A Joint Model of Text and Aspect Ratings for Sentiment Summarization

Auteur-es

Titov, Ivan; McDonald, Ryan

Nombre Auteurs

2

Titre

A Joint Model of Text and Aspect Ratings for Sentiment Summarization

Année de publication

2008

Référence (APA)

Titov, I., & McDonald, R. (2008). A Joint Model of Text and Aspect Ratings for Sentiment Summarization. Proceedings of ACL-08: HLT, 308‑316. https://aclanthology.org/P08-1036

Mots-clés

ND

URL

https://aclanthology.org/P08-1036

Accessibilité de l'article

Open access

Champ

Natural Language Processing

Type contenu (théorique Applicative méthodologique)

Applicative

Méthode

In this paper we presented a joint model of text and aspect ratings for extracting text to be displayed in sentiment summaries.
The model uses aspect ratings
We propose a statistical model which is able to discover corresponding topics in text and extract textual evidence from reviews.
We propose an unsupervised model that leverages aspect ratings that frequently accompany an online review.

Cas d'usage

Online customer reviews

Objectifs de l'article

We propose a statistical model which is able to discover corresponding topics in text and extract textual evidence from reviews supporting each of these aspect ratings – a fundamental problem in aspect-based sentiment summarization (Hu and Liu, 2004a).

Question(s) de recherche/Hypothèses/conclusion

Research question(s) : The most pressing challenge in an aspect-based summarization system is to extract all relevant mentions for each aspect [...]. When labeled data exists, this problem can be solved effectively using a wide variety of methods available for text classification and information extraction. However, labeled data is often hard to come by, especially when one considers all possible domains of products and services.

Hypothesis(es) : Instead, we propose an unsupervised model that leverages aspect ratings that frequently accompany an online review. In order to construct such model, we make two assumptions. First, ratable aspects normally represent coherent topics which can be potentially discovered from co-occurrence information in the text. Second, we hypothesize that the most predictive features of an aspect rating are features derived from the text segments discussing the corresponding aspect. Motivated by these observations, we construct a joint statistical model of text and sentiment ratings.

Conclusion(s) : Our model achieves high accuracy, without any explicitly labeled data except the user provided opinion ratings.
We demonstrated that the model indeed discovers corresponding coherent topics and achieves accuracy in sentence labeling comparable to a standard supervised model.

Cadre théorique/Auteur.es

Aspect-based sentiment summarization (Hu et Liu, 2004a ; Popescu et Etzioni, 2005 ; Gamon et al., 2005 ; Carenini et al., 2006 ; Zhuang et al., 2006)

Sentiment classification (Wiebe, 2000 ; Pang et al., 2002 ; Turney, 2002)

Aspect identification (Hu et Liu, 2004b ; Gamon et al, 2005 ; Titov et McDonald, 2008)

Text classification and information extraction (Manning et Schutze, 1999)

Summarizing sentiment by extracting and aggregating sentiment over ratable aspects (Hu et Liu, 2004a ; Popescu et Etzioni, 2005 ; Gamon et al, 2005 ; Mei et al., 2007 ; Titov et McDonald, 2008)

Joint sentiment and topic modeling (Blei et McAuliffe, 2008 ; Branavan et al., 2008)

Concepts clés

Sentiment analysis

Aspect Ratings

Données collectées (type source)

Qualitative evaluation : To perform qualitative experiments we used a set of reviews of hotels taken from TripAdvisor.com. Every review was rated with at least three aspects: service, location and rooms. Each rating is an integer from 1 to 5. The dataset was tokenized and sentence split automatically.

Quantitative evaluation : [...] We hand labeled random sentences from the dataset considered in the previous set of experiments. The sentences were labeled with one or more aspects, related to aspects service, location and rooms.

Définition des émotions

No definition

Ampleur expérimentation (volume de comptes)

Qualitative evaluation : 10,000 reviews of hotels (109,024 sentences, 2,145,313 words in total)

Quantitative evaluation : We hand labeled 779 random sentences with one or more aspects. Among them, 164, 176 and 263 sentences were labeled as related to aspects service, location and rooms, respectively. The remaining sentences were not relevant to any of the rated aspects.

Technologies associées

Multi-Aspect Sentiment model (MAS)

Multi-Grain Latent Dirichlet Allocation model (MG-LDA)

Mention de l'éthique

ND

Finalité communicationnelle

The proposed approach is general and can be used for segmentation in other applications where sequential data is accompanied with correlated signals.

The primary area of future work is to incorporate the model into an end-to-end sentiment summarization system in order to evaluate it at that level.

Résumé

Online reviews are often accompanied with numerical ratings provided by users for a set of service or product aspects. We propose a statistical model which is able to discover corresponding topics in text and extract textual evidence from reviews supporting each of these aspect ratings – a fundamental problem in aspect-based sentiment summarization (Hu and Liu, 2004a). Our model achieves high accuracy, without any explicitly labeled data except the user provided opinion ratings. The proposed approach is general and can be used for segmentation in other applications where sequential data is accompanied with correlated signals.