MultiHead Joint Entity Relation Extraction

Input : Sentence, Entity 1, Entity1_type, Entity1_start_position, Entity1_end_position,Entity 2,Entity2_type, Entity2_start_position, Entity2_end_position,relation

Output : sentence, Entity1, Entity2, Predicted Relation, True Relation

The input of our model is the words of a sentence which are then represented as word vectors(i.e., embeddings). The BiLSTM layer extracts a more complex representation of each word. Then the CRF and the sigmoid layers are able to produce the outputs for the two tasks.

The characters of each word are represented by character vectors that are learned during training. The character embeddings are fed into a BiLSTM and the two final states are concatenated. We take the representation for each word which is the character level representation of the word and then concatenate it to form the word level representation(word2vec) to obtain the complete word embedding vector.

The model has several parameters such as:

EC (entity classification) or BIO (BIO encoding scheme) Character embeddings Ner loss (softmax or CRF) that could be specified in the configuration files.

Download the zip file from the google drive link in the Joint_NER_RE/data/CoNLL04/ and place the zip file at that location.

python main.py

When running for the first time run the following commands and then run python main.py

import nltk

nltk.download('averaged_perceptron_tagger')

Code currently runs on CoNLL04 config file by default.

For other detailed explanation of how to modify the code to run for your dataset (https://github.com/Sanjithae/Joint_NER_RE/blob/master/How_to_run.txt)

1. CoNLL04
2. SemEval 2010 Task 8
3. NYT
4. DDI2013

1. Precision
2. Recall
3. F1 Score

Test Score is 70.766 in 150 epoch on CoNLL04.

Test Score is 68.95 in 150 epoch on SemEval 2010 Task 8.

https://github.com/bekou/multihead_joint_entity_relation_extraction

You can see the execution of the code in this jupyter notebook

Click here to play the video.

Giannis Bekoulis, Johannes Deleu, Thomas Demeester, Chris Develder. Joint entity recognition and relation extraction as a multi-head selection problem. Expert Systems with Applications, Volume 114, Pages 34-45, 2018

Link to the paper : https://arxiv.org/abs/1804.07847