Requirements

• python2.7 (development python3 code available in python3 branch; code still requires testing)
• gensim: pip install gensim
• tensorflow 0.8-0.12

Data Format

• One line per document
• Sentences are delimited by tabs in each document
• See examples in data/
• ACL2017 Paper dataset (AP News, BNC and IMDB)

Running the code (example.sh)

Train a word2vec model using gensim. This step is optional, you'll only need to do this if you want to initialise TDLM with pre-trained embeddings. word2vec model settings are in the python file (word2vec.py)

python word2vec_train.py

Train a model; configurations/hyper-parameters are defined in tdlm_config.py

python tdlm_train.py

All test inferences are invoked with tdlm_test.py. E.g. to compute language and topic model perplexity

python tdlm_test.py -m output/toy-model/ -d data/toy-valid.txt --print_perplexity

Print topics (to topics.txt)

python tdlm_test.py -m output/toy-model/ -d data/toy-valid.txt --output_topic topics.txt

Infer topic distribution in documents (saved as a npy file)

python tdlm_test.py -m output/toy-model/ -d data/toy-valid.txt --output_topic_dist topic-dist.npy

Generate sentences conditioned on topics

python tdlm_test.py -m output/toy-model/ -d data/toy-valid.txt --gen_sent_on_topic topic-sents.txt

tdlm_test.py arguments:

usage: tdlm_test.py [-h] -m MODEL_DIR [-d INPUT_DOC] [-l INPUT_LABEL]
                    [-t INPUT_TAG] [--print_perplexity] [--print_acc]
                    [--output_topic OUTPUT_TOPIC]
                    [--output_topic_dist OUTPUT_TOPIC_DIST]
                    [--output_tag_embedding OUTPUT_TAG_EMBEDDING]
                    [--gen_sent_on_topic GEN_SENT_ON_TOPIC]
                    [--gen_sent_on_doc GEN_SENT_ON_DOC]

Given a trained TDLM model, perform various test inferences

optional arguments:
  -h, --help            show this help message and exit
  -m MODEL_DIR, --model_dir MODEL_DIR
                        directory of the saved model
  -d INPUT_DOC, --input_doc INPUT_DOC
                        input file containing the test documents
  -l INPUT_LABEL, --input_label INPUT_LABEL
                        input file containing the test labels
  -t INPUT_TAG, --input_tag INPUT_TAG
                        input file containing the test tags
  --print_perplexity    print topic and language model perplexity of the input
                        test documents
  --print_acc           print supervised classification accuracy
  --output_topic OUTPUT_TOPIC
                        output file to save the topics (prints top-N words of
                        each topic)
  --output_topic_dist OUTPUT_TOPIC_DIST
                        output file to save the topic distribution of input
                        docs (npy format)
  --output_tag_embedding OUTPUT_TAG_EMBEDDING
                        output tag embeddings to file (npy format)
  --gen_sent_on_topic GEN_SENT_ON_TOPIC
                        generate sentences conditioned on topics
  --gen_sent_on_doc GEN_SENT_ON_DOC
                        generate sentences conditioned on input test documents

Publication

Jey Han Lau, Timothy Baldwin and Trevor Cohn (2017). Topically Driven Neural Language Model. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (ACL 2017), Vancouver, Canada, pp. 355--365.