The objective of the present invention is to estimate a word meaning with high precision by means of unsupervised learning. A word meaning estimation device (100) calculates, for each word, a probability density for when each concept that has been extracted as a candidate word meaning is employed as the word meaning, said calculation being based on the distance of the characteristics of the context of the selected word and the other words, the distance between the selected concept and the word meaning of the other words, a context bandwidth, which is the bandwidth for the characteristics of the context, a concept bandwidth, which is the bandwidth for the concept distance, and a word meaning probability, which is the probability that the selected word has the selected word meaning. In addition, the estimation device executes a calculation process multiple times, whereby the context bandwidth, the concept bandwidth, and the aforementioned probability are recalculated on the basis of the probability density that has been calculated, and the estimation device estimates, for each word, that the concept for which the calculated probability density is the highest is the meaning of the word.
展开▼
