Skip to content

Latest commit

 

History

History
96 lines (59 loc) · 3.34 KB

README.md

File metadata and controls

96 lines (59 loc) · 3.34 KB
Raw

Response-Reranking

Code repository for the paper:

Reranking Overgenerated Responses for End-to-End Task-Oriented Dialogue Systems by Songbo Hu, Ivan Vulić, Fangyu Liu, and Anna Korhonen.

This response reranker is a simple yet effective model which aims to select high-quality items from the lists of responses initially over-generated by any end-to-end task-oriented dialogue system.

Environment

The code is tested with python 3.8. Firstly, install Pytorch 1.11.0 from the official website. Then, clone this repository and install the dependencies:

>> git clone [email protected]:cambridgeltl/response_reranking.git
>> pip install -r requirements.txt

Data Preprocessing

Before training and evaluating our reranking models, unzip data.zip in the repository root directory. It contains three files and a folder: 0.7_train.json, 0.7_dev.json, 0.7_test.json, and multi-woz-processed.

>> unzip data.zip

Each JSON file contains overgenerated responses from the MinTL System. It is a list of candidate response pairs with the following fields:

  • "context_text" denotes the lexicalised dialogue context.
  • "resp_text" and "resp_nodelex" are the ground truth delexicalised/lexicalised responses to the given dialogue context.
  • "resp_gen" is the generated delexicalised response based on greedy search given the dialogue context.
  • "over_gen" is a list of 20 overgenerated delexicalised responses based on top-p sampling given the given dialogue context.

We used the preprocess script (setup.sh) from DAMD to perform delexicalisation and produce files in multi-woz-processed.

Generating Similarity Scores

Generating the cosine similarity scores with the all-mpnet-v2 encoder between the overgenerated responses and the ground truth responses:

>> PYTHONPATH=$(pwd) python ./src/generate_similarity_scores.py

Generating the cosine similarity scores with the all-mpnet-v2 encoder between the greedy search responses and the ground truth responses:

>> PYTHONPATH=$(pwd) python ./src/generate_similarity_scores_greedy.py

Training

For stage 1: response selection training:

>> PYTHONPATH=$(pwd) python ./src/train_response_selection_cross_encoder.py

For stage 2: similarity-based response reranking training:

>> PYTHONPATH=$(pwd) python ./src/train_similarity_reranking.py

For stage 2: classification-based response reranking training:

>> PYTHONPATH=$(pwd) python ./src/train_classification_reranking.py

Testing

For testing the similarity-based response reranking models:

>> PYTHONPATH=$(pwd) python ./src/eval_similarity_reranking.py

For testing the classification-based response reranking models:

>> PYTHONPATH=$(pwd) python ./src/eval_classification_reranking.py