Text　comprehension　and　information　retrieval　are　two　essential　methods　which　could　be　reinforced　by　modeling　to　semantic　similarity　in　sentences　and　phrases.　However,　there　are　general　problems　of　traditional　methods　on　LSTM　which　is　used　to　process　the　input　sentences.　Those　semantic　vectors　cannot　fully　represent　the　entire　information　sequence　and　the　information　contained　in　firstly　input　content　will　be　diluted　or　overwritten　by　the　late　r　information.　The　longer　the　input　sequence,　the　more　serious　this　phenomenon　is.　In　order　to　address　these　problems,　we　propose　new　methods　with　self-attention.　It　can　incorporate　weights　of　special　words　and　highlight　the　comparison　of　the　similarity　in　key　words.　Compared　with　normal　self-attention　which　can　only　incorporate　the　weight　of　the　key　words　into　the　naive　sentences　and　describe　position　information　on　sentences　through　position　encoding.　Our　experiment　shows　that　new　method　can　improve　the　performance　of　model.　©　2018　IEEE.