Building Chinese Affective Resources in Valence-Arousal Dimensions
First, this paper to solve the problem
More and more research has focused on the emotional state is represented as continuous values on the multi-dimensional, multi-dimensional space of the most common is Aalence-Arousal (VA) space. Compared with the previous emotional state is divided into several categories (positive and negative), the method can provide emotional dimension more fine-grained analysis. Then it is very lack of emotional resources with a VA rating, the Chinese resources even more serious. Therefore, the purpose of this paper is to establish Chinese dictionary VA (CVAW) can be used in sentiment analysis of VA and Chinese text corpus (CVAT), in order to enrich the emotional research and development at VA multilingual dimension.
Two, VA space
Valence pleasant and unpleasant representatives (ie, positive and negative) the extent and degree of excitement on behalf of Arousal and calm. Based on this representation, any emotional state can be represented as a point on the coordinate plane VA. I.e. each point represents an emotional state in the plane of the coordinates. As shown below.
Third, the existing dictionary and emotional resources corpus
SentiWordNet lexical resource for opinion mining and sentiment it allocates three levels each for the WordNet synset (negative, positive, objective).
Language queries and counting (LIWC): calculated over a wide range of text extent people use words in different categories.
The Chinese LIWC (C-LIWC) : The Chinese translation is LIWC, it is human translation of the Chinese depending on the application.
Norms for English Words (ANEW) provides 1,034 English words with ratings in the dimensions of pleasure, arousal and dominance, it provides the only real-valued VAD three dimensions.
The NTU Sentiment dictionary (NTUSD): uses both manual and automated approach to tagging, including both positive and negative emotional vocabulary.
Fourth, the establishment method
CVAW established: it is based on C-LIWC the above, the use of self-assessment manikin (SAVM), that is trained commentators who give each word a value in VA dimension, while SVA will give some pictures to improve the accuracy of their commentators. Valence range of dimension values of 1-9 (1 being the most negative, the most actively represented 9) 5 indicates no clear tendency of neutral emotions. Arousal is the similar methods. The last five comments were given as VA worth the average value of the word, in order to get CVAW dictionary.
CVAT establishment: at 720 select from six different types of network information, including a total of 2009 the highest number of emotional words appear include C-LIWC dictionary VA sentences were chosen to be assessed. These volunteers score 1-9 sentences, each sentence has been assessed at least 10 times. Executes a program after the clear listening rates to remove some of the abnormal value VA (VA is not the average value in the range of about 1.5 standard deviation) and inappropriate sentences. After removed, these sentences will no longer participate in the calculation of the average VA worth.
Fifth, assessment methods
Average adopt an integrated approach and the use of the word CVAW to estimate the value of VA sentences CVAT. Valence sentence (or arousal) equal to those of the sentence Valence (or arousal) average of the values contained in the word dictionary section of CVAW. The actual values of VA VA value obtained in this way are compared with the sentence, j calculates the average absolute error (the MAS), the root mean square error (RMSE) and the Pearson correlation coefficient r. While using ANEW to estimate 20 English forum message results are presented for comparison.
Sixth, the results of analysis
With CVAW to predict the effect in English with emotion resources to predict the resulting effects VA values CVAT in a sentence is considerable. And it was found, whether it is English or Chinese emotional resources emotional resources, arousal dimension is more difficult to predict than the valence dimension.