Tagging And Action Graph Generation For Recipes

Abstract

Processing instructions is significant to accomplish daily tasks. Instructions can be found in many different forms for a variety tasks. Machine understanding of instructions, similarly, can be beneficial for artificial agents/robots to perform a task automatically. Building systems checking if a task is carried-out in conformity with the instructions is important for many mission critical tasks, for instance factories, workers who repair the electronic devices etc. In this thesis, it is aimed to automatically extract the steps of a certain task with the aid of instructions. Instructions dataset is needed to train model and extract the steps of a task. Recipes are the examples of instructions that are easy to follow and can be found in large quantities. Understanding of how to cook a recipe step by step requires extraction of course of actions, ingredients, tools and, the relationships between each other through Natural Lan- guage Processing (NLP) Techniques. Supervised and rule-based model is proposed to clarify and extract actions and components. Instead of a fully supervised method, NLP Techniques are used to find relations between components and actions in the text. The workflow of the recipes are finally produced by a rule-based method. When compared to a state-of-the-art unsupervised method which models the task as a whole, the proposed method benefits from the output of smaller and well-studied NLP Techniques.