Adaptive Parser-Centric Text Normalization

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1159–1168, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Adaptive Parser-Centric Text Normalization Congle Zhang∗ Dept of Computer Science and Engineering University of Washington, Seattle, WA 98195, USA clzhang@cs.washington.edu Tyler Baldwin Howard Ho Benny Kimelfeld Yunyao Li IBM Research - Almaden 650 Harry Road, San Jose, CA 95120, USA {tbaldwi,ctho,kimelfeld,yunyaoli}@us.ibm.com Abstract Text normalization is an important first step towards enabling many Natural Lan- guage Processing (NLP) tasks over infor- mal text. While many of these tasks, such as parsing, perform the best over fully grammatically correct text, most existing text normalization approaches narrowly define the task in the word-to-word sense; that is, the task is seen as that of mapping all out-of-vocabulary non-standard words to their in-vocabulary standard forms. In this paper, we take a parser-centric view of normalization that aims to convert raw informal text into grammatically correct text. To understand the real effect of nor- malization on the parser, we tie normal- ization performance directly to parser per- formance. Additionally, we design a cus- tomizable framework to address the often overlooked concept of domain adaptabil- ity, and illustrate that the system allows for transfer to new domains with a minimal amount of data and effort. Our experimen- tal study over datasets from three domains demonstrates that our approach outper- forms not only the state-of-the-art word- to-word normalization techniques, but also manual word-to-word annotations. 1 Introduction Text normalization is the task of transforming in- formal writing into its standard form in the lan- guage. It is an important processing step for a wide range of Natural Language Processing (NLP) tasks such as text-to-speech synthesis, speech recognition, information extraction, parsing, and machine translation (Sproat et al., 2001). ∗This work was conducted at IBM. The use of normalization in these applications poses multiple challenges. First, as it is most often conceptualized, normalization is seen as the task of mapping all out-of-vocabulary non-standard word tokens to their in-vocabulary standard forms. However, the scope of the task can also be seen as much wider, encompassing whatever actions are required to convert the raw text into a fully gram- matical sentence. This broader definition of the normalization task may include modifying punc- tuation and capitalization, and adding, removing, or reordering words. Second, as with other NLP techniques, normalization approaches are often fo- cused on one primary domain of interest (e.g., Twitter data). Because the style of informal writ- ing may be different in different data sources, tailoring an approach towards a particular data source can improve performance in the desired do- main. However, this is often done at the cost of adaptability. This work introduces a customizable normal- ization approach designed with domain transfer in mind. In short, customization is done by provid- ing the normalizer with replacement generators, which we define in Section 3. We show that the introduction of a small set of domain-specific gen- erators and training data allows our model to out- perform a set …

Content truncated to 500 words