The traditional way of calculating word semantic similarity is based on wordnet structure, which has a huge gap between physical concept and abstract concept, and only considering concepts' hyponymy. To solve the problem, a novel word similarity calculation algorithm based on traffic field words relation network is proposed in the paper. 10 kinds of concept relationships, including concepts of hyponymy, tool-tool object relationship, standard parts-overall and so on, are used to build traffic words knowledge network. Then modified average path length parameter is used to calculate words' semantic similarity, which accords with people's judgement. The experiment based on Finkelstein's 353 word pairs shows that the algorithm achieves more accurate word semantic similarity.