Author: Yi Yang

contact: [email protected]

Python code for

• NAACL 2015 paper: Unsupervised Domain Adaptation with Feature Embeddings
• ICLR 2015 paper: Unsupervised Domain Adaptation with Feature Embeddings.

• Install gensim by
  • pip install --upgrade gensim
• If you want a faster version of this tool, you may also want to
  • install Cython by
    • pip install cython
  • compile the code by running
    • python setup.py build_ext --inplace

A demo for saving feature embeddings to a txt/bin file is available (python save_embeddings.py -h).

Given a feature file (data/twitter_feat.txt) in which each line corresponds to features of one instance, save feature embeddings to a txt file (data/twitter_embeddings.txt):

1. If features employ bag-of-word (BoW) representation (no feature templates involved)

• python save_embeddings.py --bow 1 --dim 25 data/twitter_feat.txt data/twitter_embeddings.txt

2. If features employ structured representation (extract features by feature templates), and given the feature-template mapping file (data/twitter_feat_template.txt)

• python save_embeddings.py --feature_template_file data/twitter_feat_template.txt --dim 25 data/twitter_feat.txt data/twitter_embeddings.txt

3. If features employ structured representation (extract features by feature templates), and given the template prefix file (data/twitter_template_prefix.txt)

• python save_embeddings.py --template_prefix_file data/twitter_template_prefix.txt --dim 25 data/twitter_feat.txt data/twitter_embeddings.txt

See save_features method of twproc.py for how to generate data/twitter_feat.txt and data/twitter_feat_template.txt files given files in CONLL POS format.

A light demo for part-of-speech tagging of tweets is also provided, using data from CMU Twitter NLP project.

oct27 dataset is regarded as source data, and daily547 dataset is regarded as target data. We also sample some unlabeled tweets randomly (see data/twitter folder).

Run the demo:

1. Prepare the data (extract features, select pivots, etc.) by running

• python twproc.py

2. Obtain the baseline (no adaptation) SVM tagging results by running

• python twpos.py none

3. Obtain the marginalized Denoising Autoencoders adaptation results by running

• python twpos.py mldae

4. Obtain the feature embedding adaptation results by running

• python twpos.py feat2vec

The first step will create a file data/dataset_twitter.pkl. I got results of 0.8839, 0.8889 and 0.8924 for step 2, 3 and 4. The feat2vec results may vary a litter due to the negative sampling technique. You should obtain even better results with feat2vec by using more unlabeled data.