--- dataset_info: features: - name: id dtype: int64 - name: org_index dtype: int64 - name: data_source dtype: string - name: industry dtype: string - name: text dtype: string - name: labels sequence: sequence: string - name: label_codes dtype: string splits: - name: train num_bytes: 2599501.8469831664 num_examples: 7930 - name: validation num_bytes: 346490.977586533 num_examples: 1057 - name: test num_bytes: 520228.17543030076 num_examples: 1587 download_size: 1010316 dataset_size: 3466221.0 configs: - config_name: default data_files: - split: train path: data/train-* - split: validation path: data/validation-* - split: test path: data/test-* --- # FABSA, An aspect-based sentiment analysis dataset in the Customer Feedback space (Trustpilot, Google Play and Apple Store reviews). A professionally annotated dataset released by [Chattermill AI](https://chattermill.com/), with 8 years of experience in leveraging advanced ML analytics in the customer feedback space for high-profile clients such as Amazon and Uber. Two annotators possess extensive experience in developing human-labeled ABSA datasets for commercial companies, while the third annotator holds a PhD in computational linguistics. There has been a lack of high-quality ABSA datasets covering broad domains and addressing real-world applications. Academic progress has been confined to benchmarking on domain-specific, toy datasets such as restaurants and laptops, which are limited in size (e.g., [SemEval Task ABSA](https://aclanthology.org/S16-1002.pdf) or [SentiHood](https://aclanthology.org/C16-1146/)). This dataset is part of the [FABSA paper](https://www.sciencedirect.com/science/article/pii/S0925231223009906), and we release it hoping to advance academic progress as tools for ingesting and analyzing customer feedback at scale improve significantly, yet evaluation datasets continue to lag. FABSA is a new, large-scale, multi-domain ABSA dataset of feedback reviews, consisting of approximately 10,500 reviews spanning 10 domains (Fashion, Consulting, Travel Booking, Ride-hailing, Banking, Trading, Streaming, Price Comparison, Information Technology, and Groceries). ## Task This dataset encompasses **Aspect Category Sentiment Analysis** and is suitable for both **Aspect Category Detection** (ACD) and **Aspect Category Sentiment Classification** (ACSC). ACD in sentiment analysis identifies aspect categories mentioned in a sentence. These categories are conceptual; they may not explicitly appear in the review and are chosen from a predefined list of Aspect Categories. ACSC classifies the sentiment polarities of these conceptual aspect categories. The predefined list of Aspect Categories for this dataset are: ``` category category_code 0 Account management: Account access account-management.account-access 1 Company brand: Competitor company-brand.competitor 2 Company brand: General satisfaction company-brand.general-satisfaction 3 Company brand: Reviews company-brand.reviews 4 Logistics rides: Speed logistics-rides.speed 5 Online experience: App website online-experience.app-website 6 Purchase booking experience: Ease of use purchase-booking-experience.ease-of-use 7 Staff support: Attitude of staff staff-support.attitude-of-staff 8 Staff support: Email staff-support.email 9 Staff support: Phone staff-support.phone 10 Value: Discounts promotions value.discounts-promotions 11 Value: Price value for money value.price-value-for-money ``` ## Annotation Scheme The FABSA dataset is manually labeled according to a hierarchical annotation scheme that includes parent and child aspect categories. Each aspect category comes with an associated sentiment label (positive, negative, and neutral), creating a total of (12 × 3) target classification categories. In line with prior studies, we employ a multi-label classification scheme where each review is tagged with one or more aspect + sentiment labels. Thus, a single review can cover multiple aspects and express various (sometimes contrasting) polarities. For example, ``` Customer Review: “product is very good but customer service is really bad, they never respond” Labels: [Company brand: General satisfaction, positive), (Staff support: Attitude of staff, Negative)" ``` ![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/608995d619137b3a6ba76113/Lr60Oc6NGalkOEj-_Gtzk.jpeg) ## Citation ``` @article{KONTONATSIOS2023126867, title = {FABSA: An aspect-based sentiment analysis dataset of user reviews}, journal = {Neurocomputing}, volume = {562}, pages = {126867}, year = {2023}, issn = {0925-2312}, doi = {https://doi.org/10.1016/j.neucom.2023.126867}, url = {https://www.sciencedirect.com/science/article/pii/S0925231223009906}, author = {Georgios Kontonatsios and Jordan Clive and Georgia Harrison and Thomas Metcalfe and Patrycja Sliwiak and Hassan Tahir and Aji Ghose}, keywords = {ABSA, Multi-domain dataset, Deep learning}, abstract = {Aspect-based sentiment analysis (ABSA) aims at automatically extracting aspects of entities and classifying the polarity of each extracted aspect. The majority of available ABSA systems heavily rely on manually annotated datasets to train supervised machine learning models. However, the development of such manually curated datasets is a labour-intensive process and therefore existing ABSA datasets cover only a few domains and they are limited in size. In response, we present FABSA (Feedback ABSA), a new large-scale and multi-domain ABSA dataset of feedback reviews. FABSA consists of approximately 10,500 reviews which span across 10 domains. We conduct a number of experiments to evaluate the performance of state-of-the-art deep learning models when applied to the FABSA dataset. Our results demonstrate that ABSA models can generalise across different domains when trained on our FABSA dataset while the performance of the models is enhanced when using a larger training dataset. Our FABSA dataset is publicly available.11https://github.com/kontonag86/fabsa-dataset.} } ```