相似语义-机器学习.pptVIP

4 Experiments on similarity methods Strategy 1: Similarity measure is linear and exclusively based on the shortest path length between the two words [18], using both IS-A and HAS-A links as described in Section 3.2： (10) where M is the maximum depth of the semantic hierarchy. WordNet was used in our study and, thus, M was set to 15. This value could be any other number provided that 2Mmax(l). This ensures that similarity is not negative, as a positive similarity measure is more intuitive. 4 Experiments on similarity methods Strategy 2：(Fig.2a) Similarity measure is a linear combination of shortest path length and depth. (11) where α,β belong to [0,1]; and α ,β cannot be zero at the same time. Our experiments demonstrated that the correlation coefficient between S2 and human similarity judgments reach the maximum value at α=0.05 and β=1. Fig. 2a shows the correlation value against α and β on D1. Using the optimal parameters α=0.05 and β=1,the similarities for word pairs in the test set were calculated.The correlation coefficient between S2 and human similarity judgments is 0.8315 on D0. Fig. 2: Correlation coefficient versus strategy parameters on design word set D1. 4 Experiments on similarity methods Strategy 3：(Fig.2b) The similarity measure is a nonlinear function of the shortest path length. The exponential function as defined in (3) is considered in this study, so the similarity is defined as ： (12) It is observed that the strongest correlation is reached at α=0.25.Using this optimal α, the similarities for word pairs in the test set were calculated.The correlation coefficient between S2 and human similarity judgments is 0.8911 on D0. 4 Experiments on similarity methods Strategy 4：(Fig.2c) Semantic similarity is considered to be governed by the shortest path length as well as the depth of the subsumer, that is (13) where α=0 and β 0 are par