Beyond Single Items: Exploring User Preferences in Item Sets with the Conversational Playlist Curation Dataset
This is the work of Arun Tejasvi Chaganty, Megan Leszczynski, Shu Zhang, Ravi Ganti, Krisztian Balog and Filip Radlinski from Google LLC, made available under the Creative Commons Attribution 4.0 License. A full copy of the license can be found at https://creativecommons.org/licenses/by/4.0/
Users in consumption domains, like music, are often able to more efficiently provide preferences over a set of items (e.g. a playlist or radio) than over single items (e.g. songs). Unfortunately, this is an underexplored area of research, with most existing recommendation systems limited to understanding preferences over single items. Curating an item set exponentiates the search space that recommender systems must consider (all subsets of items!): this motivates conversational approaches---where users explicitly state or refine their preferences and systems elicit preferences in natural language---as an efficient way to understand user needs. We call this task conversational item set curation and present a novel data collection methodology that efficiently collects realistic preferences about item sets in a conversational setting by observing both item-level and set-level feedback. We apply this methodology to music recommendation to build the Conversational Playlist Curation Dataset (CPCD), where we show that it leads raters to express preferences that would not be otherwise expressed. Finally, we propose a wide range of conversational retrieval models as baselines for this task and evaluate them on the dataset.
Read the paper on arXiv.
The dataset consists of 917 conversations split into a development set (450 conversations) and test set (467 conversations). We also provide canonical train and validation splits of the development set, but encourage using k-fold cross validation given the relatively small size.
Each line of the dataset represents a single conversation in JSON format consisting of the following fields:
id(string): a unique identifier for this conversation.turns(list[Turn]): A list of turns in the conversation.- tracks` (dict[string, Track]): The metadata associated with each track referenced in turns above.
goal_playlist(list[string]): The list of track ids for the final target list of “liked” tracks in this conversation.
Each turn consists of:
user_query(string): user query for this turn.system_response(string): wizard response for this turn.search_queries(list[string]): list of queries entered by the wizard this turn.search_results(list[list[string]]): list of track ids retrieved for each query in “search_queries”.liked_results(list[string]): list of track ids shown to users and liked by them.disliked_results(list[string]): list of track ids shown to users and disliked by them.
Each track consists of:
track_ids(string): identifier for the track. It corresponds to a YouTube video id.track_titles(string): title of the track.track_artists(list[string]): names of the artist on this track.track_release_titles(string): title of the release or album.track_cluster_ids(string): identifier for the cluster this track belongs to. Tracks with approximately the same title and artists were clustered together. The evaluation script respects these clusters when computing metrics.track_canonical_ids(string): the canonical track id for the cluster this track belongs to.
Here is an abbreviated example:
{
"id": "00079c9c8dd7b55a",
"turns": [
{
"user_query": "I'd like to create a playlist that is perfect for a dance party. I'd like to create a playlist that I can utilize for a dance party.",
"system_response": "Hello! I can help with that! What kind of genre or artist would you like to add to your playlist?",
"search_queries": [],
"search_results": [],
"liked_results": [],
"disliked_results": []
},
{
"user_query": "Any genre works as long as it is upbeat and something people can jump around and sing to!",
"system_response": "Here is a selection of songs that match your preferences. Any other artist or genre?",
"search_queries": [
"I Gotta Feeling",
"Uptown Funk",
"Hey Ya"
],
"search_results": [
[
"OIPmhkzN2ug"
],
[
"tYvFa2ARD24"
],
[
"Jx_O6PHdWww",
"NDXCJYxkyio"
]
],
"liked_results": [
"Jx_O6PHdWww",
"OIPmhkzN2ug"
],
"disliked_results": [
"tYvFa2ARD24"
]
}
],
"tracks": {
"NDXCJYxkyio": {
"track_ids": "NDXCJYxkyio",
"track_titles": "Hey Ya",
"track_release_titles": "Circus",
"track_artists": [
"Sean Paul"
],
"track_canonical_ids": "NDXCJYxkyio",
"track_cluster_ids": "f971a7673a776078fed9b62f5669f848"
},
"OIPmhkzN2ug": {
"track_ids": "OIPmhkzN2ug",
"track_titles": "I Gotta Feeling",
"track_release_titles": "THE E.N.D. (THE ENERGY NEVER DIES)",
"track_artists": [
"The Black Eyed Peas"
],
"track_canonical_ids": "YU7IywQ_adI",
"track_cluster_ids": "3ad671cb3f4304d2c2a2927f2748de12"
},
"tYvFa2ARD24": {
"track_ids": "tYvFa2ARD24",
"track_titles": "Uptown Funk",
"track_release_titles": "Uptown Special",
"track_artists": [
"Mark Ronson",
"Bruno Mars"
],
"track_canonical_ids": "_vAM53xoLvo",
"track_cluster_ids": "c5877449f4d33ac521d25361d952bf8b"
},
"Jx_O6PHdWww": {
"track_ids": "Jx_O6PHdWww",
"track_titles": "Hey Ya! (Radio Mix)",
"track_release_titles": "Speakerboxxx/The Love Below",
"track_artists": [
"Outkast"
],
"track_canonical_ids": "jyyt0T-4dc4",
"track_cluster_ids": "bdc91aeae488a51e31cabf6285726fc9"
}
},
"goal_playlist": [
"Jx_O6PHdWww",
"OIPmhkzN2ug"
]
}
This repository also includes a standardized evaluation script that discounts near-duplicate tracks in the evaluation metrics.
To use this script, format the model output in JSONL as follows (see
model_outputs/ for examples):
{
"docid": "<dialog_id>:<turn_index>",
"neighbor": [
{ "docid": "<track_id>" }
...
]
}
You can then run the evaluation code as follows:
python3 eval.py --model_output <model_output.jsonl> --output output.csv