A python package that can run this automated pipeline in several lines of codes, together with instructions, is included in the EntityRelationExtraction folder.
Project Objective - This project aims to formulate a better framework of obtaining entity relation dataset from text documents. Specifically, we will perform entity extraction from text documents in PubMed dataset, get the phrases connecting every pair of entities, summarize the relations, and cluster the summarizations so similar relationships can be described by a centroid term.
To install all the requirement for this repository Run
bash requirement.txt
Authors: + Ruilin Liu (rl3234)(Team Captain) + Zhifeng Zhang (zz2884) + Jessie Wang (yw3765) + Zhiqing Yang (zy2491) + Zhucheng Zhan (zz2783)
Sponsor/Mentor: - John Labarga (john.labarga@unilever.com) from Unilever
The objective of this capstone is to evaluate the feasibility and performance of a novel strategy for producing training examples for relation extractors. This strategy seeks to make a trade-off where lower fidelity is tolerated to achieve lower cost and broader scope in producing training examples.
The central hypothesis of this effort is that transformer-based sumamrization models have gotten good enough to identify a term that describes the relationship between two entities in a sentence or paragraph (what is sometimes called the verb in the relation extraction literature, when the relation extraction problem is framed as identifying (subject, verb, object) triples). There is a secondary hypothesis here that this verb is present in the text between two entities, however this hypothesis seems plasuble since it also underlies the premise of relation extraction.
The basic workflow of this approach is as follows:
-
Pass documents through an entity extractor to identify entities.
-
For each entity pair within a given proximity, use a summarization model to extract a token or small set of tokens (3 tokens at most) that describes the relationship between the entities. What "proximity" means here will need to be tuned via experimentation, in the simplest case this means two adjacent entities. These tokens or small sets of tokens become the "relation candidates".
-
The output of step 2 is expected to be sparse and not immediately usable. To condense the candidates, use semantic clustering to cluster the relation candidates. Then identify a centroid vector for each cluster. Finally, use a decoder (this could be as simple as Word2Vec) to get a token (or small set of tokens) which is most indicative of that cluster.
-
Produce a training set by tagging the entity pairs from step 2 with the centroid relationships from step 3.
-
Fine-tune a relation extractor with the training examples fro step 4, and compare with an un-tuned relation extractor.