User-factor adaptation is the general task of adapting NLP models to real-valued human attributes, or factors.
Details about user-factor adaptation and our approach to this problem can be found in:
@inproceedings{Lynn2017EMNLP,
title={Human Centered {NLP} with User-Factor Adaptation},
author={Veronica E. Lynn and Youngseo Son and Vivek Kulkarni and Niranjan Balasubramanian and H. Andrew Schwartz},
booktitle={Empirical Methods in Natural Language Processing},
location={Copenhagen, Denmark},
year={2017}
}
Please contact velynn@cs.stonybrook.edu with questions or comments.
This is all the data that was used for the prepositional-phrase attachment task.
The data is formatted as follows:
TWEET <message id> <user id>
<token index> <token> <parent index>
<token index> <token> <parent index>
...
Note that a token will only have its parent listed if it's a preposition. A parent index of 0 indicates that the parent is outside the parse tree. Note that some tweets may contain more than one preposition.
333 of these tweets were obtained from the Tweeboparser dataset. The remaining 986 are new tweets that we annotated.
This is an implementation of the top performing stance detection system from SemEval-2016 Task 6, SVM-ngrams, with user-factor adaptation added to it.
To run with adaptation, change the following variables in config.py.
adapt_type: Whether to use discrete or continuous adaptation. Options are None, "disc", "cont".factor_type: Which factor to adapt to. Options are None, "age", "gender", "personality", "userembed". TF-IDF and word2vec coming soon.
If one or both variables is set to None, no adaptation will be used. Once that's done, simply run python main.py.
Coming soon!