On improving parsing with automatically acquired semantic classes

• 80% of parsing mistakes in appositions are due to a lack of semantic information.
• We automatically gather evidence on class-instance semantic compatibility from text.
• Classes are common nouns; instances are entities characterized by name and type.
• Our best model uses both sources of evidence with smoothed conditional probability.
• Experiments reach 91.4% accuracy, a 12.9% relative improvement over the baseline.

Parsing mistakes impose an upper bound in performance on many information extraction systems. In particular, syntactic errors detecting appositive structures limit the system’s ability to capture class-instance relations automatically from texts. The article presents a method that considers semantic information to correct appositive structures given by a parser.First, we build automatically a background knowledge base from a reference collection, capturing evidence of semantic compatibility among classes and instances. Then, we evaluate three different probabilistic-based measures to identify the correct dependence on ambiguous appositive structures.Results reach a 91.4% of correct appositions which is a relative improvement of 12.9% with respect to the best baseline (80.9%) given by a state of the art parser.