/PatCLS

Codes for Adaptive Taxonomy Learning and Historical Patterns Modelling for Patent Classification

Primary LanguagePython

Adaptive Taxonomy Learning and Historical Patterns Modelling for Patent Classification

Datasets

We provide the preprocessed datasets and NLP pretrained Model at here, which should be put in the ./ folder. The datasets contain two components: patent documents and a hierarchy of IPC classification codes. For patent documents, we provide anonymous IDs for companies, IPC codes in five levels, patent documents, publication times, and descriptions for patents. Under the "Tree" category, we store the hierarchical structure for IPC codes at adjacent levels. If you wish to utilize our work with your datasets, simply ensure they are formatted similarly and train a Word2Vec model as the pre-training model to capture textual information.

To train patent classification model:

  • on USPTO-200K datasets: python train.py --US US
  • on CNPTD-200K datasets: python train.py --US CN

Introduction of our model:

  • Detailed settings such as length of historical patents, slide windows, and dimension of hidden representations are in ./utils/param_parser.py.
  • The whole design of our model is in ./Model/PatentCLS.py, now we use the Bi-LSTM to obtain textual information for patent documents, you could also replace it with BERT model.
  • ./Model/Hierarchy_Emb.py is the implement of IPC Codes Correlations Learning, we get the semantic embeddings of IPC codes in the third-level by learning the hierarchical structure of taxonomy and incorporate the semantic information into patent representations for classification.
  • ./Model/DynamicGraphModule.py is the implement of Historical Application Patterns Learning. We first build the graph in ./utils/util.py before we train our model and put the patent interaction graph into this module to learn the temporal behaviors for companies.

Performance

Datasets Precision(@1) Recall(@1) NDCG(@1) Precision(@3) Recall(@3) NDCG(@3) Precision(@5) Recall(@5) NDCG(@5)
USPTO-200K 0.8341 0.5226 0.8341 0.4856 0.7794 0.8017 0.3342 0.8537 0.8255
CNPTD-200K 0.6848 0.5795 0.6848 0.3428 0.8084 0.7540 0.2257 0.8708 0.7814

Environments: