Model descriptions

Several models are available for text relation extraction. If you are unsure which model to choose, we recommend starting with the Relation Extraction PCNN model. It provides a balance between runtime efficiency and result quality. The following sections describe each model to help you choose the one that best fits your scenario.

• Relation Extraction PCNN

  • This model uses the Piecewise Convolutional Neural Networks (PCNN) classification model. It trains the relation classification model with an added noise converter module to provide some resistance to noise. Compared with BERT-based models, training and prediction are faster. This model is suitable for scenarios with primarily Chinese content that require a balance between performance, training time, and prediction time.

• Relation Extraction Bert

  • This model uses Bidirectional Encoder Representations from Transformers (BERT), which is pre-trained on a large volume of unlabeled text. It trains the relation classification model by adding a module that fuses entity pair information. This model is suitable for scenarios with primarily Chinese content where the annotated data is clean, performance requirements are high, and training or prediction time is not a major concern.

• Relation Extraction BertNoise

  • This model uses BERT. It trains the relation classification model by adding a module that fuses entity pair information and an anti-noise module. This model is suitable for scenarios with primarily Chinese content where the annotated data is not clean (contains mislabeled or noisy data), performance requirements are high, and training or prediction time is not a major concern.

• Relation Extraction StructBERT-split

  • This model uses StructBERT, which is part of the alicemind deep language model system developed by Alibaba DAMO Academy. It uses a two-stage independent training policy for entity extraction and relation classification. The training process is time-consuming.

• Relation Extraction StructBERT-cascade [Recommended]

  • This model also uses StructBERT. It uses a joint entity and relation extraction policy. This model is less time-consuming, has better overall performance, and is suitable for data with complete entity annotations.