# Ugan

**Repository Path**: ices-e-sbot/ugan

## Basic Information

- **Project Name**: Ugan
- **Description**: Codes are publicly available for TSC paper.
- **Primary Language**: Python
- **License**: MIT
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-03-25
- **Last Updated**: 2022-08-16

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Ugan

#### Introduction
Codes are publicly available for TSC Manuscript: 
Utterances Generation and API Matching Framework based on Adversarial Learning for Service Bots

#### Architecture
For the generation task, this paper proposes an **U**tterances **G**eneration&Matching Framework (Ugan) based on **a**dversarial trai**n**ing, as shown in Fig 1.

<table>
  <tr>
    <td><img src="Egan.png" alt="the architecture of Egan"></td>
  </tr>
  <tr>
    <td align="center">Fig1: The architecture of Egan</td>
  </tr>
</table>

<!--  
For the matching task, this paper proposes an **E**xpression-API **M**atch **Mo**odel (Emmo) based on Siamese network, as shown in Fig 2.

<table>
  <tr>
    <td><img src="Emmo.png" alt="the architecture of Emmo"></td>
  </tr>
  <tr>
    <td align="center">Fig2: The architecture of Emmo</td>
  </tr>
</table>
-->

#### Results
The generation performance of different models as shown in Fig 3.

<table>
  <caption>Fig 3: The performance of different generation models</caption>
  <tr>
    <td><img src="visualization/url_bleu.png"></td>
    <td><img src="visualization/url_gleu.png"></td>
    <td><img src="visualization/url_chrf.png"></td>
  </tr>
  <tr>
    <td align="center">BLEU</td>
    <td align="center">GLEU</td>
    <td align="center">CHRF</td>
  </tr>
  <tr>
    <td align="center" colspan="3">All models only take url as input</td>
  </tr>
    
  <tr>
    <td><img src="visualization/all_bleu.png"></td>
    <td><img src="visualization/all_gleu.png"></td>
    <td><img src="visualization/all_chrf.png"></td>
  </tr>
  <tr>
    <td align="center">BLEU</td>
    <td align="center">GLEU</td>
    <td align="center">CHRF</td>
  </tr>
  <tr>
    <td align="center" colspan="3">All models both take url and description as input</td>
  </tr>
</table>

For the ablation experiment, the genereted result files are placed in './result/xxx.json', where xxx represents the training strategy and the model name.