MetaLDA is a topic model that leverages either document or word meta information, or both of them jointly [1].

Key features:

1. Incorporates both document and word meta information in binary format
2. Implemented on top of Mallet in JAVA

• Works with Mallet input format
• Runs efficiently (bit-coding and SparseLDA framework apply)
• Runs with multi-threads (DistributedLDA framework applies)

1. Requirements: JDK 1.8 or later and Maven. Other dependencies will be downloaded by Maven.
2. Clone the repository or download the code
3. Compile the code with Maven

• cd <metalda_location>
• mvn package

4. Prepare documents

   • All documents (training/testing) are in Mallet's LabeledLDA format.
   • If the input documents are already in Mallet, they can be directly fed into the model.
   • Otherwise, the documents have to be first converted into Mallet format.
     • Each raw document should in the following format:
       DOC_ID\tLABEL1\sLABEL2\sLABEL3\tWORD1\sWORD2\sWORD3\n.
     • Install Mallet then use:
       <mallet_location>/bin/mallet import-file --input <training/testing_doc_location> --output <training/testing_doc_mallet_location> --label-as-features --keep-sequence --line-regex '([^\t]+)\t([^\t]+)\t(.*)'

5. Prepare word features

   • MetaLDA uses the following sparse representation of binary word features:
     WORD\tNNZ_INDEX1\sNNZ_INDEX2\sNNZ_INDEX3
   • Use embeddings as word features
     • MetaLDA offers a function to binarise and convert word embeddings into the required word feature format. The raw input word embeddings are expected to follow the format of GloVe:
       WORD\sEMBEDDING1\sEMBEDDING2\sEMBEDDING3
     • To binarise and convert the raw word embeddings, in the root folder of MetaLDA, use:
       java -cp ./target/metalda-0.1-jar-with-dependencies.jar topicmodels.BinariseWordEmbeddings --train-docs <training_doc_mallet_location> --test-docs <testing_doc_mallet_location> --input <raw_embedding_location> --output <binary_embedding_location>
     • The function first reads the vocabularies of the training and testing documents (both in Mallet format) and then binarise the embeddings of the words in the vocabularies stored in the word embedding file, and finally saves the binarised embeddings into the required format. Note that MetaLDA does not require all the words in the training and testing documents have embeddings.

6. Train MetaLDA
   java -cp ./target/metalda-0.1-jar-with-dependencies.jar topicmodels.MetaLDATrain --train-docs <training_doc_mallet_location> --num-topics <num_topic> --word-features <binary_embedding_location> --save-folder <save_folder> --sample-alpha-method <sample_alpha_method> --sample-beta-method <sample_beta_method>

• <sample_alpha_method>:
  • 0: fixed on initial value
  • 1: alpha is a full matrix sampled with doc labels
  • 2: alpha is sampled as an asymmetric vector over topic
  • 3: alpha is sampled as a single value
• <sample_beta_method>:
  • 0: fixed on initial value
  • 1: beta is a full matrix sampled with word features
  • 2: beta is sampled as an asymmetric vector over topics
  • 3: beta is sampled as a single value
• For details of the arguments, use:
  java -cp ./target/metalda-0.1-jar-with-dependencies.jar topicmodels.MetaLDATrain --help

7. Access the saved files in the training phrase
   In the training phrase, MetaLDA saves the following files in the <save_folder>:

• top_words.txt:
  the top 50 words with the largest weights (phi) in each topic (the number of top words can be changed)
• train_alphabet.txt:
  the vocabulary of the training documents, the order of the words matches the index of phi.
• train_target_alphabet.txt:
  the vocabulary of the labels in the training documents, the order of the labels matches the index of lambda
• train_stats.mat:
  a MAT-file of Matlab that saves the training statistics. matfilerw for JAVA and FileIO of Scipy for Python are good tools to access MAT-files. Note that Matlab installation is not required although Matlab can directly load MAT-files.

8. Inference on the testing documents
   MetaLDA offers two kinds of inference:

• Ignore the words that exist in the testing documents but not in the training documents
  java -cp ./target/metalda-0.1-jar-with-dependencies.jar topicmodels.MetaLDAInfer --test-docs <testing_doc_mallet_location> --save-folder <save_folder> --compute-perplexity true
  • <save_folder>: same to the folder where the files are saved in the training phrase
  • --compute-perplexity
    • true: MetaLDA will use one half of each testing document (every first words) to sample its document-topic distribution (theta) and the other half (every second words) to compute perplexity.
    • false: MetaLDA will use all the content of each testing document to sample its document-topic distribution (theta). Perplexity will not be computed.
• Consider the words that exist in the testing documents but not in the training documents
  java -cp ./target/metalda-0.1-jar-with-dependencies.jar topicmodels.MetaLDAInferUnseen --test-docs <testing_doc_mallet_location> --save-folder <save_folder> --compute-perplexity true --word-features <binary_embedding_location>

9. Access the saved files in the inference phrase

• If MetaLDAInfer is used, MetaLDA will save the testing statistics into 'test_stats.mat' in <save_folder>
• If MetaLDAInferUnseen is used, MetaLDA will save the testing statistics into 'test_stats_unseen.mat' in <save_folder>

To wrap up the above steps, a demo bash script is offered. To run the demo, simply go to the demo folder and run ./demo.sh. For the demo, the WS dataset used in the paper is included.

[1] H. Zhao, L. Du, W. Buntine, G. Liu, "MetaLDA: a Topic Model that Efficiently Incorporates Meta information", International Conference on Data Mining (ICDM) 2017. Arxiv

If you find any bugs, please contact [email protected].