This paper introduces Compositional Vector Grammars (CVGs), a new parsing model that combines the speed of small-state probabilistic context-free grammars (PCFGs) with the semantic richness of neural word representations and compositional phrase vectors. CVGs use a syntactically untied recursive neural network to learn compositional vector representations of phrases and words. This model is more linguistically plausible as it uses different composition functions for parent nodes based on the syntactic categories of their children. The CVG improves the Stanford Parser's F1 score from 86.56% to 90.44% on the WSJ section 23 test set and is 20% faster than the current Stanford factored parser. The CVG also learns a soft notion of head words and improves performance on ambiguities requiring semantic information, such as PP attachments. The model is trained using a max-margin objective function that maximizes the score of the correct tree while minimizing the scores of incorrect trees. The CVG uses a bottom-up beam search approach to find the optimal tree, which is significantly faster than traditional methods. The model is evaluated on the Penn Treebank WSJ and shows improved performance over other parsers, including the Berkeley parser, Collins parser, and Charniak parser. The CVG's ability to transfer semantic knowledge between related sentences is demonstrated through experiments on PP attachment tasks. The model is implemented in Python and is available at nlp.stanford.edu.This paper introduces Compositional Vector Grammars (CVGs), a new parsing model that combines the speed of small-state probabilistic context-free grammars (PCFGs) with the semantic richness of neural word representations and compositional phrase vectors. CVGs use a syntactically untied recursive neural network to learn compositional vector representations of phrases and words. This model is more linguistically plausible as it uses different composition functions for parent nodes based on the syntactic categories of their children. The CVG improves the Stanford Parser's F1 score from 86.56% to 90.44% on the WSJ section 23 test set and is 20% faster than the current Stanford factored parser. The CVG also learns a soft notion of head words and improves performance on ambiguities requiring semantic information, such as PP attachments. The model is trained using a max-margin objective function that maximizes the score of the correct tree while minimizing the scores of incorrect trees. The CVG uses a bottom-up beam search approach to find the optimal tree, which is significantly faster than traditional methods. The model is evaluated on the Penn Treebank WSJ and shows improved performance over other parsers, including the Berkeley parser, Collins parser, and Charniak parser. The CVG's ability to transfer semantic knowledge between related sentences is demonstrated through experiments on PP attachment tasks. The model is implemented in Python and is available at nlp.stanford.edu.