This repository contains the data of the SIGIR 2022 Paper Personalized Abstractive Opinion Tagging.

This Github shows more details of this paper.

If you have any question, please open an issue or contact my email [email protected].

PATag is a large-scale Chinese restaurant reviews, opinion tags and user behavior dataset for personalized abstractive opinion tagging, which contains 555,297 reviews, 135,586 opinion tags and 15 million behavior actions from 68,732 users and 58,643 products(POIs). Each opinion tag is manually annotated to 28 predefined aspect categories, such as service, environment, taste and so on.

Intuitively, PATag mainly contains the following fields and sample_for_data_field.txt is an instance of PATag.

• review_id: Anonymized review id.
• user_id: Anonymized user id.
• poi_id: Anonymized product(POI) id.
• review_context: Historical reviews of user and product(POI).
• opinion_tags: Ranked aspect-opinion tag pairs.
• is_click: Whether user has clicked product(POI).
• is_order: Whether user has ordered product(POI).
• is_favor: Whether user has favored product(POI).

For ease of analysis and processing, we organize the data into multiple npy files. You can download them by Google Drive and process them with our provided data_reader.py and data_loader.py, then you have the data that POT needed.

For each user, we divide his/her reviews into historical reviews and recent reviews. There are 419,711 historical reviews and 135,586 recent reviews in PATag. We only use meaningful sentences in reviews as input, i.e. review sentences that can correspond to predefined aspects. As multiple review sentences may link to the same aspect in reviews, we only select the one with the highest confidence as the opinion tag. Briefly, we use meaningful sentences from historical reviews as input, and output the meaningful sentences with the highest confidence for a single aspect in recent reviews, i.e., opinion tag.

• UGC_Aspect_Word.npy : A dict with historical reviews id (int) as key and meaningful sentences (list) as value.
• UGC_RankedAspects.npy : A dict with recent review id (int) as key and ranked aspects (list) as value.
• UGC_RankedTags.npy : A dict with recent review id (int) as key and ranked opinion tags (list) as value.

• POI_User.npy : A dict with products(POIs) id (int) as key and users who have written reviews for it (list) as value.
• UGC_UP.npy : A dict with recent review id (int) as key and a tuple of user and product(POI) corresponding to it (tuple) as value.
• User_POI_rank.npy : A dict with user id (int) as key and products(POIs) which he/she has written reviews on (list) as value. Note that the products are sorted by its written time.
• User_PR.npy : A dict with user id (int) as key and a tuple of his/her historical reviews and corresponding products(POIs) (tuple list) as value.

• Click_UP.npy : A dict with user id (int) as key and a tuple list of products(POIs) his/her has clicked and the number of clicks (tuple list) as value.
• Buy_UP.npy : A dict with user id (int) as key and a tuple list of products(POIs) his/her has ordered and the number of orders (tuple list) as value.
• Favor_UP.npy : A dict with user id (int) as key and a tuple list of products(POIs) his/her has favored and the number of favors (tuple list) as value.

If you find our code useful, please cite our paper as follows:

@article{zhao2022personalized,
  title={Personalized Abstractive Opinion Tagging},
  author={Zhao, Mengxue and Yang, Yang and Li, Miao and Wang, Jingang and Wu, Wei and Ren, Pengjie and de Rijke, Maarten and Ren, Zhaochun},
  year={2022}
}