English WordNet Random Walk Pseudo-Corpora

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4893–4902 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4893 English WordNet Taxonomic Random Walk Pseudo-Corpora Filip Klubiˇcka1,3, Alfredo Maldonado2,3, Abhijit Mahalunkar1, John D. Kelleher1,3 1Technological University Dublin, 2Trinity College Dublin, 3ADAPT Centre Dublin, Ireland filip.klubicka@adaptcentre.ie, maldonaa@tcd.ie, abhijit.mahalunkar@mydit.ie, john.d.kelleher@tudublin.ie Abstract This is a resource description paper that describes the creation and properties of a set of pseudo-corpora generated artificially from a random walk over the English WordNet taxonomy. Our WordNet taxonomic random walk implementation allows the exploration of different random walk hyperparameters and the generation of a variety of different pseudo-corpora. We find that different combinations of the walk’s hyperparameters result in varying statistical properties of the generated pseudo-corpora. We have published a total of 81 pseudo-corpora that we have used in our previous research, but have not exhausted all possible combinations of hyperparameters, which is why we have also published a codebase that allows the generation of additional WordNet taxonomic pseudo-corpora as needed. Ultimately, such pseudo-corpora can be used to train taxonomic word embeddings, as a way of transferring taxonomic knowledge into a word embedding space. Keywords: WordNet, taxonomy, random walk, language resource, pseudo-corpus, semantic relationship 1. Introduction Semantic relationships between words or concepts have at least two key dimensions: taxonomic and thematic. Taxo- nomic relations between concepts are based on a compar- ison of the concepts’ features. Concepts that belong to a common taxonomic category share properties or functions. In contrast, thematic relations are formed between con- cepts performing complementary roles in a common event or theme, which often implies having different, albeit com- plementary, features and functions (Kacmajor and Kelleher, 2019)1. When it comes to language and language resources, as a rule of thumb the two semantic relationships are explic- itly encoded in two different kinds of resources: a natu- ral language corpus primarily reflects the thematic relation- ships between words by way of word co-occurrence. Tax- onomic relations, on the other hand, are rarely overtly ex- pressed in examples of natural language. Though research has shown that such relationships can be automatically ex- tracted from natural language corpora (Hearst, 1992), they are more accessible and more commonly modeled in the form of knowledge-engineered language resources such as knowledge bases, ontologies, taxonomies and similar se- mantic networks, where relationships are reflected via ex- plicit links between entities (i.e. nodes) in the knowledge graph. Modelling both kinds of relationships is an important task in building AI with comprehensive natural language under- standing abilities, yet most NLP models and systems, es- pecially language models and word/sentence embeddings, 1In the linguistics literature, the concepts of taxonomic and thematic relatedness would roughly correspond to what are re- spectively called paradigmatic and syntagmatic relations, and there is a nuanced discussion to be had about the extent of the overlap in the terminology. However, as we are focused on re- sources modeling taxonomic relations exclusively, delving deeper into the differences between these terms falls beyond …

Content truncated to 500 words