Title: ACAT: A Collaborative Platform for Efficient Aspect-Based Sentiment Dataset Annotation

Abstract:

The development of robust Aspect-Based Sentiment Analysis (ABSA) models hinges on the availability of high-quality annotated datasets. Currently, researchers face significant inefficiencies because existing annotation tools output data as flat files. This limitation forces teams to manually merge inputs from multiple annotators, reconstruct relational structures, and calculate reliability metrics using bespoke scripts. To address these challenges, this study presents ACAT (Aspect-based sentiment analysis Collaborative Annotation Tool), a web-based environment designed natively to support four distinct ABSA workflows: (1) Aspect-Category Sentiment Analysis, (2) Clause-Level Segmentation, (3) Aspect-Term Sentiment Analysis featuring character-level position tracking, and (4) Aspect Sentiment Triplet Extraction that preserves dual span offsets. The primary innovation of ACAT is its integrated Extract, Transform, Load (ETL) pipeline, which automatically aligns collaborative annotations and generates Inter-Annotator Agreement (IAA) metrics at the point of export, thereby producing datasets ready for training. Preliminary validation involving 1,002 restaurant reviews and two annotators with varying levels of expertise demonstrated that ACAT achieves a median annotation time of 31.58 seconds, with raw IAA scores ranging from 0.78 to 0.86 across all tested tasks.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC