Month: February 2011

A discussion that began with a graduate scholar at George Washington University in November, 2010 about semantic software technologies prompted him to follow up with some questions for clarification from me. With his permission, I am sharing three questions from Evan Faber and the gist of my comments to him. At the heart of the conversation we all need to keep having is, how far does this technology go and does it really bring us any gains in retrieving information?

1. Have AI or semantic software demonstrated any capability to ask new and interesting questions about the relationships among information that they process?

In several recent presentations and the Gilbane Group study on Semantic Software Technologies, I share a simple diagram of the nominal setup for the relationship of content to search and the semantic core, namely a set of terminology rules or terminology with relationships. Semantic search operates best when it focuses on a topical domain of knowledge. The language that defines that domain may range from simple to complex, broad or narrow, deep or shallow. The language may be applied to the task of semantic search from a taxonomy (usually shallow and simple), a set of language rules (numbering thousands to millions) or from an ontology of concepts to a semantic net with millions of terms and relationships among concepts.

The question Evan asks is a good one with a simple answer, “Not without configuration.” The configuration needs human work in two regions:

• Management of the linguistic rules or ontology
• Design of search engine indexing and retrieval mechanisms

When a semantic search engine indexes content for natural language retrieval, it looks to the rules or semantic nets to find concepts that match those in the content. When it finds concepts in the content with no equivalent language in the semantic net, it must find a way to understand where the concepts belong in the ontological framework. This discovery process for clarification, disambiguation, contextual relevance, perspective, meaning or tone is best accompanied with an interface making it easy for a human curator or editor to update or expand the ontology. A subject matter expert is required for specialized topics. Through a process of automated indexing that both categorizes and exposes problem areas, the semantic engine becomes a search engine and a questioning engine.

The entire process is highly iterative. In a sense, the software is asking the questions: “What is this?”, “How does it relate to the things we already know about?”, “How is the language being used in this context?” and so on.

2. In other words, once they [the software] have established relationships among data, can they use that finding to proceed – without human intervention- to seek new relationships?

Yes, in the manner described for the previous question. It is important to recognize that the original set of rules, ontologies, or semantic nets that are being applied were crafted by human beings with subject matter expertise. It is unrealistic to think that any team of experts would be able to know or anticipate every use of the human language to codify it in advance for total accuracy. The term AI is, for this reason, a misnomer because the algorithms are not thinking; they are only looking up “known-knowns” and applying them. The art of the software is in recognizing when something cannot be discerned or clearly understood; then the concept (in context) is presented for the expert to “teach” the software what to do with the information.

State-of-the-art software will have a back-end process for enabling implementer/administrators to use the results of search (direct commentary from users or indirectly by analyzing search logs) to discover where language has been misunderstood as evidenced by invalid results. Over time, more passes to update linguistic definitions, grammar rules, and concept relationships will continue to refine and improve the accuracy and comprehensiveness of search results.

3. It occurs to me that the key value added of semantic technologies to decision-making is their capacity to link sources by context and meaning, which increases situational awareness and decision space. But can they probe further on their own?

Good point on the value and in a sense, yes, they can. Through extensive algorithmic operations, instructions can be embedded (and probably are for high-value situations like intelligence work), instructing the software what to do with newly discovered concepts. Instructions might then place these new discoveries into categories of relevance, importance, or associations. It would not be unreasonable to then pass documents with confounding information off to other semantic tools for further examination. Again, without human analysis along the continuum and at the end point, no certainty about the validity of the software’s decision-making can be asserted.

I can hypothesize a case in which a corpus of content contains random documents in foreign languages. From my research, I know that some of the semantic packages have semantic nets in multiple languages. If the corpus contains material in English, French, German and Arabic, these materials might be sorted and routed off to four different software applications. Each batch would be subject to further linguistic analysis, followed by indexing with some middleware applied to the returned results for normalization, and final consolidation into a unified index. Does this exist in the real world now? Probably there are variants but it would take more research to find the cases, and they may be subject to restrictions that would require the correct clearances.

Discussions with experts who have actually employed enterprise specific semantic software, underscores the need for subject expertise, and some computational linguistics training coupled with an aptitude for creative inquiry. These scientists informed me that individuals, who are highly multi-disciplinary and facile with electronic games and tools, did the best job of interacting with the software and getting excellent results. Tuning and configuration over time by the right human players is still a fundamental requirement.