The invention provides a big data processing method based on machine learning, which comprises the following steps: given a retrieval sentence, filtering the words in the initial retrieval by using auniversal stop word list, and reserving meaningful retrieval words; the semantic block model is used to represent the lexical semantic vector. On the basis of semantic vector, cosine similarity is used to find out the closest words from other words for each initial search term, which can be used as extended search terms. The corresponding extended search terms in the initial search are used to replace the original search terms, and the newly generated search term sequence is used as the extended search terms. According to the permutation and combination of the extended search terms, the extended search sentences with different expressions are obtained. The invention improves the parallel frame of the MAPRUDUCE and is better adapted to the needs of text data mining. Aiming at the irregularity of social text, semantic vectors are used to represent and analyze the text data effectively, which is suitable for the analysis and calculation of social text mining of various scales.