Herramientas de generación de historias narrativas basadas en mapas personalizados
Loading...
Official URL
Full text at PDC
Publication date
2023
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
La generación automática de historias es un campo de estudio dentro del campo de la Inteligencia Artificial que tiene como objetivo generar relatos consistentes y coherentes mediante mecanismos automatizados. En este trabajo se pretende generar una herramienta capaz de crear historias narrativas con suficiente profundidad para ser útiles para escritores y creadores de todos los ámbitos, a partir de un mapa geográfico/político. Un escritor creador de mundos normalmente cuenta con una idea general de lo que quiere conseguir, así como un mapa de su mundo, pero puede ser tedioso pensar en distintas historias para cada uno de sus territorios. Nuestro proyecto pretende ofrecer una forma de generar una primera lluvia de ideas de forma totalmente automática y veloz.
En este proyecto realizamos un estudio de los distintos avances en el campo del aprendizaje automático, modelos de lenguaje y distintas técnicas relacionadas, generadores de mapas presentes en el mercado, así como generadores de historias. Mostraremos como estos últimos no presentan la posibilidad de utilizar un mapa para generar historias, opción que nuestro proyecto busca ofrecer.
Para la realización de este programa concebimos un plan analizando las distintas tecnologías del mercado. Posteriormente, realizamos un primer procesado de los datos del mapa y un primer prototipo. Al no conseguir el resultado esperado realizamos una mejora de dicho prototipo mediante distintos mecanismos, que incluyen: una puntuación de los resultados obtenidos para seleccionar el texto con mejor nota, prevención de elementos indeseados para un texto narrativo histórico, fine-tuning de uno de los modelos e investigación sobre mecanismos de cohesión.
Los resultados obtenidos al final del proyecto son mucho mejores de los obtenidos en nuestro primer prototipo, lo que demuestra la eficacia de nuestros mecanismos. Sin embargo, el trabajo realizado es altamente ampliable, pudiendo obtener muchos mejores resultados con la utilización de modelos mucho más modernos y actuales a los utilizados en nuestro desarrollo. Los mecanismos desarrollados hasta ahora mejorarán aún más cualquier resultado obtenido, permitiendo continuar en el futuro con la parte del proyecto correspondiente a la cohesión entre historias.
Automatic story generation is a field of study within the eld of Artificial Intelligencethat aims to generate consistent and coherent narratives through automated mechanisms. This project intends to generate a tool capable of generating narrative stories with enough depth to be useful for writers and creators from various domains, based on a geographic/political map. A world-building writer typically has a general idea of what they want to achieve, as well as a map of their world, but it can be tedious to think of different stories for each territory. Our project aims to provide a way to generate an initial brainstorming session automatically and swiftly. In this project, we conducted a study of the different advancements in automatic learning, language models and related techniques, market-available map generators, as well as story generators. We will demonstrate how the latter do not offer the possibility of using a map to generate stories, which is an option that our project seeks to provide. To develop this program, we devised a plan by analyzing the different technologies available in the market. Subsequently, we performed an initial processing of the map data and created a first prototype. As we did not achieve the expected results, we improved the prototype using various mechanisms, including: scoring the obtained results selecting the result with the best score, prevention of undesired elements for a historical narrative text, fine-tuning one of the models, and researching cohesion mechanisms. The results obtained at the end of the project are much better than those obtained in our first prototype, which demonstrates the effectiveness of our mechanisms. However, the work carried out is highly expandable, as much better results can be obtained by using much more modern and up-to-date models than those used in our development. The mechanisms developed so far will further enhance any achieved result, allowing for the continuation of the project's cohesion between stories in the future.
Automatic story generation is a field of study within the eld of Artificial Intelligencethat aims to generate consistent and coherent narratives through automated mechanisms. This project intends to generate a tool capable of generating narrative stories with enough depth to be useful for writers and creators from various domains, based on a geographic/political map. A world-building writer typically has a general idea of what they want to achieve, as well as a map of their world, but it can be tedious to think of different stories for each territory. Our project aims to provide a way to generate an initial brainstorming session automatically and swiftly. In this project, we conducted a study of the different advancements in automatic learning, language models and related techniques, market-available map generators, as well as story generators. We will demonstrate how the latter do not offer the possibility of using a map to generate stories, which is an option that our project seeks to provide. To develop this program, we devised a plan by analyzing the different technologies available in the market. Subsequently, we performed an initial processing of the map data and created a first prototype. As we did not achieve the expected results, we improved the prototype using various mechanisms, including: scoring the obtained results selecting the result with the best score, prevention of undesired elements for a historical narrative text, fine-tuning one of the models, and researching cohesion mechanisms. The results obtained at the end of the project are much better than those obtained in our first prototype, which demonstrates the effectiveness of our mechanisms. However, the work carried out is highly expandable, as much better results can be obtained by using much more modern and up-to-date models than those used in our development. The mechanisms developed so far will further enhance any achieved result, allowing for the continuation of the project's cohesion between stories in the future.
Description
Trabajo de Fin de Grado en Desarrollo de Videojuegos, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Curso 2022/2023.