Difference between revisions of "Cross-Lingual Mixture Model for Sentiment Classification, Xinfan Meng, Furu Wei, Xiaohua Liu, Ming Zhou, Ge Xu, Houfeng Wang, ACL 2012"
| (3 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
| − | == Citation == | + | ==Citation== |
| + | |||
Cross-Lingual Mixture Model for Sentiment Classification, Xinfan Meng, Furu Wei, Xiaohua Liu, Ming Zhou, Ge Xu, Houfeng Wang, ACL 2012 | Cross-Lingual Mixture Model for Sentiment Classification, Xinfan Meng, Furu Wei, Xiaohua Liu, Ming Zhou, Ge Xu, Houfeng Wang, ACL 2012 | ||
| − | == Online version == | + | ==Online version== |
An online pdf version is here[http://www.aclweb.org/anthology-new/P/P12/P12-1060.pdf] | An online pdf version is here[http://www.aclweb.org/anthology-new/P/P12/P12-1060.pdf] | ||
| − | == Summary == | + | ==Summary== |
This paper propose a cross-lingual mixture model (CLMM) to tackle the problem of cross-lingual sentiment classification. The motivation for this work is the lack of labeled data in target language (therefore, we want to bring labeled data in source language to help). | This paper propose a cross-lingual mixture model (CLMM) to tackle the problem of cross-lingual sentiment classification. The motivation for this work is the lack of labeled data in target language (therefore, we want to bring labeled data in source language to help). | ||
| − | + | Having a labeled source data Ds, a parallel corpus U and an optional labeled target data Dt, they maximize the log-likelihood function for the parallel corpus. | |
| − | Having a labeled source data | ||
[[File:P12-1060.png]] | [[File:P12-1060.png]] | ||
This is to say that a word in the parallel corpus is generative by 1) directly generate a Chinese word according to the polarity of the sentence OR 2) first generate an English word with the same polarity and meaning, and then translate it to a Chinese word. | This is to say that a word in the parallel corpus is generative by 1) directly generate a Chinese word according to the polarity of the sentence OR 2) first generate an English word with the same polarity and meaning, and then translate it to a Chinese word. | ||
| − | |||
At the same time, they want to maximize the log-likelihood function for the source data (and the target data, optional). | At the same time, they want to maximize the log-likelihood function for the source data (and the target data, optional). | ||
| − | |||
The words projection probability is given by the Berkeley aligner. The words generation probability given sentimental class is estimated using EM. | The words projection probability is given by the Berkeley aligner. The words generation probability given sentimental class is estimated using EM. | ||
| − | |||
Finally, the words generation probability can be used in Naive Bayes classifier. | Finally, the words generation probability can be used in Naive Bayes classifier. | ||
| − | == Evaluation == | + | ==Evaluation== |
The author evaluate CLMM's performance using [[http://malt.ml.cmu.edu/mw/index.php/Dataset:MPQA MPQA]] and [[http://malt.ml.cmu.edu/mw/index.php/NTCIR-6_Opinion NTCIR]] in mainly two cases: | The author evaluate CLMM's performance using [[http://malt.ml.cmu.edu/mw/index.php/Dataset:MPQA MPQA]] and [[http://malt.ml.cmu.edu/mw/index.php/NTCIR-6_Opinion NTCIR]] in mainly two cases: | ||
| − | |||
1) Keep the labeled data in target language (Chinese) unavailable. | 1) Keep the labeled data in target language (Chinese) unavailable. | ||
| − | |||
Can greatly improve the performance (71%) comparing with MT-SVM(52%-62%) and MT-Cotrain(59%-65%). | Can greatly improve the performance (71%) comparing with MT-SVM(52%-62%) and MT-Cotrain(59%-65%). | ||
| + | 2) Using the labeled target language (Chinese) data. | ||
| + | Still beat the baseline SVM (using the labeled data in Chinese to train the model), and can compete other state-of-art methods like the Joint-Train & MT-Cotrain (Wan, 2009), while need less time in training. | ||
| − | + | ==Discussion== | |
| − | + | If we only use MT results directly, we will suffer from two things 1) The vocabulary is quite limited, hence many words in target language can not enjoy the help from the source language. 2) The MT systems have defects, Like "too good to be true" can be positive after MT into Chinese, where leads to errors. By using CLMM in this paper, we can partly solve these two problems. But it seems like, it is still a way to find sentimental words in target language. In the experiment we can clearly see that, when we have labeled target language data, the improvement given by this method is quite limited. Not sure if other auto sentimental words expands methods can also help the system like CLMM. | |
| − | == Related papers == | + | ==Related papers== |
Xiaojun Wan. 2009. Co-training for cross-lingual senti- ment classification. | Xiaojun Wan. 2009. Co-training for cross-lingual senti- ment classification. | ||
| − | |||
Bin Lu, Chenhao Tan, Claire Cardie, and Benjamin K. Tsou. 2011. Joint bilingual sentiment classification with unlabeled parallel corpora. | Bin Lu, Chenhao Tan, Claire Cardie, and Benjamin K. Tsou. 2011. Joint bilingual sentiment classification with unlabeled parallel corpora. | ||
Latest revision as of 21:52, 1 October 2012
Citation
Cross-Lingual Mixture Model for Sentiment Classification, Xinfan Meng, Furu Wei, Xiaohua Liu, Ming Zhou, Ge Xu, Houfeng Wang, ACL 2012
Online version
An online pdf version is here[1]
Summary
This paper propose a cross-lingual mixture model (CLMM) to tackle the problem of cross-lingual sentiment classification. The motivation for this work is the lack of labeled data in target language (therefore, we want to bring labeled data in source language to help). Having a labeled source data Ds, a parallel corpus U and an optional labeled target data Dt, they maximize the log-likelihood function for the parallel corpus.
This is to say that a word in the parallel corpus is generative by 1) directly generate a Chinese word according to the polarity of the sentence OR 2) first generate an English word with the same polarity and meaning, and then translate it to a Chinese word. At the same time, they want to maximize the log-likelihood function for the source data (and the target data, optional). The words projection probability is given by the Berkeley aligner. The words generation probability given sentimental class is estimated using EM. Finally, the words generation probability can be used in Naive Bayes classifier.
Evaluation
The author evaluate CLMM's performance using [MPQA] and [NTCIR] in mainly two cases: 1) Keep the labeled data in target language (Chinese) unavailable. Can greatly improve the performance (71%) comparing with MT-SVM(52%-62%) and MT-Cotrain(59%-65%). 2) Using the labeled target language (Chinese) data. Still beat the baseline SVM (using the labeled data in Chinese to train the model), and can compete other state-of-art methods like the Joint-Train & MT-Cotrain (Wan, 2009), while need less time in training.
Discussion
If we only use MT results directly, we will suffer from two things 1) The vocabulary is quite limited, hence many words in target language can not enjoy the help from the source language. 2) The MT systems have defects, Like "too good to be true" can be positive after MT into Chinese, where leads to errors. By using CLMM in this paper, we can partly solve these two problems. But it seems like, it is still a way to find sentimental words in target language. In the experiment we can clearly see that, when we have labeled target language data, the improvement given by this method is quite limited. Not sure if other auto sentimental words expands methods can also help the system like CLMM.
Related papers
Xiaojun Wan. 2009. Co-training for cross-lingual senti- ment classification. Bin Lu, Chenhao Tan, Claire Cardie, and Benjamin K. Tsou. 2011. Joint bilingual sentiment classification with unlabeled parallel corpora.
