Section 6.6, Methods, provides a hopefully useful summary of the techniques used by top performing systems. Additionally, all of the state-of-the-art baseline semantic similarity models from section 8, STS Benchmark, are now open source. Links to the open source implementations are provided in the paper for everything except InferSent, which as open sourced after the preparation of the task paper (
InferSent on github).