Dynamic Joint Sentiment-Topic Model (dJST)

This project aims to extend the Joint Sentiment-Topic Model (JST) by a dynamic time component. With this, sentiment-topics can be analyzed over time to do trend analytics.

Getting Started

To compile the project type make in a shell. Alternatively you can import the project into Visual Studio.

Properties

// TODO describe how to set the properties

Estimation (JST)

For training on a single dataset execute: jst -est -config YOUR-PATH/train.properties This is equivalent to execute JST.

Inference/Continnous Estimation (dJST)

To continuously train on new data execute: jst -djst -config YOUR-PATH/test.properties

Output

The trained parameters/distributions will be written out in the following files:

• <iter>.others Holds summaries on the corpus statistics, like how many docs and words have been discovered. Also the training parameter details will be displayed (how many iterations, topics, senti-labels)
• <iter>.pi Contains the per-document sentiment distributions
• <iter>.phi Contains the sentiment specific topic-word distributions
• <iter>.theta Contains the per-document sentiment specific topic proportions. For each topic the sentiment distribution is shown
• <iter>.tassign Contains the sentiment label and topic assignments for words in training data. After sampling each word was assigned to a topic- and sentiment-label

Data

The data input format should be like the following:

[Doc_1 name] [token_1] [token_2] ... [token_N]
     :
     :
[Doc_M name] [token_1] [token_2] ... [token_N]

Each line is a document and it is assumed, that data preprocessing has been performed already (removal of stop words, stemming,...).

The algorithm needs prior information on the sentiment-polarity of words. The sentiment lexicon used here is MPQA and has the following format:

[word]	[neu prior prob.] [pos prior prob.] [neg prior prob.]

Authors

• Chenghua Lin - Initial work JST - linron84
• Kolja Maier extension to dJST

License

This project is licensed under the GPL License - see the LICENSE.md file for details

Acknowledgments

• Fur further documentation on classes and functions see Help

References

• Lin, He, Yulan and Lin, Chenghua and Gao, Wei and Wong, Kam-Fai Dynamic joint sentiment-topic model, ACM Transactions on Intelligent Systems and Technology, 2013.
• Lin, C., He, Y., Everson, R. and Reuger, S. Weakly-supervised Joint Sentiment-Topic Detection from Text, IEEE Transactions on Knowledge and Data Engineering (TKDE), 2011.
• Lin, C. and He, Y. Joint Sentiment/Topic Model for Sentiment Analysis, In Proceedings of the 18th ACM Conference on Information and Knowl- edge Management (CIKM), Hong Kong, China, 2009.