Hugging Face's logo

sagteam
/
pharm-relation-extraction

Text Classification
Transformers
PyTorch
xlm-roberta
Inference Endpoints
Model card Files Community
1
pharm-relation-extraction / README.md
sunny3's picture
sunny3
Create README.md
d552201
preview code
|
raw
history blame
No virus
2.59 kB

Relation_Extraction

This repository provides code and additional materials of the paper: "Extraction of the Relations between Significant Pharmacological Entities in Russian-Language Reviews of Internet Users on Medications".

In this work, we trained a model to recognize 4 types of relationships between entities in drug review texts: ADR–Drugname, Drugname–Diseasename, Drugname–SourceInfoDrug, Diseasename–Indication. The input of the model is a review text and a pair of entities, between which it is required to determine the fact of a relationship and one of the 4 types of relationship, listed above.

Data

Proposed model is trained on a subset of 908 reviews of the Russian Drug Review Corpus (RDRS). The subset contains the markup of the following types of entities: and contains pairs of entities marked with the 4 listed types of relationships:

  • ADR-Drugname — the relationship between the drug and its side effects
  • Drugname-SourceInfodrug — the relationship between the medication and the187source of information about it (e.g., “was advised at the pharmacy”, “the -Drugname-- - Drugname-Diseasename — the relationship between the drug and the disease
  • Diseasename-Indication — the connection between the illness and its symptoms (e.g., “cough”, “fever 39 degrees”) Also, this subset contains pairs of the same entity types between which there is no relationship: for example, a drug and an unrelated side effect that appeared after taking another drug; in other words, this side effect is related to another drug.

Model

Proposed model is based on the XLM-RoBERTA-large topology. After the additional training as a langauge model on corpus of unmarked drug reviews, this model was trained as a classification model on 80% of the texts from subset of the corps described above. This model showed the best accuracy on one of the folds of the cross-validation. For additional details see original paper.

How to use

See section "How to use" in our git repository for the model

Results

Here are the accuracy, estimated by the f1 score metric for the recognition of relationships on the best fold.

ADR–Drugname Drugname–Diseasename Drugname–SourceInfoDrug Diseasename–Indication
0.955 0.892 0.922 0.891