Group Dynamic Analysis and Multi-Modal Ontologies

Task Objectives

Human intelligence in the form of messages, blogs, reports, emails, etc., provides invaluable information for any COA. Event Ontology Driven Predictions from Text proposes to enhance the model of a current situation by analyzing textual sources for current evidence about known threats, enemy activities and enemy capabilities in order to combine lexical, semantic, and statistical properties of the knowledge into a text driven situation model. Lymba’s EventNet, an ontological network of WordNet events augmented with corpus driven likelihood statistics will be adapted to the military domain, and will become a vehicle for determining ranked potential actions to decision makers.

Figure 1 Situation Modeling and COA Predictions

Figure 1 Situation Modeling and COA Predictions

The key innovations of this work include: 1) Automatically extracting semantic structure from textual sources using Lymba’s NLP tools, 2) Tailoring Lymba’s existing event ontology resource, EventNet for COA predictions, 3) Scoring and ranking potential actions and outcomes for use in a Multi-INT tactical assessment tool, and 4) Transforming the knowledge inside the EventNet COA into RDF triples for WC3 compliant access by third party tools. Lymba is uniquely qualified to accomplish this task due to its mature natural language processing (NLP) tools that extract semantic structure from unstructured documents and organizes them into an ontology. In the last 10 years Lymba has participated in research programs like AQUAINT, NIMD, CASE, GALE and others. The NLP tools includes tokenization, part-of-speech tagging, named entities, events, syntactic parsing, semantic relation extraction and automatic ontology building. Lymba has also built EventNet, a resource consisting of all events from WordNet linked through probabilistic causal and other dependencies links.

Figure 1 illustrates the process of analyzing human intelligence to produce a ranked list of suggested actions based on the current situation. Given a situation task description, a situation modeling tool translates the semantics extracted from the description into a semantic representation. This is then linked to the knowledge extracted from the textual stream of human intelligence in order to predict a ranked list of the most likely actions based on the event ontology encoded in EventNet. As new documents, news wires, messages, and other intelligence are merged into the COA Event repository, existing situations are monitored for change in the status of their main events and actors and consequently any associated predictions. In Figure 2, the event “kill” from WordNet is decorated with additional cause and effect relations that are extracted from text. Each has relation has a probability associated with it based on the number of times this relationship was seen in text.

Figure 2: An EventNet Frame for Kill

Figure 2: An EventNet Frame for Kill

Under this work Lymba will develop algorithms to 1) model a dynamic situation in a semantically rich ontology backed representation of capabilities, threat, intentions, and actors, 2) tailor EventNet to the COA domain, 3) compute the most likely actions based on the current situation model and output them in a ranked list for use by other tactical assessment tools an 4) transform the knowledge captured in the COA EventNet into RDF triples. The key innovations include: 1) A COA tailored probabilistic event network for predictions, 2) a rich semantic representation of textual intelligence sources, and 3) dynamically adapting ontology backed repository of human intelligence for COA.

Task 1: Semantic Situation Modeling for Adversarial Predictions

Lymba will leverage its existing ontology building and natural language processing tools to encode a situation description into a semantic representation that includes, named entities, events, and the semantic relations that exist between them. The relation set is composed of those available from Lymba’s semantic parser and include, cause, agent, theme, instrument, is-a, part-whole, intention, location, and temporal. The resulting semantic representation will be used to query a COA EventNet in order to determine the most likely next actions to take.

Task 2: Tailoring EventNet for the COA Domain

Lymba’s EventNet is made up of WordNet events and relations that are enhanced with semantic relations extracted from web documents by Lymba’s semantic parser. EventNet is more than 100,000 events linked network. The connections between events are probabilistically weighted based on the number of times they are extracted from the web corpus. To be effective in the COA domain, EventNet needs to be customized with COA documents, reports, blogs, and any other human intelligence that is in textual form. Lymba will develop algorithms to extract semantic information from COA sources and update EventNet to operate in this domain. As new intelligence is made available, the event and relation structure of EventNet need to evolve. Algorithms will be developed to update the values of its event edges as new events, entities, and semantic relations are extracted from the stream of textual intelligence that is available to COA decision makers.

Task 3: Predicting and Ranking Actions for a COA Decision Maker

Lymba will explore algorithms that compute the actions to suggest to a decision maker or a tactical assessment tool. The key events need to be identified and the relations and participants in the events extracted. Each event structure inside the situation model will be instantiated with information from the COA EventNet ontology resulting in an event driven sub-model, where the edges each have a likelihoods associated with them. From this structure all the possible events and outcomes that are computed using the COA EventNet will be linked and scored. If for example a killing event was extracted by consulting the EventNet frame in Figure 2, injuries, death, emergency call, and rescue operations are all prediction for the analyst to consider, with injuries having the highest probability. Low scoring event outcomes can be used as clues that more information needs to be gathered, and the system can be extended to output suggestions for pro-active intelligence gathering based on edges in the graph where evidence was not fulfilled.

Task 4: Transforming COA EventNet into RDF

In order to make the semantics encoded in the COA EventNet built in Task 2 readily available to third party reasoning and tactical assessment engines, Lymba will develop algorithms to map the semantic representation of the situation models, which includes the semantic relations, concepts, events, and likelihoods, the concepts into the W3C standard for the Resource Description Framework (RDF). Lymba already has a complete RDFS schema for the information extracted by it tools and would extend this to work with the information extracted for situation models.

B. Technical Summary and Task Deliverables

Event Ontology Driven Predictions from Text will exploit textual human intelligence for the purposes of informing COA decision makers. Situation descriptions will be transformed into a semantically rich ontology backed representation from which key events, relationships, and actors will be identified as input to an action prediction engine. The action prediction engine will leverage a COA tailored event ontology, EventNet in order to compute and rank a list of suggest actions. Finally all the knowledge capture in the event ontology will be encoded in RDF triples so as to make it readily available to third party reasoning and tactical assessment tools. The deliverables for this project include algorithms to: 1) Model situations in a semantic ontology backed representation, 2) Customize EventNet for the COA domain, 3) Predict and Rank actions for a COA decision maker, and 4) Transforming COA EventNet into RDF triples.