/socialsent

Code and data for inducing domain-specific sentiment lexicons.

Primary LanguagePythonApache License 2.0Apache-2.0

SocialSent: Domain-Specific Sentiment Lexicons for Computational Social Science

Authors: William L. Hamilton and Kevin Clark

Overview

SocialSent is a package for inducing and analyzing domain-specific sentiment lexicons. A number of state-of-the-art algorithms are included, including SentProp and Densifier (http://www.cis.lmu.de/~sascha/Ultradense/). A detailed description of the algorithms in the SocialSent package, with references, is provided in the paper: Inducing Domain-Specific Sentiment Lexicons from Unlabeled Corpora.

The project website includes pre-constructed sentiment lexicons for 150 years of historical English and 250 online communities from the social media forum Reddit.

If you make use of this work in your research please cite the following paper:

William L. Hamilton, Kevin Clark, Jure Leskovec, and Dan Jurafsky. Inducing Domain-Specific Sentiment Lexicons from Unlabeled Corpora. Proceedings of EMNLP. 2016. (to appear; arXiv:1606.02820).

Install

pip install socialsent will work but has some downsides right now. In particular, if you use the pip install method, you will need to know where the installation directory is in order to modify the paths in the constants.py folder. You also won't have access to the examples that are discussed below if you install via pip.

As a preferred alternative, download the source from the GitHub repo and run python setup.py install

Note that you should re-run this command every time after you update the paths in constants.py folder.

If you run into issues related to Keras dependencies: You can use the provided environment.yml file to set up a conda environment that should solve these issues. If you still run into issues after creating the conda environment, you might need to change the Keras backend to Theano (instead of TensorFlow). Many thanks to @sashank06 for this fix.

Using the code

To use SentProp you will need to either build some word vector embeddings or download some that are pre-trained. Once this is done, you would specify the path to these embeddings in constants.py. The file constants.py also contains some links to pre-trained embeddings that were used in the paper mentioned above. Running example.sh will download some pre-trained embeddings and run SentProp on them (using the code in example.py). You can build embeddings yourself with the code in the representations directory, which is based upon code in https://github.com/williamleif/historical-embeddings This code also illustrates how to use the SocialSent methods.

The file polarity_induction_methods.py contains implementations for a suite of sentiment induction algorithms, along with some comments/documentation on how to use them. The file evaluate_methods.py also includes the evaluation script used in our published work.

NB: Right now the code uses dense numpy matrices in a (relatively) naive way and thus has memory requirements proportional to the square of the vocabulary size; with a reasonable amount of RAM, this works for vocabs of size 20000 words or less (which is reasonable for specific domain), but there are definitely optimizations that could be done, exploiting sparsity etc. I hope to get to these optimizations soon, but feel free to submit a pull request :).

NB: The random walk based implementation of the SentProp algorithm is quite sensitive to the embedding method used and pre-processing. We found it worked well on very small corpora with our "default" SVD-based embeddings on a restricted vocabulary (<50000 words), i.e. using context distribution smoothing (with a smoothing value of 0.75) and throwing away the singular values, as described in the paper. However, its performance varies substantially with embedding pre-processing. On large vocabularies/datasets and using more general embeddings/pre-processing (e.g., word2vec), the Densifier method usually achieves superior performance (as described in the paper, tables 2a-b), with less sensitivity to pre-processing.

Dependencies

The code is not currently compatible with the newest Keras distribution (1.0); only the "denisfy"/Densifier method requires this package, however. So you can either (a) set up a conda environment using the provided environment.yml file, (b) install an older Keras (0.3) manually, or (c) remove all calls to the "densify" method. I aim to update this dependency in the near future

An up-to-date Python 2.7 distribution, with the standard packages provided by the anaconda distribution is required. However, the code was only tested with some versions of these packages.

In particular, the code was tested with:

  • theano (0.8.0)
  • keras (0.3.3)
  • numpy (1.11.0)
  • scipy (0.15.1)
  • sklearn (0.18.dev0 or 0.17.1)