Browsing by Subject "text mining"
Now showing items 1-4 of 4
-
(Helsingin yliopisto, 2020)Maps of science, or cartography of scientific fields, provide insights into the state of scientific knowledge. Analogous to geographical maps, maps of science present the fields as positions and show the paths connecting each other, which can serve as an intuitive illustration for the history of science or a hint to spot potential opportunities for collaboration. In this work, I investigate the reproducibility of a method to generate such maps. The idea of the method is to derive representations representations for the given scientific fields with topic models and then perform hierarchical clustering on these, which in the end yields a tree of scientific fields as the map. The result is found unreproducible, as my result obtained on the arXiv data set (~130k articles from arXiv Computer Science) shows an inconsistent structure from the one in the reference study. To investigate the cause of the inconsistency, I derive a second set of maps using the same method and an adjusted data set, which is constructed by re-sampling the arXiv data set to a more balanced distribution. The findings show the confounding factors in the data cannot account for the inconsistency; instead, it should be due to the stochastic nature of the unsupervised algorithm. I also improve the approach by using ensemble topic models to derive representations. It is found the method to derive maps of science can be reproducible when it uses an ensemble topic model fused from a sufficient number of base models.
-
(Helsinki University Press, 2020)Historical scholarship is currently undergoing a digital turn. All historians have experienced this change in one way or another, by writing on word processors, applying quantitative methods on digitalized source materials, or using internet resources and digital tools. Digital Histories showcases this emerging wave of digital history research. It presents work by historians who – on their own or through collaborations with e.g. information technology specialists – have uncovered new, empirical historical knowledge through digital and computational methods. The topics of the volume range from the medieval period to the present day, including various parts of Europe. The chapters apply an exemplary array of methods, such as digital metadata analysis, machine learning, network analysis, topic modelling, named entity recognition, collocation analysis, critical search, and text and data mining. The volume argues that digital history is entering a mature phase, digital history ‘in action’, where its focus is shifting from the building of resources towards the making of new historical knowledge. This also involves novel challenges that digital methods pose to historical research, including awareness of the pitfalls and limitations of the digital tools and the necessity of new forms of digital source criticisms. Through its combination of empirical, conceptual and contextual studies, Digital Histories is a timely and pioneering contribution taking stock of how digital research currently advances historical scholarship.
-
(Helsingin yliopisto, 2020)Text mining methods provide a solution to the task of extracting relevant information from large text datasets. These methods can be applied to extract the relevant parts of Suomi24 internet health discussion to analyze how people discuss and negotiate their health through words, which represents medication or symptoms. Semantic similarities between these two concepts can be examined by learning the word vector representations from data and exploring the vector space using Word2Vec, a popular word embedding method. This thesis reviews how the training of word similarity models is affected by increasing corpus size using text retrieval methods.The effects of corpus size are examined by comparing the measured cosine similarity distances between word vectors representations in two different vector spaces. Word vector representations are learned using two different sized corpora. The first corpus includes only messages from the health discussion area of Suomi24. The second corpus includes the same messages as the first corpus, but also messages from other discussion areas, which include health related words. Cosine similarities are evaluated on using concept vocabularies including relevant health related words. Increasing the number of training examples by almost 30% did not have a drastic effect on the qualities of the training data. The results did not indicate a distinct connection between corpus size and the measured cosine similarity distances between word vector representations of health related words.
-
(2021)In this research, we present our work participation for the DrugProt task of BioCreative VII challenge. Drug-target interactions (DTIs) are critical for drug discovery and repurposing, which are often manually extracted from the experimental articles. There are >32M biomedical articles on PubMed and manually extracting DTIs from such a huge knowledge base is challenging. To solve this issue, we provide a solution for Track 1, which aims to extract 10 types of interactions between drug and protein entities. We applied an Ensemble Classifier model that combines BioMed-RoBERTa, a state of art language model, with Convolutional Neural Networks (CNN) to extract these relations. Despite the class imbalances in the BioCreative VII DrugProt test corpus, our model achieves a good performance compared to the average of other submissions in the challenge, with the micro F1 score of 55.67% (and 63% on BioCreative VI ChemProt test corpus). The results show the potential of deep learning in extracting various types of DTIs.
