Nidhogg-zh/Multi-Questioner_Dialogue

Multi-Questioner Dialogue (MQDialog) Dataset

Dataset Details

Dataset Description

The Multi-Questioner Dialogue (MQDialog) dataset is designed to facilitate research in questioner-aware personalization. It contains dialogues with various questioners for each reponder. The dataset is derived from English and Chinese scripts of popular TV shows and real-world conversations. It includes dialogues where selected leading actors act as responders, while other characters or contacts serve as questioners. The dataset contains a diverse set of 12 responders and 173 questioners. The dataset supports research on dialogue generation, response evaluation, and questioner-aware personalization in multi-turn conversations.

Dataset Sources

• English scripts: The Big Bang Theory, Friends, and Modern Family.
• Chinese scripts: My Own Swordsman and Empresses in the Palace.
• Real-world conversations (WeChat): Records from a single user, focusing on two-person chats. (Not public, but you can extract the data using the code we provided)

Direct Use

The dataset is suitable for:

• Training and evaluating questioner-aware multi-turn dialogue systems.
• Studying personality-aligned response generation.
• Benchmarking the performance of dialogue models with multi-questioner setups.

Dataset Structure

• Responders: 12 leading actors from TV scripts and a single WeChat user.
• Questioners: 173 individuals interacting with the responders, the detailed information is listed in the Table.
• Splits: Randomly divided into training (3761 dialogues per responder on average) and testing (917 dialogues per responder on average).

Language	Data Source	# Questioners	Questioner Examples	Responder	# train	# test
English	The Big Bang Theory	14	Priya, Barry, Howard, Leonard, etc.	Sheldon	4805	1101
		12	Bernadette, Penny, Raj, Stuart, etc.	Leonard	4607	1014
	Friends	12	Amy, Chandler, Charlie, Joey, etc.	Rachel	3768	870
		20	Ben, Mike, Gary, Paul, etc.	Ross	3839	960
	Modern Family	9	Alex, Cameron, Dylan, Gloria, etc.	Claire	1161	281
		8	Haley, Jay, Luke, Mitchell, etc.	Phil	881	246
Chinese	My Own Swordsman	16	Bai Sanniang, Guo Furong, Mo Xiaobei, etc.	Tong Xiangyu	3200	831
		16	Bao Daren, Ji Wuming, Zhu Wushuang, etc.	Bai Zhantang	2995	857
		8	Li Dazui, Xing Butou, Yan Xiaoliu, etc.	Lv Xiucai	1635	409
	Empresses in the Palace	17	Cao Guiren, Mei Zhuang, Liu Zhu, etc.	Zhen Huan	1229	350
		11	Consort Hua, Empress, Huan Bi, etc.	Emperor	704	200
	WeChat Records	30	Author's contacts	Author	-	-

*Note: The last response from the responder serves as the ground truth, while preceding dialogues constitute the dialogue history. We provide a compact version of the training set because, during training, the answers within the dialogue history can be used to compute the loss, eliminating the need for the answer in the last sentence.

Data Files & Code

For each responder, dialogues with different questioners are stored in the corresponding folder, diags_two_role_{responder_name}. Intermediate results from data processing are also provided. The final datasets used for questioner-aware personalization are:

• {script_name}_diags_{responder_name}_{questioner_name}_{responder_name}_response_L512_dev.json
• {script_name}_diags_{responder_name}_{questioner_name}_{responder_name}_response_L512_train.json

Additionally, dialogues with different questioners are clustered based on query similarity. The clustering results are stored in the diags_two_role_{responder_name}_clustered folder.

We have provided the preprocessed raw data for each scripts named with {script_name}_dialgs.json. To extract dialogues for one responder, please run the python file extract_two_role_diag_{responder_name}.py under each subfolder.

Related functions:

• get_role_list(): get whole role name
• extract_diag_between_two_role(): extract and only reserve diags between two roles
• clean_diag(): remove duplicates, remove conversations with only one person, and remove empty values
• clean_diag_with_repeated(): remove conversations with only one person, and remove empty values
• split_train_and_dev(): split training set and validation set
• split_diag_with_sliding_window(): construct diags with limited length through a sliding window
• extract_diag_for_target_from_role_conv(): only reserve diags that the response is from target role

Data Instances

Below is an example from the dataset, it contains conversations between the target_role (i.e. responder) and the input_role (i.e. questioner).

{
    "id": "episode_14_chunk_6_index_0_part2_piece_0",
    "conversations": [
        {
            "from": "Bernadette",
            "value": "Did you hear? Isn’t it terrible?"
        },
        {
            "from": "Leonard",
            "value": "Have you seen him?"
        },
        {
            "from": "Bernadette",
            "value": "They wouldn’t let me in. Oh my Howie."
        },
        {
            "from": "Leonard",
            "value": "It’ll be okay. It’ll be okay."
        }
    ],
    "target_role": "Leonard",
    "target_role_short": "Leonard",
    "input_role": "Bernadette",
    "input_role_short": "Bernadette",
    "role_pair_id": 8,
    "cluster_id": 2 (Only in the clustered data)
}

Dataset Creation

Curation Rationale

MQDialog was created to address the need for a multilingual, multi-questioner dataset that reflects questioner-aware personalized response generation in diverse conversational contexts.

Data Collection and Processing

• Scripts: Extracted dialogues between a responder (leading actor) and questioners (other characters), ensuring a clean dataset by removing errors, repeated content, and irrelevant entries.
• Real-world records: Focused on one-on-one conversations, with new dialogue sessions defined by a time gap (e.g., 3 hours).
• Filtering: Questioners with fewer than 20 interactions were excluded to ensure meaningful analysis.

Recommendations

• Use the dataset in conjunction with other corpora to mitigate cultural or linguistic biases.
• Ensure responsible use of the data, particularly when training models for real-world applications.