Source-linked AI summary
W2VLDA: Almost Unsupervised System for Aspect Based Sentiment Analysis
Aitor García-Pablos, Montse Cuadros, German Rigau
TL;DR
ABSA systems often depend on costly labelled data and language-specific resources for each domain and language. W2VLDA combines topic modelling with other unsupervised methods and minimal seed-word configuration to analyze unlabelled reviews. It performs aspect and sentiment classification across several languages and domains, while also separating aspect terms from opinion words.
Problem
Supervised ABSA generally requires manually labelled data and language-specific resources for each target domain and language, making broad coverage costly and time-consuming.
Method
W2VLDA combines topic modelling, continuous word embeddings, and a Maximum Entropy classifier, configured with unlabelled reviews and minimal seed words.
Results
W2VLDA performs aspect classification, sentiment polarity classification, and aspect-term/opinion-word separation across customer reviews in multiple languages and domains.
Takeaways & Limitations
The system can be applied across languages and domains with almost no adaptation and without language- or domain-specific resources.
Takeaways & Limitations
The approach requires a reasonable amount of domain documents to build its statistical model and assigns every vocabulary word to either the aspect-term or opinion-word class.
Abstract
from arXiv · showhide
With the increase of online customer opinions in specialised websites and social networks, the necessity of automatic systems to help to organise and classify customer reviews by domain-specific aspect/categories and sentiment polarity is more important than ever. Supervised approaches to Aspect Based Sentiment Analysis obtain good results for the domain/language their are trained on, but having manually labelled data for training supervised systems for all domains and languages are usually very costly and time consuming. In this work we describe W2VLDA, an almost unsupervised system based on topic modelling, that combined with some other unsupervised methods and a minimal configuration, performs aspect/category classifiation, aspect-terms/opinion-words separation and sentiment polarity classification for any given domain and language. We evaluate the performance of the aspect and sentiment classification in the multilingual SemEval 2016 task 5 (ABSA) dataset. We show competitive results for several languages (English, Spanish, French and Dutch) and domains (hotels, restaurants, electronic-devices).
1. Introduction
W2VLDA addresses the data and resource demands of supervised ABSA with an almost unsupervised, multilingual and multidomain approach. Using minimal seed-word configuration and unlabelled reviews, it jointly models aspects, opinion-word polarity, and sentences.
- Best-performing ABSA systems generally require manually labelled data and language-specific resources for a particular domain and language.
- Weakly supervised systems avoid labelled training data but usually require language-specific seed lists or preprocessing tools, and often report only English results.
- W2VLDA simultaneously classifies aspects, separates aspect-terms from opinion words, and classifies sentiment polarity at the word level.
- The system outputs weighted words for each aspect and polarity, together with weighted sentences for every selected aspect and polarity.
- The system combines topic modelling, continuous word embeddings, and a Maximum Entropy classifier to model customer reviews.
- W2VLDA uses unlabelled domain-language corpora and a minimal configuration of one seed word per desired aspect, plus positive and negative examples.
2. Related work
Related work covers supervised, lexicon-based, topic-modeling, embedding-based, and jointly modeled approaches to opinion and aspect analysis. W2VLDA combines minimal topic guidance with embeddings to perform three subtasks across languages and domains without pre-existing language- or domain-dependent resources.
- Prior opinion-mining research addresses subtasks including aspect detection and detecting opinions about evaluated product aspects.
- Common approaches identify frequent nouns, lexical patterns, and dependency relations, while other work derives word polarity through sentiment lexicons.
- Bootstrapping and polarity propagation typically leverage base dictionaries and pre-existing linguistic resources.
- LDA models documents as distributions over topics and words, supporting unsupervised text modeling, clustering, and extensions that jointly represent topics and sentiment.
- Topic-model outputs are anonymous word distributions, so mapping them to meaningful domain categories generally requires expert inspection or an existing resource.
- Word embeddings have also been used in supervised sentiment-analysis systems, while some topic models automatically separate aspect terms from opinion words.
- W2VLDA combines topic modelling, continuous word embeddings, and minimal topic definition to perform aspect classification, sentiment classification, and aspect-term/opinion-word separation together.
3. System description
W2VLDA performs aspect, opinion-word, and polarity classification with minimal seed-word configuration, combining topic modelling, embeddings, and unsupervised classification. Its LDA-based model separates word types and polarity while biasing topic and word distributions toward seed-defined aspects.
- W2VLDA performs aspect detection, opinion classification, and positive/negative polarity classification while separating opinion words from aspect terms.
- Topics and sentiment configuration: The configuration uses one seed per desired aspect plus one general positive and one general negative seed for each language and domain.Stopword lists from Apache Lucene are also used for better results.
- Aspect-term and opinion-word separation: Brown clusters from an unlabeled domain corpus provide context features for a MaxEnt classifier that separates aspect terms from opinion words.Seed words are treated as gold examples, bootstrapped from the corpus, represented through context clusters, and used to train the classifier.
- Aspect-term and opinion-word separation: Every vocabulary word is assigned to either the aspect-term or opinion-word class, because adding an “other” class would require manually labelled training instances.The method assumes ambiguous words will be distributed across both classes and lose relevance during topic modelling.
- Combining everything in a topic model: The core model extends LDA with aspect-term/opinion-word and positive/negative variables, using classifier probabilities and biased hyperparameters.Posterior inference uses Gibbs sampling to sample topic, word-type, and opinion-word polarity assignments.
- Combining everything in a topic model: Word embeddings bias topic and word hyperparameters through cosine similarity between vocabulary words and aspect seed words.The implementation uses word2vec representations; α controls document-topic distributions, β controls topic-word distributions, and δ controls document polarity distributions.
4. Evaluation
W2VLDA is evaluated across three subtasks, multiple languages and domains, using unlabelled corpora, seed configurations, and comparisons with topic-modelling and supervised baselines. Results show competitive topic and polarity classification, stable seed-word effects, and generally effective aspect-term/opinion-word separation.
- Resources and experimental setting: The evaluation covers topic classification, sentiment classification, and aspect-term/opinion-word separation across multilingual and domain-specific datasets.Experiments include SemEval2016 restaurant reviews and examples spanning English, Spanish, French, and multiple domains.
- Resources and experimental setting: The resulting topic distributions separate aspect terms, positive words, and negative words for English, Spanish, and French domains.Tables 2 and 3 report these distributions, while Table 4 gives high-probability sentence examples across topics, domains, and languages.
- Comparison with other LDA-based approaches: W2VLDA obtains slightly better overall results than other LDA-based systems without requiring human relabelling of the resulting topics.The comparison uses the proposed approach and alternative topic-modelling systems.
- Comparison with other LDA-based approaches: W2VLDA obtains comparable results on the small dataset and better results on the big dataset while using one seed word per polarity.The comparison is reported in Figure 5.
- Multilingual evaluation on SemEval2016: W2VLDA outperforms the baselines with 95% confidence for aspect classification in all languages except Dutch, where significance reaches 80%.Evaluation uses balanced subsets, five random samplings, 10-fold cross-validation, and statistical significance testing.
- Multilingual evaluation on SemEval2016: Polarity classification likewise outperforms the baselines with 95% confidence except for Dutch, where the reported confidence is lower.The paper suggests excessive oversampling or less representative unlabelled Dutch data as possible reasons and leaves minimum data requirements for future work.
- Assessing seed-word impact: Aspect classification remains stable across seed combinations, while polarity varies only slightly with domain-aspect seeds and remains stable across polarity-seed combinations.Domain-aspect seed experiments report standard deviation below 5%; polarity results vary by only 0.8% with aspect seeds and 1.2% across polarity-seed runs.
- Aspect-term/opinion-word separation: Aspect-term/opinion-word separation produces generally high correct-assignment proportions compared with random assignment.The reported experiment varies the number of Brown clusters for English.
5. Conclusions and future work
W2VLDA performs aspect and sentiment classification with almost no supervision and without language- or domain-specific resources. It combines unsupervised approaches, supports multilingual and multidomain application, and automatically separates aspect terms from opinion words.
- W2VLDA performs aspect and sentiment classification with almost no supervision and without language- or domain-specific resources.
- The system combines word embeddings and Latent Dirichlet Allocation to bootstrap information from a domain corpus.
- A single seed word per desired aspect and polarity is the only user-provided supervision.
- W2VLDA can be applied across languages and domains with almost no adaptation because topics and polarities are paired with user-selected aspect names.
- The system automatically separates aspect terms from opinion words to provide clearer domain-aspect vocabulary information.
- The evaluation covers aspect classification across several domains, comparisons with LDA-based approaches, and multilingual SemEval 2016 task 5 ABSA data.