Content　curation　social　networks　(CCSNs),　which　provide　users　a　platform　to　share　their　interests　by　multimedia　information,　are　the　most　rapidly　growing　social　networks　in　recent　years.　Since　large-scale　multimodal　data　have　been　generated　by　CCSN　users,　learning　multimodal　representations　for　contents　have　become　the　key　to　the　progress　of　many　applications　such　as　user　interest　analysis　and　recommender　system　for　curation　networks.　Learning　representations　for　CCSNs　faces　a　vital　challenge:　the　sparsity　of　multimodal　data.　It　is　difficult　for　most　existing　approaches　to　learn　effective　representations　for　multimodal　CCSNs　because　they　didn＇t　provide　a　solution　on　how　to　model　sparse　and　noisy　multimodal　data.　In　this　paper,　we　propose　a　2-step　approach　to　learn　accurate　multimodal　representations　from　sparse　multimodal　data.　First,　we　propose　a　novel　Board-Image-Word　(BIW)　graph　to　model　the　multimodal　data.　Benefited　from　the　unique　board-image　relation　on　CCSNs,　embeddings　of　images　and　texts　which　endow　semantic　relations　are　learned　from　the　network　topology　of　the　BIW　graph.　As　the　second　step,　a　deep　vision　model　with　modified　loss　function　are　trained　by　minimizing　the　distance　between　the　visual　features　of　contents　and　their　corresponding　semantic　relation　embeddings　to　learn　representations　which　incorporate　visual　information　and　graph-based　semantic　relations.　Experiments　on　the　dataset　from　Huaban.com　demonstrate　that　under　the　circumstance　of　sparser　text　modality,　our　method　significantly　outperformed　multimodal　DBN,　DBM　and　unimodal　representation　learning　methods　on　pin　classification　and　board　recommendation　tasks.　©　2019　IEEE.