Spark-gram: Mining frequent N-grams using parallel processing in Spark

Mining sequence patterns in form of n-grams (sequences of words that appear consecutively) from a large text data is one of the fundamental parts in several information retrieval and natural language processing applications. In this work, we present Spark-gram, a method for large scale frequent sequence mining based on Spark that was adapted from its equivalent method in MapReduce called Suffix-σ. Spark-gram design allows the discovery of all n-grams with maximum length σ and minimum occurrence frequency τ, using iterative algorithm with only a single shuffle phase. We show that Spark-gram can outperform Suffix-σ mainly when τ is high but potentially worse when the value of σ grows higher.