Roger Schank did some interesting work on language processing with Conceptual Dependency (CD) in the 1970s. He then moved somewhat out of the field, being in Education these days. There were some useful applications in natural language generation (BABEL), story generation (TAILSPIN) and other areas, often involving planning and episodes rather than individual sentences.

Has anybody else continued to use CD or variants thereof? I am not aware of any other projects that do, apart from Hovy's PAULINE, which uses CD as representation for the story to generate.


2 Answers 2


Is anybody still using Conceptual Dependency Theory?

Yes. Many people. Conceptual dependencies are central to the conveyance of ideas in natural language.

Here are just a few publications in this century building off of Schank's work or travelling in parallel with his direction in related fields.

I met Roger Schank in Hartford, in 1992, during a lecture series sponsored by the AI labs of United Technologies Research Center and a few other Fortune 500 companies in the region. His entire lecture was a series of stories in AI research. I remember every story 26 years later.

The toy NLP implementations you see in the field today pale in comparison with the story based reasoning and memory systems proposed by Dr. Schank as a probable explanation of observations that can be made about human vocal communications.

It is easy to guess the reason he moved into education. His natural language and artificial intelligence ideas were about a century early and over the heads of most of the people that were at the lecture alongside me.

If you and I find his story-based reasoning and memory proposals compelling, we are probably a century too early and a bit over the heads of most in the present day NLP field. Most of those in labs in the 1980s found Schank irritating, and people who fit comfortably into today's technology culture find him irrelevant.

Some of those I interacted with on a project from the University of Michigan in Ann Arbor don't find his work irrelevant though, and their work is in the directions he indicated. Unfortunately the client NDA restricts me from commenting further about that project.

The reason we should not and ultimately will not abandon the idea that we communicate in stories is because it is correct. When a person says, "It makes me want to puke," or, "I love you too," the direct parse of those sentences using "modern" techniques are not closely related to a correct reconstruction of the idea in the mind of the speaker. Both sentences reference a conceptual heap of interdependence that we call a story.

If two "party girls" are in the ladies room at a Borgore concert and one says, "Hand me a roll," the interpretation of the word, "roll," is conceptually dependent. If the speaker is in a stall it means one thing. If at the sink it means another.

There will always be some segment of the research community that understands this. Those that do not may construct money-saving automatons that will answer your business's phone calls, but they will not give you a heads up on a customer relations pattern that points to a policy issue.

These toy NLP agents, until they develop the capabilities Dr. Schank proposed, will not recognize from phone conversations with clients that a product or service enhancement is an opportunity waiting to be exploited, and they won't tell you a story that will convince you that you would benefit from being the first to capitalize on the opportunity.


Although this model played an important role in contributing to our present understanding of NLP and NLU, it is no longer useful in production systems and currently no successful commercial product follows this approach.

In CDT the goal was to design an AI system that could draw logical inferences from sentences. In this system the goal was to make the meaning independent of the words used in the input.

CDT modeled sentences by using tokens such as: locations, time, real world actions and real world objects. However as computational power became more common and less expensive, interest diverted to statistical models which were now outperforming the previous rule based systems.

The problem with rule based approaches such as CDT is that they require manual development of linguistic rules which can be costly and which usually don't generalize well to other languages.

On the other hand, statistical approaches use human language resources (multilingual textual corpora) more efficiently. Rather than using a rule based approach, statistical models make soft probabilistic decisions based on attaching real weights to the features making up the input data. (Wikipedia NLP)

This efficient use of human language resources leads to a model that is more accurate and robust especially when given unfamiliar input or input that contains errors. Statistical models also generalise well to other languages.

  • $\begingroup$ Thanks for your reply; I am aware of statistical models and their properties, but for this question I was interested in CDT only! $\endgroup$ Jan 11, 2018 at 9:11
  • $\begingroup$ The topic was a joy to research and answer. I introduced statistical models at the end for comparison, however I totally get your point. $\endgroup$
    – Seth Simba
    Jan 11, 2018 at 9:40

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .