This paper presents effect of preprocessing on Long Short Term Memory (LSTM) basedrnsentiment analysis for Amharic language. Sentiment analysis or opinion mining is anrnapproach used to analyze user generated textual contents to a way that is importantrnfor decision making. User generated textual contents are found everywhere such as,rnsocial media posts, product reviews blogs and form. Developing sentiment analysis isrna challenging task due to different writing styles and variation of word meanings. Tornanalyze the sentiment of these textual contents, several approaches use labeled lexicons.rnIn the preprocessing step of the approaches, Emojis are removed and words are stemmed.rnHowever, Emojis are usually used to express opinions.rnIn this research, we propose to use Emojis to automatically label texts for sentimentrnanalysis. In addition, we investigate the impact of using unstemmed words on sentimentrnanalysis. To evaluate the proposed labeling scheme on sentiment anaslysis, we conductedrnan experiment using 9,138 Amharic textual comments. The results show that integratingrnEmojis with lexicons for labeling gives 0.55% higher accuracy than using only lexicons.rnTo investigate the effect of using stemming as part of preprocessing strategy, LSTMrnbased Amharic sentiment analysis with and without stemming is conducted using 1077rncomments. Result shows that applying stemming drops the accuracy of the sentimentrnanalysis by 6.43% while using long short-term memory based sentiment analysis, andrn0.43% while using bi-gram multinomial naive bayes.rnKeyword: - Amharic sentiment analysis, Emoji, Natural Language Processing (NLP),rnSentiment analysis, Stemming.