The code for Lexical Sememe Prediction via Word Embeddings and Matrix Factorization(IJCAI2017)
Memory: at least 8GB, 16GB or more is recommended.
Storage: at least 15GB, 20GB or more is recommended.
-
Prepare a file that contains pre-trained Chinese word embeddings(of Google Word2Vec form).We recommend that the amount of words be at least 200,000 and the number of dimentions be at least 200. It will achieve much better result using a large ( 20GB or more is recommended.) corpus to train your embeddings for running this program.
-
Rename the word embedding file as 'embedding_200.txt' and put it under the directory.
-
Run data_generator.sh, the program will automatically generate evaluation data set and other data files required during training.
-
Run SPSE.sh/SPWE.sh/SPASE.sh . The corresponding model will be automatically training and evaluated. (As for SPASE model, it will take pretty much time for training. On a typical computer with CPU of 12 cores, it takes 3 days. For better performance, we suggest that user rewrite it in C++. Model.cpp serves as a simple example.)
-
Run Ensemble_Model.sh after you have run SPSE.sh and SPWE.sh.
(Please check Ensemble_Model.sh, you will get more information about how to run other combinations of models (only support combining 2 models at once)
hownet.txt is an Chinese knowledge base with annotated word-sense-sememe information
After you have run data_generator.sh, you will see 'hownet.txt_test' and 'hownet.txt_answer' file under the directory. These two files make the evaluation set. The size of the evaluation set is 10% of the full size of the part of embedding_200.txt which is anotated in hownet.txt. The evaluation set is generated by random choices.
Feel free to get insight of the files which are named after 'output_', these files contain the sememe predictions for evaluation set.
You can also use pickle library in python to load the files which are named after 'model_'. For more information, please refer to Ensemble_model.py.
If you use the code, please cite this paper:
Ruobing Xie, Xingchi Yuan, Zhiyuan Liu, Maosong Sun. Lexical Sememe Prediction via Word Embeddings and Matrix Factorization. The 26th International Joint Conference on Artificial Intelligence (IJCAI 2017).