TL;DR
This paper introduces a novel method for automatically extracting opinion words and their sentiment polarities from text, outperforming existing state-of-the-art techniques. By combining pre-trained word vectors with matrix factorization, it effectively captures nuanced semantic relationships between words, leading to more accurate and domain-adaptable opinion lexicons.
What Problem Does This Paper Address?
Sentiment analysis, a crucial task in understanding public opinion and user feedback, relies heavily on opinion lexicons – lists of words with associated sentiment polarities (positive or negative). However, creating and maintaining these lexicons is challenging because the sentiment of a word can vary significantly depending on the domain. For example, "unpredictable" might be positive in a movie review but negative in a car review. Existing methods often struggle with this domain dependency. Dictionary-based approaches are too general, while corpus-based methods that rely on seed words are sensitive to the quality of those seeds and may not fully consider the semantic relationships between words within a document. This paper aims to address these limitations by developing a more robust and domain-aware opinion lexicon extraction method.
Methods
The proposed method, Word Vector and Matrix Factorization (WVMF), builds upon previous work like CONR but introduces two key improvements. Firstly, instead of relying solely on pointwise mutual information (PMI) to capture word relationships, which can be sparse, WVMF leverages pre-trained word vectors (trained on Google News) for a more comprehensive similarity measurement. These word vectors encode rich semantic information. Secondly, WVMF incorporates additional features, such as inverse document frequency (IDF), to better assess a word's sentiment contribution to a document.
The core of the method involves defining a sentiment contribution matrix (C) that captures how much each word contributes to the sentiment of each document. This matrix is learned through a matrix factorization process, aiming to minimize the difference between the actual document-word relationships (R) and the reconstructed relationships from latent document (U) and word (V) feature matrices. The objective function also includes regularization terms to encourage meaningful latent representations and capture word similarities. Once the latent representations are learned, the sentiment polarity of each word is derived by comparing its contribution to positive versus negative documents.
Datasets
The authors evaluated their method on three publicly available datasets:
- IMDB: Movie reviews.
- Movie reviews: Another collection of movie reviews.
- DVD reviews: Reviews related to DVDs.
The opinion words used for testing were obtained from the MPQA (Multi-Perspective Question Answering) corpus.
Key Insights and Ideas
- Pre-trained word vectors are powerful: Using pre-trained word vectors significantly enhances the ability to capture semantic relationships between words, overcoming the sparsity issues of traditional methods like PMI.
- Domain-specific sentiment is crucial: The paper highlights the necessity of methods that can adapt to domain-specific sentiment, as universal lexicons are insufficient.
- Matrix factorization is effective for sentiment: The matrix factorization framework provides a robust way to model the complex interactions between documents, words, and their sentiment contributions.
- Combining multiple features improves performance: Incorporating features like IDF alongside word vector similarities leads to a more accurate sentiment contribution estimation.
Key Results
The experimental results demonstrate that the proposed WVMF method consistently outperforms baseline methods, including SO-PMI (seed word based), WEED (optimization based), SVD, NMF (matrix factorization based), and CONR (a strong prior method).
- Accuracy in Opinion Word Extraction: Figure 1 shows that WVMF achieves the highest accuracy in extracting opinion words across all tested domains (IMDB, Movie, DVD).
- Accuracy in Sentiment Polarity Identification: Table 1 presents a detailed comparison of sentiment polarity identification accuracy for the top k opinion words. WVMF consistently achieves the highest accuracy, for instance, achieving 96.33% accuracy for the top 10 opinion words on the IMDB dataset, compared to 93.83% for CONR. This superiority is observed across different dataset sizes (Top10, Top20, Top50, Top100, Top200).
Limitations and Future Work
Limitations: * The paper mentions that the similarity between words is measured by cosine distance with word vectors. While effective, this might not capture all forms of semantic relatedness. * The method relies on pre-trained word vectors, which themselves might have biases or limitations inherited from their training data.
Future Work: * The authors suggest that their method can be extended to other tasks such as aspect-based sentiment analysis. * Further exploration of different word vector representations or similarity measures could potentially yield even better results.
Summary updated