Tourism Dialogue Parsing Tool

Introduction

Toursim dialogue system not only needs to chat with the user and but also understand and extract the entities from dialogues for making travel plan. This repository provides a tool to automatically extract the entities critical to making travel plan such as the travel destination and activities.

Tool Usage

We provide an web page to interact with the entity extraction model. To use this interface, you need clone this repository

git clone [email protected]:mynlp/TourismDialogue.git

Then start the web page server with the following command.

bash run_server.sh

The web page could be accessed with the link

http://127.0.0.1:9200

In addition, we also provide script to extract entities in a command line. To extract entities from your own data, you could replace the test file with your own in test_seq2seq3 of run.sh and run the following command.

bash run.sh test

To train the model, you need prepare the dataset by splitting the dataset into train, dev, and test set.

bash run.sh split_dataset

Then run the extract2.py script to generate the training instances.

cd dataset
bash run.sh create_dataset

Finally train the model by running the train.py

bash run.sh train

To train the model on customized dataset, you need prepare the dataset in the following format. each utterance within the dialogue should be like

{
  "utterance": "えーとー、ちょっと紅葉が綺麗な所に行きたいんですけども。",
  "speaker": "customer",
  "annotation": [{
    "query": {
      "大ジャンル": "遊ぶ",
      "キーワード": "紅葉"
    }
  }]
}

Performance

The performance of the model is evaluated on the held out test set of the toruism dialogue dataset. Three foundation models are evaluated and two different decoding strategies are tested. The results are exhibited in the following table.