Error Corrective Learning for Semantic parsing F. Jurˇ c´ ıˇ cek, F. Mairesse, M. Gaˇ si´ c, S. Keizer, B. Thomson, K. Yu, and S. Young Engineering Department Cambridge University Trumpington Street, Cambridge, CB2 1PZ, UK {fj228, farm2, mg436, sk561, brmt2, ky219, sjy}@eng.cam.ac.uk Abstract a In this paper, we present a semantic parser which transforms initial naive semantic hy- pothesis into correct semantics by using an or- dered set of rules. These rules are learned au- tomatically from the training corpus with no linguistic knowledge. 1 Introduction Check how you cal STEP. ECL parser? Semantic parsing is important part of a spoken di- alogue system (Williams and Young, 2007; Thom- son et al, 2008). The goal of a semantic parsing is to construct formal meaning representation (seman- tics) which is directly executable by a dialogue man- ager. Such semantics is usually defined by a gram- mar, e.g. LR grammar for GeoQuery domain (Wong and Mooney, 2006) better paper???, which is de- signed by a domain expert and easy to interpret by a dialogue manager or question answering system. Semantic parsing can be understood as machine translation from a natural language to a formal lan- guage. First, we do not not have formal grammar for natural language is ungrammatical, include hes- itations, and very often only fragments of complete sentences, e.g. “Boston to Miami tomorrow”. In our approach, we adapt transformation-based learning (TBL) (Brill, 1995) to the problem of se- mantic parsing slot classification, attribute/value pair extraction. TBL attempts to find an ordered list of transformation rules to improve a baseline an- notation. The rules decompose the search space into a set of consecutive words (windows) within which align- ment links are added, to or deleted from, the initial alignment. TBL is an appropriate choice for this problem for the following reasons: 1. It can be optimized di- rectly with respect to an evaluation metric. 2. It learns rules that improve the initial prediction iter- atively, so that it is capable of correcting previous errors in subsequent iterations. 3. It provides a read- able description (or classification) of errors made by the initial system, thereby enabling alignment refine- ments. The rest of the paper is organized as follows: In the next section we describe previous work on se- mantic parsing. Section 3.2 presents am example of TBL based semantic parsing. Section 3.3 de- scribes the learning process. Section 4 compares ECL to the previously developed semantic parser on ATIS (Dahl et al, 1994) and TownInfo (Williams and Young, 2007; Thomson et al, 2008) tasks. We show that ECL is competitive to the state-of- the-art semantic parser on the ATIS task without us- ing any handcrafted linguistic knowledge. 2 Related work There has been a large amount of work done on learning to map sentences into their semantics. Many different techniques have been considered in- cluding machine translation (Wong and Mooney, 2006) techniques using inductive logic programing (?) support vector machines (Mairesse et al, 2009) and tree kernels (Kate, 2008)