NATURAL LANGUAGE PROCESSING FOR STORYTELLING AND ROLE PLAYING: A TRAINING SYSTEM BASED ON THE PROPP MODEL
Politecnico di Milano (ITALY)
About this paper:
Appears in: ICERI2010 Proceedings
Publication year: 2010
Conference name: 3rd International Conference of Education, Research and Innovation
Dates: 15-17 November, 2010
Location: Madrid, Spain
Abstract:Storytelling and Role Playing are recognized as fundamental learning activities. Reading, understanding, writing and re-writing stories, playing different roles are considered powerful strategies for sharing human knowledge. Social skills, mental attitudes, self-consciousness, and relational confidence of kids and adults can widely improve when they have the opportunity to share interactive storytelling activities.
The system we propose is able to recognize the main actors of folk tales, providing a summary of the text, and permitting users to edit the story, in a controlled way.
The list of main actors, as well as the summary, highlight the structure of the story, allowing students to easily understand its plot. Once students understood the story, they can modify it, editing the parts of the text where main actors play their own roles. This way, students can create new stories, starting from the original one, while respecting the ground structure.
The tool is based on the Propp's theory of folk tales, where he argues that such stories share a common set of so-called "actors" and "actions". As "actors" one can think of human entities like "hero" and "villain", but also of inanimate things like "magical helper". Actors performs their own well specified set of "actions", in a well specified sequence.
Following the Propp's approach, we use Natural Language Processing techniques to associate a set of verbs to different actions and to associate such verbs to the actors that can perform them.
Our model is composes of a set of rules that translates the Propp's abstract model in a computer-ready description. The rules predicate on the syntactical structure of the text, extracted by means of a dependency parser, and try to find triples as (subject(word_1), verb(word_2), object(word_3)).
As our model specifies the actor that is likely to perform a given action onto a given target actor, and the set of verbs that lexicalize such action, the algorithm is able to associate the most likely actors to the subject word_1 and the object word_3.
Such association is performed for each sentence and, in case of contradictory associations, a resolving procedure is applied.
A prototype has been implemented as a web site. At run-time, students upload a text, the system runs the parser, applies the rules and calculates the set of words that seem to impersonate the various roles, as well as the set of sentences where such roles perform their actions. These sentences represent the summary of the text, and are presented to the student in a web form where she/he is invited do edit them, thus modifying the story while preserving the general structure of the tale.
As the proposed model is not based on word frequencies (excepting for the "hero"), it is able to capture roles even if the related words are not frequent in the text. In our experiments, we compared our algorithm against pure statistical predictors, observing that our approach significantly outperform them.
Future works will be devoted to generalize the idea of "actors" and "actions" on a generic text, defining a rule-based meta-model. Moreover, we plan to learn such model from data, applying machine-learning techniques. Finally, future works will be done to make the system accessible to students with cognitive disability, as they (but also their parents, teachers or therapists) could appreciate the solution we presented for its possible impact on the learning activities they usually share.
Keywords: NLP, summarization, Propp, storytelling, role playing.