Linguistic Reduction and Knowledge Graphs for Next-Gen Chatbots

Chatbots are dynamic agents with the express capability to engage in conversational interactions. By applying innovative linguistic reduction rules to user utterances, we empower chatbots to reduce any statement or sentence into its most basic form so bots can swiftly understand it and appropriately respond.

The relationship between linguistic reduction rules and chatbots for natural language technology applications is two-fold. First, this pairing drastically simplifies chatbot applications so that no matter what text or speech the chatbot encounters, they can readily understand and respond to it. Secondly, by adding elements of knowledge graphs and taxonomies to this tandem, the resulting combination can make chatbots more useful than any current commercial offerings —including Alexa and Siri.

Reductions Simplify Language

The general concept behind this symbolic reasoning approach is that when people speak or write they use more words than necessary to produce the simplest logical statement they’re conveying. For example, there are numerous ways to ask someone his or her name, including “Could you tell me your name, please?”

Reduction rules would reduce this simple question to “What is your name?”, so bots can quickly comprehend its meaning, then use additional techniques to answer it.

Although this example seems trivial, it illustrates the basic formula that’s integral for revamping a host of business use cases from analyzing legal documents to forms for regulatory compliance and heightening call center interactions—or any other NLP application.

Recommended: Klocked Announces Siri Shortcuts Feature

Less Enables More 

Linguistic reductions can be considered a rule-based approach, one of the foundations of the symbolic Artificial Intelligence approach to NLP. Although they’re manually created, their beauty lies in their universal applicability to any natural language processing use case. This utility significantly broadens when taxonomies are involved, but even without such hierarchical vocabularies, this approach works in any domain – from pharmaceutical to finance or any other domain. These rules are based on identifying patterns in language and reducing it to its bare minimum so chatbots or computers can easily understand them.

Consider the following sentence a customer might ask: ‘I just need to know if you’re open at the present time.’ Because reduction rules are recursive, bots can apply a series of them to this question.

For instance, one might state that anytime ‘I just’ is detected, it can be reduced to ‘I’. Or, phrases beginning with ‘I need to know something’ are reducible to ‘tell me something’. By applying these and other rules, this 12-word sentence (complicated by a contraction, adverbs, a prepositional phrase, and other superfluities) is reduced to ‘are you open now?”—a much simpler sentence.

NLP Technologies Update: Arthur Releases the First NLP Model Monitoring Solution To Serve Soaring Enterprise Adoption

Reduction Rules

Think of reduction rules as containing a left side—the input or all the exhaustive ways something can be expressed in language—and a right side or output containing the reduction in its simplest form.

Chatbots combine these rules with an additional set of rules for responses to questions, which is useful for online interactions with a retailer, vendor, or certain websites. However, the response rules could easily be replaced with rules about data privacy regulations, additional regulations, or legal implications for clauses in contracts and other documents.

Generating these rules requires training data or several labeled examples of common questions, phrases, and sentences found in natural language itself. However, the benefit is definite quality control in the bots’ responses, which are suitable for each of the specifically-reduced sentences. These controlled outputs are ideal for ensuring legal teams about the accuracy and respectfulness of chatbot interactions—which isn’t always guaranteed when only machine learning techniques are involved. Outputs may also contain a multitude of responses that mean the same thing for a given reduction, so bots have vibrant personalities.

Building Chatbot Intelligence 

The real magic of doing NLP with chatbots and reduction rules comes with combining them with domain-specific knowledge graphs powered by taxonomies. In this case, the questions customers ask, or the information found in legal documents, would serve as a query to run against a knowledge base to devise timely answers. The query itself would be stored in the knowledge graph (which relies on a uniform taxonomy) to constantly build knowledge to support highly sophisticated use cases, including anecdotal applications.

Whereas most chatbots breakdown after an initial sentence, storing each of the reductions in an episode about how a dishwasher stopped working, for example, as triples in a knowledge graph allows firms to query them for holistic understanding of everything in that story. Applying this technique to maintenance records of aircraft repairs, for example, could deliver insight into common repair scenarios and predictive maintenance to eliminate failure altogether.

By enhancing chatbots with linguistic reduction rules and knowledge graphs, it is possible to give chatbots truly conversational sophistication exceeding anything existent today.

About the Authors

Jans Aasman is a Ph.D. psychologist, expert in Cognitive Science and CEO of Franz Inc., an early innovator in Artificial Intelligence and leading provider of Semantic Database technology and Knowledge Graph solutions.

Dr. Richard Wallace is a knowledge engineer with over 35 years of experience in data science and artificial intelligence. He is a co-founder of Pandorabots, Inc., a leading chatbot A.I. company. In the 2000’s Dr. Wallace was a three-time winner of the Loebner Prize in artificial intelligence for the “most human computer.”

[To share your insights with us, please write to sghosh@martechseries.com]