A Fast Algorithm of Computing Word Similarity

Computing distributional similarity is an effective strategy for finding synonyms. The time complexity of the naive nearest-neighbor approach of computing distributional word similarity is O(n*n*m), it is inefficient for accurately representing synonymy using large corpus. We find a parse property of triple that the growth rate of average triples number of each word leveled off as corpus's size increases. Using this property we design a fast algorithm for computing word similarity whose time complexity is O(n*n). We demonstrate the efficiency of this algorithm based on the English Gig word corpus.