In　recent　years,　opinion　retrieval　attracted　a　growing　research　interest　as　online　users＇　opinions　become　more　and　more　valuable　for　market　survey,　political　polls,　etc.　The　goal　of　opinion　retrieval　is　to　find　relevant　and　opinionate　documents　according　to　a　user＇s　query.　Compared　with　previous　lexicon-based　generative　model　for　opinion　retrieval　considering　that　the　sentiment　words　are　equal　for　a　query,　which　cannot　reflect　different　sentiment　words＇　relevant　opinion　strength,　we　propose　a　graph-based　approach　by　using　HITS　model　to　capture　the　sentiment　words＇　relevant　opinion　strength.　Then　the　weights　are　incorporated　into　the　weighted　lexicon-based　generative　model　for　opinion　retrieval.　Experimental　results　on　two　datasets　show　the　effectiveness　of　the　proposed　generative　model.　Compared　with　the　baseline　approach,　improvements　of　4%　and　11%　have　been　obtained　on　two　real　datasets.　©　2014　IEEE.