[Paper Review] An Automatic Contextual Analysis and Clustering Classifiers Ensemble approach to Sentiment Analysis

Murtadha Talib AL-Sharuee, Fei Liu+1 more|arXiv (Cornell University)|May 29, 2017

Sentiment Analysis and Opinion Mining29 references3 citations

TL;DR

This paper proposes ACACCE, an unsupervised, domain-independent sentiment analysis method that combines automatic contextual analysis with an ensemble of modified k-means classifiers. By leveraging SentWordNet for polarity-based centroid initialization and applying multiple term weighting schemes with majority voting, the approach achieves high accuracy, stability, and generalization across diverse datasets without human-labeled data or training.

ABSTRACT

Products reviews are one of the major resources to determine the public sentiment. The existing literature on reviews sentiment analysis mainly utilizes supervised paradigm, which needs labeled data to be trained on and suffers from domain-dependency. This article addresses these issues by describes a completely automatic approach for sentiment analysis based on unsupervised ensemble learning. The method consists of two phases. The first phase is contextual analysis, which has five processes, namely (1) data preparation; (2) spelling correction; (3) intensifier handling; (4) negation handling and (5) contrast handling. The second phase comprises the unsupervised learning approach, which is an ensemble of clustering classifiers using a majority voting mechanism with different weight schemes. The base classifier of the ensemble method is a modified k-means algorithm. The base classifier is modified by extracting initial centroids from the feature set via using SentWordNet (SWN). We also introduce new sentiment analysis problems of Australian airlines and home builders which offer potential benchmark problems in the sentiment analysis field. Our experiments on datasets from different domains show that contextual analysis and the ensemble phases improve the clustering performance in term of accuracy, stability and generalization ability.

Motivation & Objective

• To address the limitations of supervised sentiment analysis, including high annotation cost and domain dependency.
• To develop a completely automatic, unsupervised method that eliminates the need for human-labeled training data.
• To improve clustering stability and accuracy in sentiment analysis by integrating linguistic contextual processing with ensemble learning.
• To introduce new benchmark datasets for Australian airlines and home builders in sentiment analysis research.
• To enhance generalization and robustness across diverse domains through multi-weighting scheme integration and non-stochastic centroid initialization.

Proposed method

• The method employs a two-phase architecture: (1) contextual analysis to preprocess text using linguistic rules for negation, intensifiers, contrast, and spelling correction.
• Spelling correction is applied to normalize text, and sentiment lexicon SentWordNet 3.0 is used to assign polarity to adjectives and adverbs.
• The k-means base classifier is modified by initializing centroids using pre-clustered positive and negative features from SentWordNet, ensuring non-random, stable initialization.
• Multiple vector space models are constructed using different term weighting schemes (e.g., TF-IDF, binary, frequency) to represent documents.
• An ensemble of these k-means classifiers is formed using a majority voting mechanism with adaptive weight schemes to improve classification reliability.
• A seed assembly strategy is used to interpret cluster outputs, enhancing group judgment and interpretability of results.

Experimental results

Research questions

• RQ1 Can an unsupervised, fully automatic sentiment analysis method achieve competitive accuracy without requiring labeled training data?
• RQ2 How does contextual analysis—handling negation, intensifiers, and contrast—impact clustering performance in sentiment classification?
• RQ3 To what extent does ensemble learning with multiple weighting schemes improve stability and generalization in unsupervised sentiment clustering?
• RQ4 Can the proposed method maintain consistent performance across diverse domains, such as product reviews, airlines, and home builders?
• RQ5 How does non-stochastic centroid initialization affect the stability and reproducibility of k-means in sentiment clustering?

Key findings

• ACACCE achieves higher accuracy than baseline methods on the Kitchen dataset, demonstrating superior performance in challenging domains.
• The method shows stable performance across multiple datasets, with consistent results across repeated runs due to non-stochastic centroid initialization.
• Contextual analysis significantly improves clustering accuracy by correctly handling sentiment modifiers such as negation and intensifiers.
• The ensemble of classifiers with diverse weighting schemes enhances generalization, particularly when using global term weighting over document-specific weighting.
• The use of SentWordNet to pre-cluster features into positive, negative, and neutral groups enables effective and interpretable initialization of k-means centroids.
• The method is robust and domain-independent, with no reliance on labeled data, making it suitable for large-scale, real-time sentiment analysis applications.