ryanzhumich / sirnn

SIRNN

Theano implementations of the Speaker Interaction RNNs(SIRNN) in the following paper:

Addressee and Response Selection in Multi-Party Conversations with Speaker Interaction RNNs.

The code and data is based on Addressee and Response Selection for Multi-Party Conversation

Dependencies

• Python 2.7
• Theano 0.9.0

Data

• To run experiments with 10 candidates, concatentate train-data.cand-10.1.gz and train-data.cand-10.2.gz into train-data.cand-10.gz: zcat train-data.cand-10.1.gz train-data.cand-10.2.gz | gzip > train-data-cand-10.gz
• Download 300-d Glove embedding and save it as data/glove.840B.300d.txt

Usage

• SIRNN with 2 candidates and 15 context length: cd code; ./run_interleaveGRUcollab_match_CAND2_Nc15.sh