Extração de informação não-supervisionada por segmentação de texto

Abstract

In this work we propose, implement and evaluate a new unsupervised approach for the problem of Information Extraction by Text Segmentation (IETS). Our approach relies on information available on pre-existing data to learn how to associate segments in the input string with attributes of a given domain relying on a very effective set of content-based features. The effectiveness of the content-based features is also exploited to directly learn from test data structure-based features, with no previous human-driven training, a feature unique to our approach. Based on our approach, we have produced a number of results to address the IETS problem in a unsupervised fashion. In particular, we have developed, implemented and evaluated distinct IETS methods, namely ONDUX, JUDIE and iForm. ONDUX (On Demand Unsupervised Information Extraction) is an unsupervised probabilistic approach for IETS that relies on content-based features to bootstrap the learning of structure-based features. Structure-based features are exploited to disambiguate the extraction of certain attributes through a reinforcement step, which relies on sequencing and positioning of attribute values directly learned on-demand from the input texts. JUDIE (Joint Unsupervised Structure Discovery and Information Extraction) aims at automatically extracting several semi-structured data records in the form of continuous text and having no explicit delimiters between them. In comparison with other IETS methods, including ONDUX, JUDIE faces a task considerably harder, that is, extracting information while simultaneously uncovering the underlying structure of the implicit records containing it. In spite of that, it achieves results comparable to the state-of- the-art methods. iForm applies our approach to the task of Web form filling. It aims at extracting segments from a data-rich text given as input and associating these segments with fields from a target Web form. The extraction process relies on content-based features learned from data that was previously submitted to the Web form. All of these methods were evaluated considering different experimental datasets, which we use to perform a large set of experiments in order to validate our approach and methods. These experiments indicate that our proposed approach yields high quality results when compared to state-of-the-art approaches and that it is able to properly support IETS methods in a number of real applications.