Brief of Project

This repository is the source code of the paper CharacterChat: Learning towards Conversational AI with Personalized Social Support (https://arxiv.org/abs/2308.10278) CharacterChat is a social support conversation system, consisting of a llama-based model to generate persona- and memory-based response, and a interpersonal matching model to dispatch a most compatible supporter agent from MBTI-1024 Bank for the help-seeker.

Getting Start

The overflow of this work is as follows: (1) we build the MBTI-1024 Bank (a group of virtual characters with complex persona and memory) by MBTI-based decomposition on ChatGPT (2) we collect the MBTI-S2Conv, a persona- and memory-based conversation dataset by improved role-playing prompting (3) we develop the CharaterChat, a social support conversation system, consisting of a llama-based model to generate persona- and memory-based response and a interpersonal matching model to dispatch a most compatible supporter agent from MBTI-1024 Bank for the help-seeker.

1. ChatGPT Data

In this part, you will get conversation records, role settings and other necessary information step by step from ChatGPT. For this, we designed a pipeline. The code and data that need to be used in this pipeline are placed in the src and data folders respectively.

1.1 OpenAI api

In order to use chatGPT, we called the interface of openai in the experiment. This means you better have several openai api keys (one is enough, but multiple can be faster). For details about the openai api key, please refer to the following link: https://platform.openai.com/docs/api-reference/introduction

After you get api keys, please copy them to the api_key.txt. Each api_key occupies one line, separated by \n.

1.2 Pipeline

1.2.1. get profiles

bash src/gen_profile.sh

Through this operation, you can obtain the settings of multiple roles. You can modify the prompt in the src/profile/make_data to get more customized role settings.

Input File: data/profile_input.json

Output File: data/profile_output.json

1.2.2. get profile transformed

bash src/gen_profile_trans.sh

After obtaining the character setting, further processing of the setting is required, such as obtaining second-person descriptions and obtaining additional information. Through this operation, a more complete role setting can be obtained.

Input File: data/profile_output.json

Output File: data/profile_trans_output.json

1.2.3. get conversations

bash gen_dial.sh

Now, the conversation records between characters can be fetched. The default setting is that each character talks to 10 random characters. You can modify src/dial/make_data.py to change it. Also, because conversations consume a lot of tokens, we recommend testing in small amounts before getting large amounts of data.

Input File: data/dial_input.json

Output File: data/dial_output.json

1.2.4. evaluate conversations

bash run_eval_dial.sh

After getting the conversations, we scored each conversation by 3 criteria. You can see src/query_data.py for details, and you can use custom evaluation criteria.

Input File: data/eval_dial_input.json

Output File: data/eval_dial_output.json

1.3 Notice

Although there are many restrictions in the prompt, as a generative language model, chatGPT does not guarantee to return ideal results. Therefore, please clean the data after getting it to ensure that the data is correct, complete and can be parsed.

In addition, due to the large amount of data, only part of the data is provided in this repository as an example of the data format. If you want to get full data of our experiment, you can view this link.

https://drive.google.com/drive/folders/15mBi-y08RL-GTKItIPAbAl13LV6SZzci?usp=drive_link

2. Model

After getting the dialogue data, we can train our model.

Model	Backbone	Dataset
model_base	LLaMA-2-7b	model_base_dataset.json
model_supporter	LLaMA-2-7b	model_supporter_dataset.json
model_seeker	LLaMA-2-7b	model_seeker_dataset.json
info_selecter	BERT-base-uncased	model_info_selecter_dataset.json
persona_score	BERT-base-uncased	model_persona_score_dataset.json

2.1 Backbone Models

Our model is mainly based on two models: LLaMA-2-7b and BERT-base-uncased. Therefore, first you need to download these two models.

After the download is complete, please put models in model/models/.

2.2. Train Model

2.2.1. model_base

bash model/train_model_base.sh

The training dataset for model_base includes ESconv and EmpatheticDialogues. model_base is the base model of model_supporter and model_seeker.

After training, merge the trained adapter and the base model to get the full model_base.

python model/merge_lora.py model/models/Llama-2-7b-hf model/output/model_base model/models/model_base

2.2.2. model_supporter

bash model/train_model_supporter.sh

The training dataset of model_supporter is processed conversation records from previous pipeline. It will talk with others like a supporter.

After training, merge the trained adapter and the base model to get the full model_supporter.

python model/merge_lora.py model/models/model_base model/output/model_supporter model/models/model_supporter

2.2.3. model_seeker

bash model/train_model_supporter.sh

Same as model_supporter, the training dataset of model_supporter is processed conversation records from previous pipeline. It will talk with others like a seeker.

After training, merge the trained adapter and the base model to get the full model_seeker.

python model/merge_lora.py model/models/model_base model/output/model_seeker model/models/model_seeker

2.2.4. info_selecter

bash model/train_model_info_selecter.sh

This model is used to select the currently used memory based on the conversation history. The training dataset for this model comes from conversation records. In each round of conversation, the selected memory is a positive sample, and the unselected memory is a negative sample.

2.2.5. persona_score

bash model/train_model_persona_score.sh

This model was used to assess the degree of personality fit between the two parties in the conversation. The training dataset of this model comes from the conversation evaluation results, average score of the three scoring criteria is used as matching score.

2.3. Notice

The above datasets and models are not included in this repository, you can download our datasets and models through links below.

DATASET

https://drive.google.com/drive/folders/1n5UlZ4vm3bSBZv6c4kIvIQJHFrSZgON3?usp=drive_link

MODEL

https://drive.google.com/drive/folders/1We80A24Cpz2WPd4hPKA2Mv8Ac4hop9nC?usp=drive_link https://drive.google.com/drive/folders/1gCXbmFUS0f6SnhWGmU-Y5Bt4l3AcR390?usp=drive_link

3. demo

3.1. chat_demo

bash model/run_chat_demo.sh

In this demo, you can chat with the model_supporter and model_seeker.

3.2. web_supporter

bash run_web_supporter.sh

A web page that allows you talk to model_supporter.

kou-bin/CharacterChat