# Multi-Instance Multi-Label Learning Networks for Aspect-Category Sentiment Analysis

**Authors:** Yuncong Li, Cunxiang Yin, Sheng-hua Zhong, Xu Pan  
**arXiv:** [2010.02656](https://arxiv.org/abs/2010.02656)  
**Date:** 2020-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.02656v1/

## Key points

- The paper proposes a new approach to aspect-category sentiment analysis (ACSA).
- This approach considers the sentiments of words indicating an aspect category in a sentence.
- Existing methods generate an aspect category-specific sentence representation and predict sentiment polarity based on it, ignoring the fact that the sentiment of an aspect category mentioned in a sentence is an aggregation of the sentiments of its corresponding words.
- The authors introduce AC-MIMLLN, which treats sentences as bags, words as instances, and the words indicating an aspect category as key instances.
- The model predicts the sentiments of all instances and then identifies key instances for each aspect category before aggregating their sentiments to obtain overall sentiment towards each aspect category.
- Experimental results on three public datasets demonstrate that AC-MIMLLN outperforms existing state-of-the-art models in terms of accuracy and F1-score.
- The proposed method has potential applications in various domains such as product reviews, social media analysis, and customer feedback analysis.

## Summary

The paper titled "Multi-Instance Multi-Label Learning Networks for Aspect-Category Sentiment Analysis" proposes a novel approach to aspect-category sentiment analysis (ACSA) that improves upon previous methods by considering the sentiments of words indicating an aspect category in a sentence. Most existing methods generate an aspect category-specific sentence representation and predict the sentiment polarity based on it, ignoring the fact that the sentiment of an aspect category mentioned in a sentence is an aggregation of the sentiments of its corresponding words. To address this limitation, the authors introduce AC-MIMLLN, which treats sentences as bags, words as instances, and the words indicating an aspect category as key instances. The model first predicts the sentiments of all instances and then identifies key instances for each aspect category before aggregating their sentiments to obtain overall sentiment towards each aspect category. Experimental results on three public datasets demonstrate that AC-MIMLLN outperforms existing state-of-the-art models in terms of accuracy and F1-score. The proposed method has potential applications in various domains such as product reviews, social media analysis and customer feedback analysis.