Autotext: autoML for text classification

Non-experts in Machine Learning research have an increasing demand for easy-to-use methods to model solutions that use the large amounts of data available today, where such solutions are expected to perform at least as well as one build by a human with profound knowledge of ML and statistics. AutoML...

Descripción completa

Detalles Bibliográficos
Autor: Jorge Madrid
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2019
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:inglés
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/1950
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/1950
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Inspec/AutoML
info:eu-repo/classification/Inspec/Text classification
info:eu-repo/classification/Inspec/Meta - learning
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
info:eu-repo/classification/cti/120323
Descripción
Sumario:Non-experts in Machine Learning research have an increasing demand for easy-to-use methods to model solutions that use the large amounts of data available today, where such solutions are expected to perform at least as well as one build by a human with profound knowledge of ML and statistics. AutoML is the area that investigates the automation of Machine Learning. Some state-of-the-art methods for approaching this problem are already available, nonetheless none of them concentrate on the challenges of Natural Language Processing. This work comprises an extensive study in text classification where 81 different problems were approached with the most commonly used algorithms. Leveraging the obtained metadata from these experiments, a method that automatically builds pipelines for classifying text documents is proposed. This method contemplates the optimization of a classification model and its hyper-parameters as well as the selection of the representation vector for a text in a given set of unprocessed text documents. A characterization for tasks is introduced as part of this method, but the reach of this novel description is not limited to AutoML problems. Results in our experimentation show that the proposed AutoML method and the novel characterization outperform previous approaches for automating text classification and under certain circumstances obtain comparable results to state-of-theart models. Being one of the first works to explore AutoML in NLP, several further questions can be derived from this thesis with potential impact for both fields.