With　the　rapid　development　of　Ethereum,　vast　amounts　of　data　are　recorded　on　the　blockchain　through　transactions,　encompassing　diverse　and　extensive　textual　information.　While　Long　Short-Term　Memory　(LSTM)　models　have　shown　remarkable　effectiveness　in　sentiment　analysis　tasks　in　recent　years,　they　often　encounter　situations　where　different　features　have　equal　importance　when　processing　such　textual　data.　Therefore,　this　study　introduces　a　Bidirectional　LSTM　model　with　a　Multi-Head　Attention　mechanism　(MABLSTM)　designed　for　sentiment　analysis　tasks　in　Ethereum　transaction　texts.　BLSTM　consists　of　two　distinct　and　independent　LSTMs　that　consider　information　flow　from　two　directions,　capturing　contextual　information　from　both　the　past　and　the　future.　The　outputs　from　the　BLSTM　layer　are　enhanced　using　a　multi-head　attention　mechanism　to　amplify　the　importance　of　sentiment　words　and　blockchain-specific　terms.　This　paper　evaluates　the　effectiveness　of　MABLSTM　on　Ethereum　transaction　data　through　experiments　conducted　on　an　Ethereum　transaction　dataset,　comparing　MABLSTM　with　CNN,　SVM,　ABLSTM　and　ABCDM.　The　results　demonstrate　the　effectiveness　and　superiority　of　MABLSTM　in　sentiment　analysis　tasks.　This　approach　accurately　analyzes　sentiment　polarity　in　Ethereum　transaction　texts,　providing　valuable　information　for　Ethereum　participants　and　researchers　to　support　decision-making　and　emotional　analysis.