• No se han encontrado resultados

PROCESO DE SELECCIÓN Y RECLUTAMIENTO DE PERSONAL CONTENIDO DEL MANUAL

4.1. Manual de Reclutamiento, Selección e Inducción del personal

When looking at the underlying computational models to encode narratives and narrative spaces used as input for applications included in the previous enumeration, we often observed hybrid ap- proaches or applications that use different models to handle different parts of the generation and storytelling, still there are a few key approaches when it comes to computational models for narrative generation and storytelling. These are described below:

Planning-based: These models include logic and plan-like models representing a story space or the rules defining a simulation, narrative theory, agent behavior or author’s goals82;101;115;120. Despite some existing popular representations (such as extensions to PDDL), because of the specialization of the different systems, these tend to be ad-hoc, manually authored, and not reusable.

Frame-based: These models include a wide variety of frame-based representations of a story such as description logics or semantic networks to represent events in a story. These are used often in analogy-based and modification approaches83;102;121, and also tend to be manually authored and not reusable.

Plot-points: Plot-points usually represent the different events in a story into a branching or graph- like structure, where each node represents aplot-point(an important event in the story). These models are also usually domain-dependent and have been mostly used for drama manage- ment22;24.

Rewriting Systems: Similar to the previous, these models encode narratives in a series of hier- archical abstractions. These models implement rules similar to those of a formal grammar, which have been used to study cognitive and narratology theories73, and can be sampled for story generation122;123.

Shallow Annotations: These take the form of a templates or involve shallow annotations over an existing text or story. These are typically used along other models in the final realization of the narrative. Used mostly for text generation121;122, can be used by instantiating assets in other mediums (e.g., side-quests124).

Other Specialized Models: There are many other types of models focusing on modeling specific features or dimensions of a narrative such as the characters and their social relationships or the locations and the spatial configuration of the story environment6;90;125. These specialized models tend to be manually tuned for specific applications and may include rule-based systems or functions that maximize or minimize some desired target. These can also be used in con- junction with other models, usually through annotations and extensions that bind the models together22;121.

2.5.2

Narrative Analysis and Narrative Information Extraction

Automated narrative analysis and narrative information extraction has been used to study folklore and sociocultural phenomena. This area of research, related to the digital humanities focuses on modeling narratives in order to study literature or validate narrative theories. We identified several distinctive areas of research and application:

Annotation and Markup Languages: From plot to discourse, there are several components of interest in a narrative. Moreover, natural language is ambiguous and may encode several over- lapping features such as focalization, reader’s mental model, embedded narratives, character affect states, beliefs, desires and goals. There have been efforts to standardize the process of adding computer-readable annotations to natural language to allow computational narrative systems to process narratives26;126. There have been also work on annotation tools127;128. Despite these aiming to be general purpose, there is no consensus on annotation formalisms, and researchers develop ad-hoc solutions to suit their particular needs that are time-consuming and expensive to generate.

Narrative Information Extraction: A second body of work focuses on automatically extracting narrative information from text using natural language processing techniques. Most of this work focuses on extracting particular components of a narrative from a previously defined on- tology. These include automatically identifying characters, narrative and plot structure, and character relationships15;18;129. Related work also used crowdsourcing approaches10 and/or

commonsense databases and domain-specific knowledge to extract specific information130–132. Applications range from those tangentially related to computational narrative (such as doc- ument indexing and retrieval) or study, summarization and visualization of stories (using graphical and animated content)9;58.

Note that the work reported in chapters4,5and6 falls within this area of research.

Story Understanding: Story understanding (a.k.a. machine reading or story comprehension) extends the idea of narrative information extraction but goes beyond extracting information from stories and strives to understand the entire story in order to reason and answer questions about it133;134. These efforts usually involve linking the extracted information to ontologies and common-sense databases135. To the best of our knowledge, despite some early interest in this area136;137, recent work has focused on either story reasoning (ignoring natural language processing138), question answering from text (with limited inference and without a focus on broad understanding139) or knowledge representations for extracted story information134;135.

Automated Literature Analysis: This work builds upon narrative information extraction but is intended to visualize, summarize or capture, often not a single story but a specific set of stories or texts. For example, comparing the works of different authors128, validating narrative theo- ries84 or analyzing trends and recurring patterns (such as dramatic arcs in popular fiction85, story structure140or character interactions18;128).

Let us now describe the computational models of narrative used in narrative information extrac- tion and analysis work:

Shallow Annotations: Equivalent to the shallow annotation models described in Section 2.5.1, these are close to the text representation of the story, dependent on the discourse of the story but relatively straightforward to acquire automatically. These are the representation that narrative information extraction pipelines similar to Voz can extract and/or annotate and those can be used either as some intermediate/hidden model (for example, for indexing and retrieval13) or to provide a summary or visual representation of a story85;141.

Symbolic Semantic Models: These are rich models that often include semantic information that has been extracted and/or inferred from the text, such as those described when discussing

annotation and markup languages. The main difference from shallow models is that semantic models usually employ comprehensive ontologies that tie the different annotations together. Despite some limited work9;15;58, these are challenging to automatically extract from text and the most feasible alternative is to use manual or semi-automatic annotation tools127;128.

Catalogs and Other Specialized Models: These focus on modeling specific features or dimen- sions of a narrative (e.g., it is characters and social networks8;13) and are acquired by narrative information extraction pipelines that exploit off-the-shelf NLP. Despite being difficult to ex- tract mostly because of open problems in NLP these are used in a breadth of applications related to narrative information extraction and automated literature analysis. These models can take the form ofcatalogs (a list of locations7;142) or graph-like structures (a social network based on character interactions8;19;128). The biggest difference between these and semantic models is that these do not intend to be comprehensive and unlike shallow annotations, these are disconnected from the discourse (i.e., text of the narrative).

2.5.3

Statistical Models for Computational Narrative

As in many other areas of research within computer science, there is an emergent trend within com- putational narrative that involves applying statistical methods in different computational narrative tasks. For the specific task of story generation, Markov-models have been used to learn the prob- ability distribution of a specific corpus and have then been used to replicate the style of a specific genre or author70. Similarly, some examples in recent research explore the use of recurrent neural networks to both encode narrative events in a similar fashion as word embeddings encode text con- tent and then use long short-term memory (LSTM) recurrent neural networks for learning the event structure of stories from a corpus and generating new stories by sampling the learned model69. More recent work combine word embeddings, Markov models and recurrent neural networks to explore additional training and sampling methods for learning from a corpus and automatically generating

stories70;143.

The key difference between the approaches in this class and the approaches described in the previous sections is that the models described previously are interpretable and explicit, that is, human authors and system designers specifically choose which features of the narrative to model and how to populate the models’ contents (e.g., if we want to model the social network between characters, we have an explicit representation that we can look at that includes the characters in the story and a set of edges to define their social interactions). For the systems in this class of black-box approaches, however, the models that are used are not easily inspectable, it is hard to know what is being modeled, the author does not explicitly decide what will be modeled (besides defining the network structure) and it is up to the algorithms to find an internal representation of what needs to be modeled based on training data.

With the increase of computational power and the access to huge corpus of information, neu- ral networks have seen a drastic increase in their applications with successfully to many areas of research. Related to computational narrative, natural language processing and natural language generation, recurrent neural networks trained from text are being used in for tasks such as sum- marization, machine translation and conversational agents.67;67;68;144. The use in computational narrative systems in the context of digital entertainment is limited but there are some examples in recent research that explore the use of recurrent neural networks (e.g., LSTMs and word embedding representations)69;70.