Handling Normalization Issues for Part-of-Speech Tagging of Online Conversational Text

Handling Normalization Issues for Part-of-Speech Tagging of Online Conversational Text G´eraldine Damnati1, Jeremy Auguste2, Alexis Nasr2, Delphine Charlet1 Johannes Heinecke1, Fr´ed´eric B´echet2 (1) Orange Labs, Lannion, France {geraldine.damnati, delphine.charlet, johannes.heinecke}@orange.com (2) Aix-Marseille Univ, Universit´e de Toulon, CNRS, LIS {jeremy.auguste, alexis.nasr, frederic.bechet}@univ-amu.fr Abstract For the purpose of POS tagging noisy user-generated text, should normalization be handled as a preliminary task or is it possible to handle misspelled words directly in the POS tagging model? We propose in this paper a combined approach where some errors are normalized before tagging, while a Gated Recurrent Unit deep neural network based tagger handles the remaining errors. Word embeddings are trained on a large corpus in order to address both normalization and POS tagging. Experiments are run on Contact Center chat conversations, a particular type of formal Computer Mediated Communication data. Keywords: Part of Speech Tagging, Computer Mediated Communication, Spelling Error Correction 1. Introduction Contact Center chat conversation is a particular type of noisy user generated text in the sense that it is a for- mal Computer Mediated Communication (CMC) interac- tion mode. It shares some normalization issues with other CMC texts such as chatroom conversations or social media interactions but unlike the aforementioned cases, the pro- fessional context implies some specificities. For instance, contact center logs are hardly prone to Internet slang. An- other characteristic is that they are dyadic conversations with asymmetric levels of orthographic or grammatical er- rors. Agents may write with mistakes but are usually re- cruited for their linguistic skills, and can rely on predefined utterance libraries. Customers on the other hand can make mistakes for several different reasons, be it their educa- tional background, linguistic skills, or even the importance they pay to the social perception of the errors they would make. Some of them will make no mistake at all while some others will misspell almost every word. The purpose of this paper is to perform POS tagging on this particular type of Noisy User Generated text. Our goal is to study to which extent it is worth normalizing text before tagging it or directly handling language deviations in the design of the tagger. We will show that a good compromise is to handle some of the errors through lexical normaliza- tion but also to design a robust POS tagger that handles orthographic errors. We propose to use word embeddings at both levels: for text normalization and for POS tagging. 2. Related work Text normalization has been studied for several years now, with different perspectives over time. When studying SMS style language, researchers tried to handle new phenomena including voluntary slang shortcuts through phonetic mod- els of pronunciation (Toutanova and Moore, 2002; Kobus et al., 2008). Recently, the effort has been more particularly set on Social Media text normalization with specific chal- lenges on Twitter texts (Baldwin et al., 2015), which has been shown to be more formal (Hu et al., 2013) that what is commonly expected. The typology of errors is slightly different and most recent works …

Content truncated to 500 words