RelationExtraction is a TensorFlow implementation for extracting relations between named entities.
- Implemented using TensorFlow's tf.estimator APIs, which reduces the need for tedious coding.
- Utilizes the Dataset API for efficient data input.
- Includes both PCNN (Piecewise Convolutional Neural Network) and LSTM (Long Short-Term Memory) models.
- Supports both single relation and multi-relation extraction on a single example.
- Introduces a novel method for handling "not available" relations.
The idea behind this program is inspired by the papers pcnn and [bilstm](./Attention-BasedBidirectionalLongShort-TermMemoryNetworksfor RelationClassificatio.pdf). The reference implementation used for this project is based on OpenNRE. However, certain modifications were made to meet specific requirements:
- Multi-relation extraction is supported, resulting in a slightly different multi-instance attention layer compared to OpenNRE.
- The dominant situation of "not available" relations is addressed by treating them as non-relations (all 0 labels), preventing confusion in the neural network.
- Through careful coding and optimization, this program has achieved great success in our knowledge base project and has demonstrated excellent performance.
- Please note that the code for generating tfrecords is not provided, as it was done using Spark.
- Python>=3.5
- TensorFlow>=1.10
I use a parameter to switch models
--ner_procedure birnn,mi_att means bilstm model is working
--ner_procedure pcnn,mi_att means pcnn model is working
you can assemble both to get better performance
1.add elmo to replace embedding layer 2.add bert to replace embedding layer
I also use NYT10 Dataset for comparation
| Model | steps | F1 | precision | recall |
|---|---|---|---|---|
| pcnn | 8000 | 0.52 | 0.63 | 0.41 |
| birnn | 18000 | 0.59 | 0.606 | 0.575 |