apropos13 / bayesian-and-lstm-networks-in-nlg

A comparison of Bayesian and LSTM Networks in Natural Language Generation

Repo for the advanced ML class 290C.

Includes code and a report of a thorough comparison of a dynamic Bayesian Network and an LSTM Network in Natural Language Generation. We use the semantically conditioned LSTM network proposed in http://www.emnlp2015.org/proceedings/EMNLP/pdf/EMNLP199.pdf and compare it with a DBN proposed in http://mi.eng.cam.ac.uk/~ky219/papers/mairesse-acl10.pdf.

The code base includes our original implementation of the DBN (bagel folder), while we use the implementation of Wen et al (with minor modifications in the loading of the data) for the LSTM network (RNNLG folder).

For the comparison of the two networks, we use the BAGEL dataset after some post processing (data wrangling folder).

A detailed approach, results, insights and conlcusions are included the Report.pdf file.