yao8839836/kge-lda

Knowledge Graph Embedding LDA. AAAI 2017

KGE-LDA

The implementation of Knowledge Graph Embedding LDA in our paper:

Liang Yao, Yin Zhang, Baogang Wei, Zhe Jin, Rui Zhang, Yangyang Zhang, and Qinfei Chen. "Incorporating Knowledge Graph Embeddings into Topic Modeling." In Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17), pp. 3119-3126. AAAI Press, 2017.

Require Java 7+ and Eclipse.

Implementation

KGE-LDA(a): KGE-LDA/src/topic/LinkLDA_KGE.java

KGE-LDA(b): KGE-LDA/src/topic/CorrLDA_KGE.java

Main entries

The main entries for KGE-LDA(a), KGE-LDA(b), Corr-LDA, CI-LDA (Link-LDA), CTM and LDA on the three datasets 20NG, NIPS and Ohsumed are in KGE-LDA/src/result20ng/, KGE-LDA/src/resultnips/ and KGE-LDA/src/resultohsumed/. The main entries are ResultXXXXXX.java

The main entries for GK-LDA on the three datasets 20NG, NIPS and Ohsumed are in GKLDA-master/Src/src/launch. The main entries are ResultGKLDAXXX.java

The main entries for LF-LDA on the three datasets 20NG, NIPS and Ohsumed are in LFTM/src/test/. The main entries are ResultLFLDAXXX.java

Reproduce results

To reproduce the results in the paper:

(1) Decompress /KGE-LDA/data.zip and /KGE-LDA/file.7z in the same fold.

(2) Download the three Wikipedia index files 20ng_word_wiki_small_index.zip (http://pan.baidu.com/s/1qXDVoVq and https://drive.google.com/open?id=1yA3tJNIFoZYQQCMMieQCzN0ABUscJNYB), nips_word_wiki_index_small.rar (http://pan.baidu.com/s/1hs9HZve and https://drive.google.com/file/d/1z7y003TsprENKJ3V92sG_SHwst5MFWBN/view?usp=sharing) and ohsumed_23_word_wiki_index.zip (http://pan.baidu.com/s/1miATVkO and https://drive.google.com/open?id=127zVR2GH7AXvIL1xRbXWWzQaSDwzxtLG), decompress them and put the decompressed folders to the GKLDA-master/Src/file/, /KGE-LDA/file/ and /LFTM/file/. The 4,776,093 Wikipedia articles are at (http://pan.baidu.com/s/1slaTPoT and https://drive.google.com/open?id=1JTLu-AqNhf7xUWTgKKm8360wT0OQve7O), I extracted them from http://deepdive.stanford.edu/opendata/.

(3) Run the main entries.

Others

(1) The raw text datasets are in three folds under /KGE-LDA/data/.

(2) The tokenized documents are in /KGE-LDA/file/20ng/, /KGE-LDA/file/nips/ and /KGE-LDA/file/ohsumed/. Documents after stopwords removing are in /KGE-LDA/file/xxx_remove_stop. Documents after stopwords and rare words removing are in /KGE-LDA/file/xxx_remove_rare.

(3) The input text for the KGE-LDA model should be like (please decompress data.zip): data//corpus_20ng.txt. Each line represents a document. Each number is an index of a word in the vocabulary.

(4) The vocabulary should be like (please decompress data.zip): data//vocab_20ng.txt. Each line is a word, the first word is 0 in data//corpus_20ng.txt, the second word is 1 in data//corpus_20ng.txt, the third word is 2 in data//corpus_20ng.txt...

(5) The linked entities in WordNet(via NLTK) of each document are in /KGE-LDA/file/20ng_wordnet/, /KGE-LDA/file/nips_wordnet/ and /KGE-LDA/file/ohsumed_wordnet/. Their ids are in /KGE-LDA/file/xxx_wordnet_id/. Each file name represents the index of the document.

(6) file/runnltk.py is an example of entity linking for Ohsumed dataset.

(7) To tokenize your own documents, you also need to download the model file of Stanford CoreNLP (http://pan.baidu.com/s/1bpDqa7d) and add it to the class path.

(8) The unique entities ids for 20NG are in /KGE-LDA/knowledge/WN18/entity_appear.txt, the unique entities ids for NIPS are in /KGE-LDA/knowledge/WN18/entity_appear_nips.txt, the unique entities ids for Ohsumed are in /KGE-LDA/knowledge/WN18/entity_appear_ohsumed.txt

(9) I used this implementation of TransE to obtain entity embeddings: https://github.com/thunlp/KB2E. See the Readme of the project for more details about how to prepare knowledge graphs and obtain embeddings.

(10) The 50 dimensional entity embeddings for 20NG are in /KGE-LDA/knowledge/WN18/entity2vec_appear.bern, the 50 dimensional entity embeddings for NIPS are in /KGE-LDA/knowledge/WN18/entity2vec_appear_nips.bern, the 50 dimensional entity embeddings for Ohsumed are in /KGE-LDA/knowledge/WN18/entity2vec_appear_ohsumed.bern. To get all 50 dimensional entity embeddings in WN18, see the three files: /KGE-LDA/knowledge/WN18/entity2vec.bern, /KGE-LDA/knowledge/WN18/entity2id.txt, /KGE-LDA/knowledge/WN18/num_synset.txt

kge-lda's Issues

Cool

3q

code

Excuse me, this code is damaged after decompressing. Can you upload it again？thanks

what is switch?

thanks for your code,but i want to know what is switchLDA_KGE and switchLDA.

问题

大佬好，我想问一下entity_appear.txt这个文件是怎么得到的，非常感谢