Methods of Definition Extraction and Linking for Food Recipes

This dissertation deals with definitional contexts extraction and automatic definitions linking in the Italian and English language. Definitional contexts extraction is a task that is not limited to glossaries and encyclopaediae, but has been addressed also in the field of Natural Language Processing. In this research, the objective is to identify definitional contexts in food-related Wikipedia articles. To set the basis of the work, we built two ad-hoc corpora out of the Italian and English dumps of Wikipedia. We trained two BERT models in a supervised fashion with a manually annotated dataset. The F1-measures of 96.08 and 97.66 testify the high performance. We then fed each model with 30 Wikipedia articles randomly extracted from the two corpora, one with Italian and one with English articles. We obtained the best results by restricting the selection to the first sentence of the article whose BERT positive score is above 0.6. The task of automatic definitions linking is loosely based on the wikification process. Rather than linking a term to its corresponding Wikipedia article, we aim at linking a term to its corresponding definition in a Wikipedia article. To lay the foundation of the task, we built two ad-hoc corpora from a cooking website in its Italian and English version. We created a pipeline for automatic definitions linking and carried out a successful experiment using the title of a recipe as input text, the output of which is a minimalistic HTML version of the input, whose terms are linked to their corresponding Wikipedia articles. The definitions linking is one of the two missing steps in the pipeline and discussed in the conclusions.