Constructing a Linearly Combined Similarity Measure with High Accuracy for Assessing the Similarity between Linguistic Items

Selecting local similarity measures and weighting their contributions to construct a linearly combined similarity measure with high accuracy is a key problem in assessing the similarity between linguistic items. Focusing on this problem, a number of approaches have been presented during the past few decades. Each approach can construct a linearly combined measure with high accuracy in its specific case. However, constructing such a measure for arbitrary cases remains a challenge. In this paper, an approach for constructing different linearly combined measures with high accuracy in different cases is proposed. This approach uses the Pearson correlation coefficient between the computed and judged similarities to quantify the accuracy of a linearly combined measure. For different cases, different local measures are selected and different weights are assigned by maximizing this coefficient. Thus the approach can ensure high accuracy in arbitrary cases. The effectiveness of the approach is theoretically proved and a set of experiments are carried out to verify the result of this proof. The proof and experiment results show that the linearly combined measure constructed by the approach has high accuracy and the weight assignment and local measure selection ways are helpful to improve the accuracy of the linearly combined measure.