which may be used in other case studies. They can be used either with generic text or biomedical text. These methods, in conjunction with ML pipelines already developed by the Omnia team, allowed the training of several traditional ML models. We were able to achieve a small improvement on performance, compared to the challenge baseline, when applying these traditional ML models on the same dataset. Regarding DL, testing with a CNN model, it was clear that the BioWordVec pre-trained embedding achieved the best performance of all pre-trained embeddings. Additionally, we explored the application of more complex DL models. These models achieved a better performance than the best challenge submission. BioLinkBERT managed an improvement of 0.4 percent points on precision, 4.9 percent points on recall, and 2.2 percent points on F1. |
|---|---|
| Main Authors: | Silva, Tiago Rafael Ferreira Miranda da |
| Subject: | Deep learning Document classification Machine learning Biomedical text mining Text mining Aprendizagem profunda Classificação de documentos Aprendizagem máquina Mineração de texto biomédico Mineração de texto |
| Year: | 2022 |
| Country: | Portugal |
| Document type: | master thesis |
| Access type: | open access |
| Associated institution: | Universidade do Minho |
| Language: | English |
| Origin: | RepositóriUM - Universidade do Minho |
| Summary: | The previous few decades have seen an enormous volume of articles from the scientific commu nity on the most diverse biomedical topics, making it extremely challenging for researchers to find relevant information. Methods like Machine Learning (ML) and Deep Learning (DL) have been used to create tools that can speed up this process. In that context, this work focuses on examining the performance of different ML and DL techniques when classifying biomedical documents, mainly regarding their relevance to given topics. To evaluate the different techniques, the dataset from the BioCreative VI Track 4 challenge was used. The objective of the challenge was to identify documents related to protein-protein interactions altered by mutations, a topic extremely important in precision medicine. Protein-protein interactions play a crucial role in the cellular mechanisms of all living organisms, and mutations in these interaction sites could be indicative of diseases. To handle the data to be used in training, some text processing methods were implemented in the Omnia package from OmniumAI, the host company of this work. Several preprocessing and feature extraction methods were implemented, such as removing stopwords and TF-IDF, which may be used in other case studies. They can be used either with generic text or biomedical text. These methods, in conjunction with ML pipelines already developed by the Omnia team, allowed the training of several traditional ML models. We were able to achieve a small improvement on performance, compared to the challenge baseline, when applying these traditional ML models on the same dataset. Regarding DL, testing with a CNN model, it was clear that the BioWordVec pre-trained embedding achieved the best performance of all pre-trained embeddings. Additionally, we explored the application of more complex DL models. These models achieved a better performance than the best challenge submission. BioLinkBERT managed an improvement of 0.4 percent points on precision, 4.9 percent points on recall, and 2.2 percent points on F1. |
|---|