Episodic Memory
OWL file
link to Episodic Memory ontology
Authors
Bart Schreuder Goedheijt
Short description of ontology
The ontology is based on the affection-based episode ontology. In this model, they combined Ontology-based unified robot knowledge (OUR-K) with temporal episode ontology in order to model event episode knowledge.
As the temporal episode ontology, this episodic memory ontology is built around the 5W1H (When, Where, Who, Why, What and How) principle of Han et al. (2013). Using a model that can understand Natural Language and can decompose an input text to match the concepts of 5W1H, information about an episode can be stored in the ontology.
Related work
report of this research: http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A1164181&dswid=7337
Ontology design
Ontology development
The goal of the research was to capture two types of memory: the augmented memory of the child (activity reports and observed behaviour) and the episodic knowledge of the PAL actor.
The module is built in different iterations. The goal for the first iteration is to implement the ability to memorize goals and progress in order that they can be used to give summarized feedback. This would increase the effectiveness as mentioned in the Goal-Setting Theory (see § 3.2). Besides the Goal-Setting Theory, the Nudge Theory could be used in order to push the behaviour change in the right direction (see § 3.8). The second iteration would include the memorization and reasoning over blood sugar measurements in order to improve intrinsic motivation (see § 3.1). The third iteration would contain the ability to memorize and reason over the activities that the children entered in the diary, which could be solved by using an Affection-based Episode Ontology (see § 3.6).
After building a foundation of knowledge about different theories and their relation with past experiences that are shared between the PAL actor and the child, a set of scenarios and use cases was developed in the specification phase. Based on these scenarios and use cases, requirements and claims were extracted.
Ontology structure
The Episode class is the main element that defines the memory. The concept of 5W1H (When, Where, Who, Why, What and How) is achieved using the relations. ‘Who’ is the child that has been mapped to the episode. ‘When’ is mapped in an Event, which is a subclass of the already existing ontology of a Happening. The ‘What’ is an Entity, that is also stored in the superclass Happening. The ‘How’ is mapped to the Happening in the form of an already existing class Activity, which involves a certain action like cooking, playing, or performing the quiz. The ‘Why’ was the most difficult part to map into an ontology. For now, it’s part of the EpisodeTag. The EpisodeTag has several instances which define the ‘What’ and implies the ‘Why’. For example, the EpisodeTag contains an instance ‘GoalWasTooDifficult’. The ‘What’ in the tag is a goal, but the specific instance of this goal is stored in the Event. In addition to the GoalWasTooDifficult, there are other instances like ‘GoalAchieved’ and ‘GoalAlmostAchieved’. These EpisodeTags are important, as they are used to define what triggers them. This is done using the relation to EpisodeTrigger. A trigger is, for example, ‘OnChildLogin’ or ‘QuizCompleted’. The relation between the Trigger and Tag is made to define when to mention a specific episode.
The current PAL ontology contains three different mappings of an emotion. They all have their pros and cons. To cope with this problem, they will soon be replaced by a new ontology of an emotional state, which would become the only mapping of emotion in the PAL domain. The emotion is currently mapped as the already existing class ‘Mood’, but it is in fact a placeholder.
This ontology is used in a more elaborated module where input is processed to gather memories and store them into episodes in the ontology. Based on the rules, each episode is tagged with possible triggers. If there is memory available for a specific trigger (e.g. child login), another part of the module, called the ‘conversation handler’ forms an action based on the stored information and passes this on to the user session (e.g. it refers to an episode).
Evaluation and results
The ontology of episodic memory as defined in this project (figure 5.3) was found suitable for storing memories about goals as defined in the first iteration. As both goals and measurements are part of the PAL ontology, the implementation of the second iteration using this ontology will not be a problem. The third iteration is, however, a bit tricky as it involves gathering episodes from an event that the child is talking about. Right now, the ‘Why’ in combination with a ‘What’ is stored in an EpisodeTag. It would be impossible to map all the possible cases that a child can talk about. The same applies to the ‘How’ which would require mapping all the actions in the ontology. This is however less of a problem than the EpisodeTag as an action is still defined atomically instead of as a combination of entities. The EpisodeTag should therefore be replaced by a separate ‘What’ and ‘Why’. The reasoning using the EpisodeTrigger should check the precondition based on multiple relations instead of the single EpisodeTag. This would make it possible to have a goal that was too difficult, but also, for example, a test at school that was too difficult. In both cases, the ‘Why’ is the same, as they were both too difficult. The moment to mention each episode could still be different, so just the ‘What’ or ‘Why’ would not be sufficient.
The current mapped Emotion is just a placeholder as a new ontology of the emotional state of a child is being created. As soon as this improved ontology is done, it should replace the mood in the episode ontology.