Generating referring expressions in a 3D environment
Loading...
Official URL
Full text at PDC
Publication date
2014
Advisors (or tutors)
Méndez Pozo, Gonzalo
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Citation
Abstract
Desde hace muchos años, el ser humano ha querido entender a las demás personas y su forma de pensar. Esto no resulta sencillo, ya que cada individuo, debido a condiciones como el sexo, la edad, la lengua materna, la cultura o los conocimientos y las vivencias previas, analiza las cosas de forma distinta. Algunas de las consecuencias de las diferentes formas de ver el mundo se ven reflejadas a la hora de la comunicación.
En este proyecto se ha investigado cómo las personas componen descripciones para referirse a alguien en concreto, y se han desarrollado algoritmos que son capaces de reproducir este comportamiento. En el campo de la Generación de Expresiones de Referencia ya hay muchos avances documentados y experimentos diversos, y este proyecto se ha centrado en investigar el comportamiento de las personas al analizar situaciones reales y cotidianas. Se ha creado una aplicación que reúne este conocimiento, y se ha demostrado que funciona proponiendo situaciones cargadas de objetos y personajes. Con los algoritmos creados, se han realizado descripciones fáciles de entender (usando lenguaje natural como lo haría un humano) que permiten reconocer personajes de forma rápida, y que tienen un bajo porcentaje de fallo para cualquier tipo de persona. Para ello se han considerado las características de la situación y del personaje a describir, y se han diseñado algoritmos que se adaptan a ellos.
Para generar las mejores situaciones para poner a prueba a nuestros algoritmos, se ha creado con el motor Unity 3D una habitación cerrada y llena de personas, como es la cafetería de la Facultad de Informática, y se ha empleado una gran variedad de personajes suficientemente diferenciables en cada situación. Con ello se han podido cronometrar las respuestas de usuarios reales y se ha determinado qué descripciones funcionan mejor en determinadas situaciones. Con la información recopilada tras pruebas reales con voluntarios, se han afinado aún más los algoritmos, creando el tipo de descripciones que personas reales quieren oír.
For years, humans have tried to understand themselves and their way of thinking. This is not easy, because each person, due to factors such as their sex, age, native language, culture or their knowledge and previous experiences, analyses things diferently. Some of the consequences of these diferent points of view are refected in how we communicate. In this project we have researched the way in which people produce descriptions to refer to someone in particular, and the details they notice in specific situations. We have developed algorithms that are able to reproduce this behaviour. There is already a lot of progress in the field of the Generation of Referring Expressions documented in books and experiments, and this project focuses on analysing people's descriptions in real life, everyday situations. We have created an application that uses all this knowledge, and we have demonstrated that it works by creating situations filled with objects and characters. With the algorithms we have designed, we have built descriptions which are easy to understand (using natural language in the way that a human would) and which allow the user to recognise any character easily with a low failure rate for any type of person. For this purpose, we have considered the details of the situation and of the character that has to be described, and we have designed algorithms that adapt to them. In order to generate the best situations to test our algorithms, we have created a large closed room full of people with the 3D engine Unity. This room is the canteen in the IT building in our university, and we have used a large variety of characters that are diferent enough from one another. With this we have been able to time the answers of real users and determine which descriptions work best in diferent kinds of situations. With the information we gathered from our surveys with volunteers, we have been able to improve our algorithms even more, creating the kind of descriptions that real peoplelike to hear.
For years, humans have tried to understand themselves and their way of thinking. This is not easy, because each person, due to factors such as their sex, age, native language, culture or their knowledge and previous experiences, analyses things diferently. Some of the consequences of these diferent points of view are refected in how we communicate. In this project we have researched the way in which people produce descriptions to refer to someone in particular, and the details they notice in specific situations. We have developed algorithms that are able to reproduce this behaviour. There is already a lot of progress in the field of the Generation of Referring Expressions documented in books and experiments, and this project focuses on analysing people's descriptions in real life, everyday situations. We have created an application that uses all this knowledge, and we have demonstrated that it works by creating situations filled with objects and characters. With the algorithms we have designed, we have built descriptions which are easy to understand (using natural language in the way that a human would) and which allow the user to recognise any character easily with a low failure rate for any type of person. For this purpose, we have considered the details of the situation and of the character that has to be described, and we have designed algorithms that adapt to them. In order to generate the best situations to test our algorithms, we have created a large closed room full of people with the 3D engine Unity. This room is the canteen in the IT building in our university, and we have used a large variety of characters that are diferent enough from one another. With this we have been able to time the answers of real users and determine which descriptions work best in diferent kinds of situations. With the information we gathered from our surveys with volunteers, we have been able to improve our algorithms even more, creating the kind of descriptions that real peoplelike to hear.
Description
Trabajo de Fin de Grado en Ingeniería Informática (Universidad Complutense, Facultad de Informática, curso 2013/2014)