In this project, it was a question of designing a model capable of recognizing terms related to our field which is: INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES.
We built our gold set according to the steps you will see below, then we trained the spacy model on our patent to see how it recognizes terms. The model we got recognized some irrelevant and out of context terms, so we used prodigy to correct these errors. Then we re-trained and re-evaluated on the test set. However, we also tried to fine tune Bert's variants, more precisely roberta and distillbert, but we couldn't get to the end due to lack of computing power. The idea was to also make the transformers and compare them to the basic models of spacy.
The second part of the project deals with the extraction of relationships between entities. We considered an entity as a verb or a verb phrase that links two entities in the style subject verb object. We used the noun chunks approach but it turns out that the relations found do not always make sense, this is surely because the noun chunks are not all related to our domain. That's why we used our model to extract the entities linked to our domain. And then extract the verb phrases that link the entities according to the rule defined above.