Sentiment analysis is an application of artificial intelligence that determines the sentiment associated sentiment with a piece of text. It provides an easy alternative to a brand or company to receive customers' opinions about its products through user generated contents such as social media posts. Training a machine learning model for sentiment analysis requires the availability of resources such as labeled corpora and sentiment lexicons. While such resources are easily available for English, it is hard to find them for other languages such as Arabic. The aim of this research is to build an Arabic sentiment lexicon using a corpus-based approach. Sentiment scores were propagated from a small, manually labeled, seed list to other terms in a term co-occurrence graph. To achieve this, we proposed a graph propagation algorithm and compared different similarity measures. The lexicon was evaluated using a manually annotated list of terms. The use of similarity measures depends on the fact that the words that are appearing in the same context will have similar polarity. The main contribution of the work comes from the empirical evaluation of different similarity to assign the best sentiment scores to terms in the co-occurrence graph.