Análise do grupo constituinte

Constituency parsing analyzes a sentence according to a phrase-structure grammar, organizing words into nested constituents such as noun phrases (NP), verb phrases (VP), and prepositional phrases (PP). The output is a tree whose leaves are words and whose internal nodes are phrase labels, all rooted at a single S (sentence) node. For example, 'The cat sat on the mat' splits into an NP ('The cat') and a VP ('sat on the mat'), which itself contains a verb and a PP. This differs from dependency parsing, which links words directly to one another rather than grouping them into phrases. Classic approaches used the CYK algorithm with probabilistic context-free grammars; modern systems use neural networks trained on treebanks like the Penn Treebank. Stern et al.’s Minimal Span-Based Neural Constituency Parser scores contiguous word spans and phrase labels independently, then constructs a tree with structured inference. The paper reports 91.79 labeled-bracket F1 on its Penn Treebank setup; scores depend on evaluation protocol and model configuration.