| Resumen |
Learning relationships between nodes in a directed graph is a task that has been widely studied and it has been applied to a large number of topics and research areas. We establish a definition of particular kind of relationship, called analogy in a directed multigraph. An analogy can be defined for a certain pair of concepts, and the paths connecting them are called explanation of this analogy. We experiment with a structure built from real oneiric stories obtained from psychoanalytic descriptions (e.g. mother is represented as a bull; book represents power). Analogies found by the analysts are automatically identified by means of linguistically motivated patterns. Analogies have degrees of similarity based on the words used to describe them: represents, is a, is like a, can be a, refers to, etc. Once they are identified and graded, they are represented in the multidigraph, allowing us to provide different hypotheses in how these analogies can be explained. In order to enrich the concept graph, we added information from ConceptNet and WordNet. In addition, we propose a learning method for association rules that, given the degree of the analogy and a starting concept, allow reaching a destination concept. For example, starting from “dream”, we obtain the path <dream, psychic, neurosis, symptom>, being "dream is a symptom" a description previously given by a psychoanalyst, that was not included when training the algorithm. We evaluated 100 analogies on 171 concepts with 8,034 properties using Leave One Out cross validation, and found that the correct analogy was found within the all the possible paths for 94% of the analogies, restricted to 85% if only the top 20% possible paths are considered. This implies that, by using our method, it is possible to learn analogies between two concepts by reconstructing paths of different lengths based on local decisions considering concept, property and degree of analogy. |
