A probabilistic formulation of keyword spotting
- Autores: Joan Puigcerver Pérez
- Directores de la Tesis: Enrique Vidal Ruiz (dir. tes.), Alejandro Hector Toselli Rossi (dir. tes.)
- Lectura: En la Universitat Politècnica de València ( España ) en 2018
- Idioma: español
- Tribunal Calificador de la Tesis: Jean-Marc Ogier (presid.), Francisco Casacuberta Nolla (secret.), Andreas Fischer (voc.)
- Programa de doctorado: Programa de Doctorado en Informática por la Universitat Politècnica de València
- Materias:
- Enlaces
- Tesis en acceso abierto en: RiuNet
- Resumen
Keyword Spotting, applied to handwritten text documents, aims to retrieve the documents, or parts of them, that are relevant for a query, given by the user, within a large collection of documents. The topic has gained a large interest in the last 20 years among Pattern Recognition researchers, as well as digital libraries and archives.
This thesis, first defines the goal of Keyword Spotting from a Decision Theory perspective. Then, the problem is tackled following a probabilistic formulation. More precisely, Keyword Spotting is presented as a particular instance of Information Retrieval, where the content of the documents is unknown, but can be modeled by a probability distribution. In addition, the thesis also proves that, under the correct probability distributions, the framework provides the optimal solution, under many of the evaluation measures traditionally used in the field.
Later, different statistical models are used to represent the probability distribution over the content of the documents. These models, Hidden Markov Models or Recurrent Neural Networks, are estimated from training data, and the corresponding distributions over the transcripts of the images can be efficiently represented using Weighted Finite State Transducers.
In order to make the framework practical for large collections of documents, this thesis presents several algorithms to build probabilistic word indexes, using both lexicon-based and lexicon-free models. These indexes are very similar to the ones used by traditional search engines.
Furthermore, we study the relationship between the presented formulation and other seminal approaches in the field of Keyword Spotting, highlighting some limitations of the latter. Finally, all the contributions are evaluated experimentally, not only on standard academic benchmarks, but also on collections including tens of thousands of pages of historical manuscripts. The results show that the proposed framework and algorithms allow to build very accurate and very fast Keyword Spotting systems, with a solid underlying theory.
