PROPUESTA DE LA INVESTIGACIÓN
CURSO INTERMEDIO DE AUDITORÍA
Earlier in this dissertation, in Sections2.4and2.6, we outlined two of the main lines of research within computational narrative: 1) storytelling and story generation applications; and 2) analyzing or model existing narratives in order to study existing literature or validate narrative theories. Despite both being actively developed, there are very few instances of work across both areas of research. In Section2.5we surveyed the different computational models used within different approaches. Figure
Short, Simple, Single-domain Rich, Complex, Open-domain Full-fledged Computational Narrative Systems Support Role (Space Gen., Author Aids, Dialogue) Visualizations (Text-to-image, animations) Planning, Frame-based Semantic Markup, Ontology
Shallow Annotations, Catalogs, Verb Frames
[Elson, Finlayson] [Coyne, Johansson] [Finlayson] [Reagan] [Hartsook, Valls-Vargas] [Elson, Srivastava] [McCoy] [Meehan, Riedl] Multi-paragraph, Limited-domain [Riloff]
Narrative Analysis and Narrative Information Extraction
Narrative Generation and Storytelling
Numerical Models
[Valls-Vargas]
[Li & Riedl]
Intermediate Models of Narrative
Input Text Final Output
Figure 7.1: Overview of input and output in computational narrative applications in the main two areas of research within computational narrative. Black arrows indicate existing approaches (solid arrows indicate automated approaches, dashed arrows indicate manual and semi-automatic approaches). The pink arrow shows the related work by Li et al.10, and the blue arrow indicate our work for text-based end-to-end computational narrative systems as described in Section7.2; these are highlighted since they are end-to-end approaches combining both areas of work.
7.1 provides a brief visual summary of this work. The left-hand side of the figure represents work on narrative analysis and the right-hand side represents work on narrative generation. The four vertical axes represent the complexity of the input and output of these systems. The center of the figure contains a set of computational models of narrative employed by the different pieces of work shown in the figure and the existing or potential mappings between them that we discuss in the next section.
7.1.1
Mapping Computational Narrative Models
Despite the lack of consensus on the most appropriate models to encode narratives or represent different features (or even human narrative mental models)75, we observed some commonalities between the underlying models used in these different areas of computational narrative (those used in narrative generation and those used in narrative analysis described in the previous section), which could lead to the construction of end-to-end systems (where by “end-to-end” systems, we mean computational narrative systems that extract narrative models directly from text and use those to generate new narratives) that use narrative analysis techniques to generate the models required by narrative generation techniques.
For example, mappings between models based on shallow annotations has been explored in specific support modules within computational narrative applications (such as text generation using text modification3;4;175) but do not truly enable the creation of end-to-end systems. We can propose a few other unexplored intuitive mappings, that is, we can establish a mapping between the output models acquired by information extraction systems (shown on the left in Figure7.1) and the models required as input by story generation and storytelling systems (shown the second from the left in Figure 7.1), instead of having to hand-author them. For example, we could use other algorithms to extract specialized models such as: social networks and character relationship information19;128 for a social simulation125 or character-centric multi-agent story generation system104;176; or use a sequence of events and extracted location information to generate virtual environments for a story to happen7;90; or; use automatically annotated text as templates for natural language generation systems.
The most coveted mapping would be between automatically extracted rich semantic annotations or even symbolic models of a narrative to planning/frame-based models. Since these planning and frame-based models are often manually crafted, the required features for storytelling or story generation applications are manually crafted as well. There are, however, a few examples of work that attempt to build end-to-end systems. An example is the work of McIntyre et al.114. They attempt to automatically extract semantic frame knowledge from a corpus of stories, sample their knowledge base, assemble a frame-based story representation and use a text realization component to generate text. They then use a generate-and-and-rank iterative process with an interest and a coherence model to refine the generated output. More recent work by Li et al.10 relies on a set of crowd-sourced short reports describing a given theme (e.g., bank robbery). They extract Shankian script-like structures15 from each and assemble them into aplot graph that joins common events. These graph-like structures are related to a set of plot points or planning operators used to describe a story space. Despite not being full-fledged symbolic representation nor completely eliminating human intervention, this approach could exploit existing text reports.
Other examples include text-to-image systems that extract a semantic representation from short,
simple and limited domain text. Similar than the previous work, there is a mapping step that compiles and links the different extracted representation components (e.g., the catalog of participants and the sequence of events) with commonsense and domain-specific databases. Then, instead of handing this internal model to a storytelling system, it is used to generate images and/or animations for the final user9;58. We identified some limited work related to interactive learning systems and digital entertainment with similar characteristics104;176.
In our survey, we have seen some research efforts in the direction towards end-to-end systems. Currently, these seem to be limited in terms of the kind of text they can process (i.e. only short simple text), their scope (i.e. only domain-specific applications) and/or their output (can only render a limited inventory of objects or generate a small specific set of stories)9;10;58. Another prominent class of examples implementing end-to-end computational narrative systems includes de statistical approaches described in Section2.5.3which include Markov models and neural networks. These approaches and models can be used to implement end-to-end systems that connect narrative analysis and generation tasks, specially in text-based systems. Some common downsides of these methods are the often large training data requirements68 and the fact that there are still open research challenges in keeping character, word, and sentence level neural networks coherent for story generation69. Moreover, the lack of explainability and the fact that the learned models are often difficult (if possible at all) to tweak by human authors is in conflict with the often desired property of authorial control over the output177.