抄録
In machine translation-mediated communication, when each party has different cultural and language backgrounds, a particular word could be related to different meanings and different images. This could cause failures to establish mutual understanding. To solve this problem, we have proposed an image feature-based method to automatically determine words that could cause misunderstanding. This method calculates and determines the optimal threshold by comparing the result from the automated method to human judgment. We applied this method to 500 concepts and compared the judgments using 400 concepts for threshold optimization and 100 concepts for testing our proposed method. We found that 0.55 was the optimal threshold with 76 percent accuracy. Moreover, we conducted the chi-square test to determine whether the accuracy is significantly different among hypernyms of the concepts and the result statistically did not recognize any significant difference.