Hierarchical and integrated error recovery based on bidirectional chart parsing technique
Kyongho Min · UNSWorks (University of New South Wales, Sydney, Australia) · 2022
A system in this thesis employs a hierarchical error recovery strategy: a parser for well-formed sentences, a second parser for repairing single-error sentences, and a third parser for repairing multiple-error sentences. CHAPTER (CHArt Parser for Two-stage Error Recovery) covers the first two stages, and MERCHANT (Multiple Error Recovery with CHArts and Need-arc Trees) covers the third stage. CHAPTER performs automatic syntactic and semantic parsing of ill-formed sentences, integrating various levels of information: lexical, syntactic, surface case, and semantic, using an augmented context-free grammar (augmented with syntactic feature constraints). It is composed of four layers of processes: morphological, syntactic, surface case, and semantic, and all subsystems are controlled by a single integrated-agenda system. At the lexical level, for spelling error correction, the inferred syntactic information (i.e., syntactic category and features) and semantic information are used to reduce the number of alternative corrections; each correction also has a penalty score determined by a Pythagorean metric for the Qwerty keyboard layout. CHAPTER can use the semantic information to correct real-word errors, if a concept associated with the erroneous word violates semantic selectional restrictions. At the syntactic level, for error recovery, the second parser employs generalised top-down chart parsing in bidirectional mode, and separates an error correction phase from an error detection phase. The best correction among alternative repairs is selected by heuristics (which are expressed in terms of penalty scores, based on error types and weight of a repaired constituent in a local tree). Surface case processing maps a surface structure of a sentence to its deep structure using a transformational grammar model. The surface case processing aims to extract maximal syntactic information from the surface structure, to help interpret the meaning of the sentence. Semantic processing interprets the meaning of a sentence using semantic selectional restrictions: act templates based on a concept hierarchy and meta-concepts represented by a type of boolean expression. For ill-formed sentences, the semantic processing filters out meaningless repairs suggested by the syntactic recovery system.