Semi-supervised latent variable models for sentence-level sentiment analysis

Auteur-es

Täckström, Oscar; McDonald, Ryan

Nombre Auteurs

2

Titre

Semi-supervised latent variable models for sentence-level sentiment analysis

Année de publication

2011

Référence (APA)

Täckström, O., & McDonald, R. (2011). Semi-supervised latent variable models for sentence-level sentiment analysis. Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers - Volume 2, 569‑574. https://aclanthology.org/P11-2100

Mots-clés

ND

URL

https://aclanthology.org/P11-2100

Accessibilité de l'article

Open access

Champ

Natural Language Processing

Type contenu (théorique Applicative méthodologique)

Applicative

Méthode

Creation of two variants of a semi-supervised model for fine sentiment analysis.
For the following experiments, we used the same data set and a comparable experimental setup to that of Tackström and McDonald (2011).3 We compare the two proposed hybrid models (Cascaded and Interpolated) to the fully supervised model of McDonald et al. (2007) (FineToCoarse) as well as to the soft variant of the coarsely supervised model of Tackström and McDonald (2011) (Coarse).

Cas d'usage

ND

Objectifs de l'article

In this paper, we demonstrate how combining coarse-grained and fine-grained supervision benefits sentence-level sentiment analysis. coarsely supervised model with a fully supervised model.
The proposed model is a fusion of a fully supervised structured conditional model and its partially supervised counterpart. This allows for highly efficient estimation and inference algorithms with rich feature definitions.

Question(s) de recherche/Hypothèses/conclusion

Research question(s) : These approaches all rely on the availability of fine-grained annotations, but T ̈ackström and McDonald (2011) showed that latent variables can be used to learn fine-grained sentiment using only coarse-grained supervision. While this model was shown to beat a set of natural baselines with quite a wide margin, it has its shortcomings. Most notably, due to the loose constraints provided by the coarse supervision, it tends to only predict the two dominant fine-grained sentiment categories well for each document sentiment category, so that almost all sentences in positive documents are deemed positive or neutral, and vice versa for negative documents.

Hypothesis(es) : As a way of overcoming these shortcomings, we propose to fuse a coarsely supervised model with a fully supervised model.
The proposed model is a fusion of a fully supervised structured conditional model and its partially supervised counterpart. This allows for highly efficient estimation and inference algorithms with rich feature definitions. We describe the two variants as well as their component models and verify experimentally that both variants give significantly improved results for sentence-level sentiment analysis compared to all baselines

Conclusion(s) : We introduced two simple, yet effective, methods of combining fully labeled and coarsely labeled data for sentence-level sentiment analysis. First, a cascaded approach where a coarsely supervised model is used to generate features for a fully supervised model. Second, an interpolated model that directly optimizes a combination of joint and marginal likelihood functions. Both proposed models are structured conditional models that allow for rich overlapping features, while maintaining highly efficient exact inference and robust estimation properties. Empirically, the interpolated model is superior to the other investigated models, but with sufficient amounts of coarsely labeled and fully labeled data, the cascaded approach is competitive.

Cadre théorique/Auteur.es

Sentence-level sentiment analysis (Pang et Lee, 2008)

Fine-grained supervision (Wiebe et al., 2005)

Exploiting document structure for sentiment analysis (Pang et Lee, 2004 ; McDonald et al., 2007 ; Yessenalina et al., 2010 ; Nakagawa et al., 2010 ; Sauper et al., 2010 ; Tackström et McDonald, 2011 ; Sauper et al., 2010)

Methods for blending discriminative and generative models (Lasserre et al., 2006 ; Suzuki et al., 2007 ; Agarwal et Daumé, 2009 ; Sauper et al., 2010)

Concepts clés

Supervised learning

Sentiment analysis

Données collectées (type source)

For our experiments we constructed a large balanced corpus of consumer reviews from a range of domains.
A training set was created by sampling reviews from five different domains: books, dvds, electronics, music and videogames. Document sentiment labels were obtained by labeling one and two star reviews as negative (NEG), three star reviews as neutral (NEU), and four and five star reviews as positive (POS).

Définition des émotions

No definition

Ampleur expérimentation (volume de comptes)

To assess the impact of fully labeled versus coarsely labeled data, we took stratified samples without replacement, of sizes 60, 120, and 240 reviews, from the fully labeled folds and of sizes 15,000 and 143,580 reviews from the coarsely labeled data. On average each review consists of ten sentences.

Technologies associées

Semi-supervised model

Fully supervised model

Partially supervised model

Mention de l'éthique

ND

Finalité communicationnelle

In this paper, we demonstrate how combining coarse-grained and fine-grained supervision benefits sentence-level sentiment analysis – an important task in the field of opinion classification and retrieval (Pang and Lee, 2008).

Résumé

We derive two variants of a semi-supervised model for fine-grained sentiment analysis. Both models leverage abundant natural supervision in the form of review ratings, as well as a small amount of manually crafted sentence labels, to learn sentence-level sentiment classifiers. The proposed model is a fusion of a fully supervised structured conditional model and its partially supervised counterpart. This allows for highly efficient estimation and inference algorithms with rich feature definitions. We describe the two variants as well as their component models and verify experimentally that both variants give significantly improved results for sentence-level sentiment analysis compared to all baselines.