Semantic Parsing via l0-norm-based Alignment

Proceedings of Recent Advances in Natural Language Processing, pages 355–361, Hissar, Bulgaria, Sep 7–9 2015. Semantic Parsing via ℓ0-norm-based Alignment Zhihua Liao Foreign Studies College Hunan Normal University Changsha, China liao.zhihua61@gmail.com Qixian Zeng Foreign Studies College Hunan Normal University Changsha, China wanyouto@qq.com Qiyun Wang Foreign Studies College Hunan Normal University Changsha, China qywang@qq.com Abstract In this paper, we explore the IBM Model with a ℓ0-norm prior to the se- mantic parsing which parses a sentence to its corresponding meaning representa- tion, and compare two supervised proba- bilistic Combinatory Categorial Grammar (PCCG) online learning approaches that are Unification-Based Learning (UBL) method and Factored Unification-Based Learning (FUBL) one. Specially, we extend manually GeoQuery and ATIS datasets from English to Chinese pinyin- format string. The experiment on such benchmark datasets in both English and Chinese with two different meaning rep- resentations (i.e., lambda-calculus and variable-free expressions) demonstrates that both methods adopted this IBM Model with ℓ0-norm outperform trivially those that used the IBM Model without ℓ0- norm, and also shows small improvements of around 0.1% ∼ 0.7% of F1 for the two algorithms on nearly all conditions. 1 Introduction Learning the mapping from natural language sen- tences to formal meaning representations has be- come one of the main targets in natural language processing. Recent research has focused on learn- ing the semantic parsers directly from corpora that consist of sentences paired with their mean- ing representations (Artzi and Zettlemoyer, 2011; Artzi and Zettlemoyer, 2013; Kwiatkowski et al., 2010; Kwiatkowski et al., 2011; Lu et al., 2008; Zettlemoyer and Collins, 2005; Zettlemoyer and Collins, 2007; Zettlemoyer and Collins, 2009; Zettlemoyer and Collins, 2012). They usually em- ploy corpus-based probabilistic methods. Further- more, some research work has been explored to learn to map any natural language to a wide va- riety of logical expressions of linguistic mean- ing (Kwiatkowski et al., 2011; Liao and Zhang, 2013). For example, the training data can con- sist of Turkish, Spanish, Japanese and English sen- tences paired with lambda-calculus expressions or variable-free logical ones. Our approach is inspired by the principle of minimum description length (Barron et al., 1998; Ashish et al., 2012). The main motivation is that through adding a ℓ0-norm prior this extension of the IBM model can enable it to encourage the spar- sity in word-to-word alignment model. It uses an efficient training algorithm based on projected gra- dient descent. In this paper, we will apply this method to the semantic parsing. Our work fo- cus on the Initialization procedure that the weights for lexeme features are initialized according to coocurrance statistics between words and logical constants. They are implemented with the modifi- cation of GIZA++ toolkit which is viewed as the drop-in replacement for GIZA++ (Ashish et al., 2012). We evaluate our approach on two benchmark corpora (i.e., GeoQuery and ATIS) annotated with Chinese pinyin-format string. The GeoQuery cor- pus has complex sentence and meaning repre- sentation pairs whereas the ATIS corpus contains spontaneous and unedited text so that it is diffi- cult to analyze …

Content truncated to 500 words