Skip to content
Improving topic models LDA and DMM (one-topic-per-document model for short texts) with word embeddings (TACL 2015)
Branch: master
Clone or download
Latest commit 2d074e5 May 7, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
jar Inference of topic distribution on unseen corpus Nov 11, 2016
lib LFTM commit Jul 2, 2015
src Inference of topic distribution on unseen corpus Nov 11, 2016
test Inference of topic distribution on unseen corpus Nov 11, 2016
License.txt Inference of topic distribution on unseen corpus Nov 11, 2016
README.md
build.xml LFTM commit Jul 2, 2015

README.md

LF-LDA and LF-DMM latent feature topic models

The implementations of the LF-LDA and LF-DMM latent feature topic models, as described in my TACL paper:

Dat Quoc Nguyen, Richard Billingsley, Lan Du and Mark Johnson. Improving Topic Models with Latent Feature Word Representations. Transactions of the Association for Computational Linguistics, vol. 3, pp. 299-313, 2015. [.bib] [Datasets] [Example_20Newsgroups_20Topics_Top50Words]

The implementations of the LDA and DMM topic models are available at http://jldadmm.sourceforge.net/

Usage

This section describes the usage of the implementations in command line or terminal, using the pre-compiled LFTM.jar file.

Here, it is expected that Java 1.7+ is already set to run in command line or terminal (for example: adding Java to the path environment variable in Windows OS).

The pre-compiled LFTM.jar file and source codes are in the jar and src folders, respectively. Users can recompile the source codes by simply running ant (it is also expected that ant is already installed). In addition, the users can find input examples in the test folder.

File format of input topic-modeling corpus

Similar to the corpus.txt file in the test folder, each line in the input topic-modeling corpus represents a document. Here, a document is a sequence words/tokens separated by white space characters. The users should preprocess the input topic-modeling corpus before training the topic models, for example: down-casing, removing non-alphabetic characters and stop-words, removing words shorter than 3 characters and words appearing less than a certain times.

Format of input word-vector file

Similar to the wordVectors.txt file in the test folder, each line in the input word-vector file starts with a word type which is followed by a vector representation.

To obtain the vector representations of words, the users can use the pre-trained word vectors learned from large external corpora OR the word vectors which are trained on the input topic-modeling corpus.

In case of using the pre-trained word vectors learned from the large external corpora, the users have to remove words in the input topic-modeling corpus, in which these words are not found in the input word-vector file.

Some sets of the pre-trained word vectors can be found at:

Word2Vec: https://code.google.com/p/word2vec/

Glove: http://nlp.stanford.edu/projects/glove/

If the input topic-modeling corpus is too domain-specific, the domain of the external corpus (from which the word vectors are derived) should not be too different to that of the input topic-modeling corpus. For example, when applying to the biomedical domain, the users may use Word2Vec or Glove to learn 50 or 100-dimensional word vectors on the large external MEDLINE corpus instead of using the pre-trained Word2Vec or Glove word vectors.

Training LF-LDA and LF-DMM

$ java [-Xmx2G] -jar jar/LFTM.jar –model <LFLDA_or_LFDMM> -corpus <Input_corpus_file_path> -vectors <Input_vector_file_path> [-ntopics <int>] [-alpha <double>] [-beta <double>] [-lambda <double>] [-initers <int>] [-niters <int>] [-twords <int>] [-name <String>] [-sstep <int>]

where hyper-parameters in [ ] are optional.

  • -model: Specify the topic model.

  • -corpus: Specify the path to the input training corpus file.

  • -vectors: Specify the path to the file containing word vectors.

  • -ntopics <int>: Specify the number of topics. The default value is 20.

  • -alpha <double>: Specify the hyper-parameter alpha. Following [1, 2], the default value is 0.1.

  • -beta <double>: Specify the hyper-parameter beta. The default value is 0.01. Following [2], you might also want to try beta value of 0.1 for short texts.

  • -lambda <double>: Specify the mixture weight lambda (0.0 < lambda <= 1.0). Set the mixture weight lambda to be 1.0 to obtain the best topic coherence. In case of document clustering/classification evaluation, fine-tune this parameter to obtain the highest results if you have time; otherwise try both values 0.6 and 1.0 (I would suggest to set lambda 0.6 for normal text corpora and 1.0 for short text corpora if you don't have time to try both 0.6 and 1.0).

  • -initers <int>: Specify the number of initial sampling iterations to separate the counts for the latent feature component and the Dirichlet multinomial component. The default value is 2000.

  • -niters <int>: Specify the number of sampling iterations for the latent feature topic models. The default value is 200.

  • -twords <int>: Specify the number of the most probable topical words. The default value is 20.

  • -name <String>: Specify a name to the topic modeling experiment. The default value is “model”.

  • -sstep <int>: Specify a step to save the sampling output (-sstep value < -niters value). The default value is 0 (i.e. only saving the output from the last sample).

NOTE that (topic vectors are learned in parallel, so) running LFTM code with multiple CPU/core machine to obtain a significantly faster training process, e.g. using a multi-core computer, or set number of CPUs requested for a remote job to be equal to number of topics.

Examples:

$ java -jar jar/LFTM.jar -model LFLDA -corpus test/corpus.txt -vectors test/wordVectors.txt -ntopics 4 -alpha 0.1 -beta 0.01 -lambda 0.6 -initers 500 -niters 50 -name testLFLDA

Basically, with this command we run 500 LDA sampling iterations (i.e., -initers 500) for initialization and then run 50 LF-LDA sampling iterations (i.e., -niters 50). The output files are saved in the same folder as the input training corpus file, in this case in the test folder. We have output files of testLFLDA.theta, testLFLDA.phi, testLFLDA.topWords, testLFLDA.topicAssignments and testLFLDA.paras, referring to the document-to-topic distributions, topic-to-word distributions, top topical words, topic assignments and model hyper-parameters, respectively. Similarly, we perform:

$ java -jar jar/LFTM.jar -model LFDMM -corpus test/corpus.txt -vectors test/wordVectors.txt -ntopics 4 -alpha 0.1 -beta 0.1 -lambda 1.0 -initers 500 -niters 50 -name testLFDMM

We have output files of testLFDMM.theta, testLFDMM.phi, testLFDMM.topWords, testLFDMM.topicAssignments and testLFDMM.paras.

In the LF-LDA and LF-DMM latent feature topic models, a word is generated by the latent feature topic-to-word component OR by the topic-to-word Dirichlet multinomial component. In practical implementation, instead of using a binary selection variable to record that, I simply add a value of the number of topics to the actual topic assignment value. For example with 20 topics, the output topic assignment 3 23 4 4 24 3 23 3 23 3 23 for a document means that the first word in the document is generated from topic 3 by the latent feature topic-to-word component. The second word is also generated from the topic 23 - 20 = 3, but by the topic-to-word Dirichlet multinomial component. It is similar for the remaining words in the document.

Document clustering evaluation

Here, we treat each topic as a cluster, and we assign every document the topic with the highest probability given the document. To get the clustering scores of Purity and normalized mutual information, we perform:

$ java –jar jar/LFTM.jar –model Eval –label <Golden_label_file_path> -dir <Directory_path> -prob <Document-topic-prob/Suffix>

  • –label: Specify the path to the ground truth label file. Each line in this label file contains the golden label of the corresponding document in the input training corpus. See the corpus.LABEL and corpus.txt files in the test folder.

  • -dir: Specify the path to the directory containing document-to-topic distribution files.

  • -prob: Specify a document-to-topic distribution file or a group of document-to-topic distribution files in the specified directory.

Examples:

The command $ java -jar jar/LFTM.jar -model Eval -label test/corpus.LABEL -dir test -prob testLFLDA.theta will produce the clustering score for the testLFLDA.theta file.

The command $ java -jar jar/LFTM.jar -model Eval -label test/corpus.LABEL -dir test -prob testLFDMM.theta will produce the clustering score for testLFDMM.theta file.

The command $ java -jar jar/LFTM.jar -model Eval -label test/corpus.LABEL -dir test -prob theta will produce the clustering scores for all the document-to-topic distribution files having names ended by theta. In this case, the distribution files are testLFLDA.theta and testLFDMM.theta. It also provides the mean and standard deviation of the clustering scores.

Inference of topic distribution on unseen corpus

To infer topics on an unseen/new corpus using a pre-trained LF-LDA/LF-DMM topic model, we perform:

$ java -jar jar/LFTM.jar -model <LFLDAinf_or_LFDMMinf> -paras <Hyperparameter_file_path> -corpus <Unseen_corpus_file_path> [-initers <int>] [-niters <int>] [-twords <int>] [-name <String>] [-sstep <int>]

  • -paras: Specify the path to the hyper-parameter file produced by the pre-trained LF-LDA/LF-DMM topic model.

Examples:

$ java -jar jar/LFTM.jar -model LFLDAinf -paras test/testLFLDA.paras -corpus test/corpus_test.txt -initers 500 -niters 50 -name testLFLDAinf

$ java -jar jar/LFTM.jar -model LFDMMinf -paras test/testLFDMM.paras -corpus test/corpus_test.txt -initers 500 -niters 50 -name testLFDMMinf

Acknowledgments

The LF-LDA and LF-DMM implementations used utilities including the LBFGS implementation from MALLET toolkit, the random number generator in Java version of MersenneTwister, the Parallel.java utility from Mines Java Toolkit and the Java command line arguments parser. I would like to thank the authors of the mentioned utilities for sharing the codes.

References

[1] Yue Lu, Qiaozhu Mei, and ChengXiang Zhai. 2011. Investigating task performance of probabilistic topic models: an empirical study of PLSA and LDA. Information Retrieval, 14:178–203.

[2] Jianhua Yin and Jianyong Wang. 2014. A Dirichlet Multinomial Mixture Model-based Approach for Short Text Clustering. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 233–242.

You can’t perform that action at this time.