Techniques for categorizing item listings based on parsing item listing titles are described. According to various embodiments, listing titles of one or more item listings on a marketplace website are accessed, the item listings being associated with a particular product category in a product category structure of the marketplace website. Words in each of the listing titles may then be converted to semantic tokens in a token symbol space, based on a tokenization process. Thereafter, n-gram modeling may be performed on the tokens corresponding to each of the listing titles of the item listings in the particular product category. One or more dominant n-gram models associated with the listing titles of the item listings in the particular product category may then be identified.