Predicting Informativeness of Semantic Triples

Proceedings of Recent Advances in Natural Language Processing, pages 1124–1129 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_126 1124 Predicting Informativeness Of Semantic Triples Judita Preiss University of Salford Salford M5 4WT, United Kingdom j.preiss@salford.ac.uk Abstract Many automatic semantic relation extraction tools extract subject-predicate-object triples from unstructured text. However, a large quan- tity of these triples merely represent back- ground knowledge. We explore using full texts of biomedical publications to create a training corpus of informative and important semantic triples based on the notion that the main con- tributions of an article are summarized in its abstract. This corpus is used to train a deep learning classifier to identify important triples, and we suggest that an importance ranking for semantic triples could also be generated. 1 Introduction Subject-predicate-object triples are used in numer- ous natural language processing areas, including question answering (e.g. Hristovski et al. (2015)), ontology building (e.g. Du and Li (2020)) and liter- ature based discovery (e.g. Hristovski et al. (2006)). While they can be thought of as representing the minimum unit of semantic expression, there is a large degree of variability in the amount of new (not commonly known) content they convey. On the one hand, they sometimes represent what can be termed background knowledge, for example “New Zealand - ISA - country” or “pharmaceuti- cal services - TREATS - health personnel”, while on the other, they may describe very specific find- ings such as pimobendan TREATS hypertrophic cardiomyopathy or LCN2 protein, human - ASSO- CIATED WITH - chronic kidney disease. We use biomedical publications to test the hypothesis that training data consisting of such, important, triples can be created from abstracts, and train a deep learning algorithm to identify these high impor- tance triples from a list of all triples appearing in a paper. The system could also be adjusted to output a weight instead of a binary decision, allowing for an importance ranking of semantic triples within an article. The paper begins with an overview of related work in Section 2, the experimental set-up follows in Section 3, with the results and discussion in Section 4 and conclusions drawn in Section 5. 2 Background A number of tools for automatically extracting se- mantic relations – (subject, relation, object) triples – from unstructured text exist (Yuan and Yu, 2018). However, as Papadopoulos et al. (2020) point out, the majority of works incorporating these do not perform much pre- or post- processing and there- fore include many potentially uninformative triples, and works proposing to extend currently existing collections of semantic relations often speak of extending the set of relations, not refining the rela- tions present (e.g. Koroleva et al. (2020)). Evaluations of semantic relation extraction sys- tems are often very comprehensive, e.g. Kilicoglu et al. (2020) present a detailed independent evalua- tion of SemRep – a biomedical domain tuned triple extraction tool – and discover common sources of error for this tool, but such evaluations do not quantify the quality of the triple that is retrieved by the system. It is …

Content truncated to 500 words