Agolo PY24 TACFI: Entity Analytics
AFWERX · AFWERX TACFI · AFWERX
Award
Description
One consistent challenge faced by today’s warfighter is the need to quickly understand more about entities (organizations, people, technologies, programs, etc.) mentioned in text. But beyond simple descriptions, the AF user needs to understand those entities in the context of their analysis - to get the “gist” of how that entity plays a role in what they are researching. The AF user may also be inexperiencedas compared to a senior level analyst or subject matter expert and miss crucial relevant data. The process Reviewing these experts is time-consuming and error-prone. NER today is a fairly mature process, at least for typical entity types (e.g. people, organizations or locations). But, simply identifying an entity type is not sufficient for complex analyses. Imagine a system identifying a “William Smith” in text. Is it Will Smith, the actor? William Smith, the former US Attorney General? William Smith, the former NHL goalie? Or one of thousands of people who share that name? Early approaches used domain-specific rules engines or models to build entity linking for specific entity types and contexts. But this approach doesn’t scale well. During the Phase II, Agolo built an Entity Linking model that leverages four different algorithms, improving the results based on context. Both DAF and DOD users have expressed interest in our expanded capability to discern relationships between identities. At present, Agolo supports the seeding of predefined explicit relationships, which exist by default in public identities coming from Wikidata and other sources. Beyond these predefined relationships, Agolo presently extracts implicit relationships, without “understanding” the type of relationship. For example, if a sentence said Mary Jones and her co-worker, Bob Smith, attended the meeting, the system would identify that there is a relationship between Jones and Smith, although it is not able to determine the appropriate representative label to assign to that relationship. While knowing the existence of a relationship between two entities is valuable, explicit relationship extraction provides the context of how those two identities are related. Knowing that can inform whether a relationship is meaningful or irrelevant to the analysis. This context also enables complex analytics, and makes the number of unique relations between any two entities countable and therefore comparable. This capability also extends to improve Agolo’s contextual disambiguation and identity linking capabilities, and allows for the representation of all identities in a visual graph representation as nodes connected to each other. Agolo will perform a significant effort to ensure that its platform can support the potential scale of DOD customers. Agolo has done performance testing with a knowledge graph with 1 million documents and 3 million unique identities. The performance testing will increase that to 3 million documents and 5 million+ unique identities.