Experiments using varying sizes and machine translated data for sentiment analysis in Twitter
Alexandra Balahur Dobrescu, José M. Perea‐Ortega · Dialnet (Universidad de la Rioja) · 2013
espanolEn este articulo presentamos varios experimentos para la tarea de ana- lisis de sentimientos a nivel global dentro de la campana de evaluacion TASS. El objetivo de esta tarea es evaluar la polaridad global de textos cortos en espanol extraidos de Twitter. Para abordar esta tarea se ha aplicado un enfoque basado en aprendizaje automatico probando diferentes combinaciones de caracteristicas. Se han empleado varios diccionarios y un corpus traducido automaticamente para entrenamiento, adaptando al espanol un enfoque inicial disenado para trabajar con textos en ingles. Ademas, se probaron en cascada cuatro clasificadores separados para determinar el sentimiento desde clases de polaridad mas generales a mas precisas. Aunque esta es nuestra primera participacion, los enfoques propuestos se podrian considerar buenas estrategias para generar corpus de entrenamiento para sistemas de clasificacion de la polaridad en espanol EnglishIn this paper we present several experiments for the task entitled sen- timent analysis at global level within the TASS evaluation campaign. The aim of this task is to assess the global polarity of Spanish short texts extracted from Twit- ter. To tackle this task, an approach based on machine learning by trying different feature combinations was applied. Several in-house built dictionaries and machine- translated data for training were employed by adapting an approach designed for English to Spanish. Additionally, four separate classifiers were tested in cascade to determine the sentiment from the general to the finer-grained classes of polarity. Although this is our first participation, the proposed approaches might be conside- red good strategies to generate learning data for polarity classification systems in Spanish