A Deep Learning Architecture for Semantic Address Matching

Codes in this repository are for our IJGIS paper A Deep Learning Architecture for Semantic Address Matching.

Data

Data are available at:

• Shenzhen address corpus (part)
• Semantic address matching dataset

Citation

Please cite the following reference if you use the codes 😊.

@article{Lin+Kang+Wu+Du+Liu:2019,
  author = {Yue Lin and Mengjun Kang and Yuyang Wu and Qingyun Du and Tao Liu},
  title = {A deep learning architecture for semantic address matching},
  journal = {International Journal of Geographical Information Science},
  volume = {0},
  number = {0},
  pages = {1-18},
  year  = {2019},
  publisher = {Taylor & Francis},
  doi = {10.1080/13658816.2019.1681431}
}

Release version of the codes can also be cited as

Details

Below is an overview of each file in this repository.

• geo_config.py Hyperparameter settings for the ESIM
• geo_data_prepare.py Tokenize the corpus and convert each address element into index
• geo_data_processor.py Process the labeled address dataset and divide it into training, development and test sets
• geo_ESIM.py Implementation of the enhanced sequential inference model (ESIM)
• geo_similarity.py Calculate statistical characteristics of the labeled address dataset
• geo_test.py Output predictive results of the ESIM on the test set
• geo_token.py Tokenize with the Jieba library
• geo_train.py Train the ESIM and evaluate its accuracy on the development set
• geo_word2vec.py Train word vectors of address elements
• other_CRF.py Tokenize using CRF [Comber, S.; Arribas-Bel, D. (2019) “Machine learning innovations in address matching: A practical comparison of word2vec and CRFs”. Transactions in GIS, 23 (2): 334–348.]
• other_crf_w2v.py Train word vectors of address elements (CRF tokenizer)
• other_string.py String similarity-based address matching methods: measure the string relevance
• other_w2v_cls.py Use word2vec embeddings directly for classification: calculat cosine similarity