TDParse - Multi-target-specific sentiment recognition on Twitter

From the EACL 2017 paper, TDParse utilises the syntactic information from parse-tree in conjunction with the left-right context of the target and achieves the state-of-the-art performance on both the benchmarking single-target corpus and new multi-target election data.

Our approximated version of the LSTM models proposed in COLING 2016, can be found at here.

Dependencies

• Python 2.7
• sklearn >= 0.18.1
• gensim == 0.13.4
• networkx == 1.11
• ftfy >= 4.1.1
• TweeboParser >= April 1, 2016

Data

You can find our election corpus at here.

Usage

Run TDParse

## e.g. using LibLinear with parameter tuning:
./run.sh lidong tdparse liblinear scale,tune,pred ../data/lidong/parses/lidong.train.conll ../data/lidong/parses/lidong.test.conll
## or without parameter tuning; adding your choice of C-parameter in the end:
./run.sh lidong tdparse liblinear scale,pred ../data/lidong/parses/lidong.train.conll ../data/lidong/parses/lidong.test.conll 0.01

Run Naive-seg

## e.g. using scikit-learn implementation of Linear SVM
./run.sh election naiveseg sklearnSVM

Reference

"TDParse - Multi-target-specific sentiment recognition on Twitter" - Bo Wang, Maria Liakata, Arkaitz Zubiaga, Rob Procter, to be published in EACL 2017

Acknowledgement

Thanks to Duy-Tin Vo and Yue Zhang, the authors of "Target-dependent Twitter Sentiment Classification with Rich Automatic Features", for sharing their code which I have built my implementation upon.