generalisation-example

This is the bot we used for the experiments in our paper on "Few-shot
Generalisation Across Dialogue Tasks".

Installation

To install all the requirements needed to use this bot, please run:

pip install -r requirements.txt

Running the bot and experiments

To train any of the individual models, run one of:

• The embedding policy:

  make train-redp

• The keras policy with a binary state featurizer

  make train-lstm-bin

• The keras policy with a label tokenizer state featurizer:

  make train-lstm-feat

To run the bot on the commandline, run:

make run

To train the comparison models for the experiments from our paper, run:

make train-compare

To evaluate these models, run:

make evaluate-compare

Parameters used

Data:

• exclusion percentages: [0, 5, 25, 50, 70, 90, 95, 100]
• augmentation_factor: 0
• runs: 5

Embedding policy:

• epochs: 2000
• attn_shift_range: 5
• embed_dim: 20
• both attentions
• rnn size: 64
• everthing else default

Keras policy:

• binary featurizer/ label token featurizer
• batch size: 32
• rnn size: 64
• epochs: 400
• max history: 38

Overview of the files

data/core/train/ - training data from the hotel + restaurant domain
data/core/test - test data from the hotel domain
services - dummy API services for hotel/restaurant recommendation
actions.py - actions file containing the hotel/restaurant search actions