A unified lexical processing framework based on the Margin Infused Relaxed Algorithm. A case study on the Romanian Language

Proceedings of Recent Advances in Natural Language Processing, pages 91–97, Hissar, Bulgaria, 7-13 September 2013. A Unified Lexical Processing Framework Based on the Margin Infused Relaxed Algorithm. A Case Study on the Romanian Language Tiberiu Boroș Research Institute for Artificial Intelligence, “Mihai Drăgănescu”, Romanian Academy tibi@racai.ro Abstract General natural language processing and text-to-speech applications require certain (lexical level) processing steps in order to solve some frequent tasks such as lemmatization, syllabification, lexical stress prediction and phonetic transcription. These steps usually require knowledge of the word’s lexical composition (derivative morphology, inflectional affixes, etc.). For known words all applications use lexicons, but there are always out-of-vocabulary (OOV) words that impede the performance of NLP and speech synthesis applications. In such cases, either rule based or data-driven techniques are used to automatically process these OOV words and generate the desired results. In this paper we describe how the above mentioned tasks can be achieved using a Perceptron with the Margin Infused Relaxed Algorithm (MIRA) and sequence labeling. 1 Introduction Natural Language Processing (NLP) applications and Text-to-Speech (TTS) synthesis systems require a set of pre-processing steps that include tasks such as lemmatization, syllabification, lexical stress prediction and phonetic transcription. Because these all these tasks require knowledge of the word composition (derivative morphology, inflectional affixes, part of speech, etc.) we will refer to them as lexical processing steps. This paper presents a unified lexical processing framework based on the Margin Infused Relaxed Algorithm (MIRA) (Crammer and Singer, 2003) designed to solve the basic text-preprocessing tasks involved in both text-to- speech (TTS) synthesis and general NLP applications. Assuming that all existing systems use lexicons for known words, we focused our research in handling the difficult problems generated by presence of out-of vocabulary (OOV) or previously unseen in the training data words that negatively impact the performance of the above mentioned tasks. Our current research is focused on the Romanian language, but the methods presented here are data-driven and with proper lexicons and feature templates, they can be used for other (Latin based) languages as well. We show how we achieved state-of-the-art results on Romanian by using the MIRA framework. 2 Lexical processing with MIRA There are various methods proposed in the literature for each of the previously mentioned lexical subtasks. For each of them, we will offer a short literature review of available methods and we will compare our results with the current state-of-art systems. The previously proposed methods vary from rule-based to data-driven and different authors employ different classifiers (in data-driven approaches), such as Maximum Entropy Classifiers, Classification and Regression Trees, Support Vector Machines (SVM), Structured SVMs, Conditional Random Fields, etc. While these are all powerful methodologies, we chose the Perceptron classifier with the MIRA update learning as our sequence labeling classifier because of its robustness and its ability to obtain highly accurate results that compare to the ones obtained using CRFs. All the lexical processing methods that we propose, share the following similarities: - All of them are reformulated as sequence labeling tasks; - …

Content truncated to 500 words