As you mentioned in the readme.md, the result was hard to reach the author's version. Have you fed the pre-trained embedding as input?
ref: https://github.com/taoshen58/DiSAN/issues/7