BinLu et al. ACL2011

Citation

Joint Bilinguial Sentiment Classification with Unlabeled Parallel Corpora, Bin Lu, Chenhao Tan, Claire Cardie and Benjamin K. Tsou, ACL 2011

Online version

Joint Bilingual Sentiment Classification with Unlabeled Parallel Corpora

Summary

This paper address the Sentiment analysis problem on sentence level for multiple languages. They propose to leverage parallel corpora to learn a MaxEnt-based EM model that consider both languages simultaneously under the assumption that sentiment labels for parallel sentences should be similar.

The experimented on 2 dataset: MPQA Multi-Perspective Question Answering and NTCIR-6 Opinion

Evaluation

This paper compared their method with other 3 kind of state-of-the-art baseline algorithms.

 1. The first kind of baseline algorithms are training separate classifiers on different languages. For this kind, the authors used MaxEnt, SVM and Monolingual TSVM
 2. The second kind of baseline is Bilingual TSVM
 3. The third kind is semi-supervised learning strategy Co-training

Discussion

This paper addresses the problem of bilingual sentiment classification. It leverages some parallel corpus, or pseudo-parallel corpus which is generated from automatic translation software like Google Translate, to build a MaxEnt model that maximize the joint probability p(y1, y2|x1, x2; w1, w2) under the assumption that the same idea expressed by different languages should have similar polarity.

The strong points of the paper includes:

 1. It maximizes the joint probability so that the model can consider different languages simultaneously and will not biased to one language. 
    Moreover, it takes the translation quality into consideration so that it will not be severely damaged by poor translation quality and can leverage some pseudo-parallel corpus.
 2. It takes EM algorithm to leverage more unlabeled parallel data, which is much more earlier to get.

The weak point of the paper includes:

 1. The baseline algorithms is too weak. It mostly compares their algorithm with some algorithm that didn't take special consideration about this configuration, so it's not surprising        
    that the proposed algorithm can out-perform the baselines.
 2. There is a limitation caused by translation. Current translation algorithms can barely give meaningful translation for documents, and parallel corpus on document level is also rare.
    This make this algorithm hard to go above the sentence level.

Related papers

In sense of multilingual sentiment analysis, there several works like:

• Paper:Learning multilingual subjective language via cross-lingual projections:[1]
• Paper:Multilingual subjectivity: Are more languages better?:[2]
• Paper:Cross-language text classification using structural correspondence learning.:[3]

In sense of semi-supervised learning, related papers include:

• Paper:Combining labeled and unlabeled data with co-training:[4]
• Paper:Text classification from labeled and unlabeled documents using EM.:[5]

Study plan

• Article:Expectation Maximization Algorithm:Expectation-maximization algorithm
• Article:Maximum Entropy Model:Maximum Entropy model
• Paper:Combining labeled and unlabeled data with co-training:[6]
• Paper:Learning multilingual subjective language via cross-lingual projections:[7]