Behavioral Software Engineering was suggested as a more general term by Ken Stephens. I have to agree
Computational Analysis of Verbal Behavior (CAVB) is (or should be) the computer based analysis and contingency management of verbal behavior, particularly text-emitting behavior. It corresponds to (or would if anyone were doing it) the "cognitive" areas of Computational Linguistics and Natural Language Processing (NLP), which can be considered subfields of Artificial Intelligence.
My particular interest was in the arena of search software on the Internet. While I have training in Behavior Analysis, I have never worked as as a Behavior Analyst, nor do I expect to. I have worked as a software engineer for a couple of decades, and have focused on search software for most of that time.
One tactic taken by developers of search software is to use Natural Language Processing (NLP) to improve search results. The clicking behavior and text-emitting behavior of the user is analyzed to determine what responses on the part of the system are most likely to be reinforcing to the user, and which are therefore most likely to result in reinforcers for the information provider. NLP and other computational techniques can, at the same time, reduce costs to the search provider by reducing the amount of searching done by users. The sooner the user is presented with the appropriate results, the sooner the user stops searching (searching is VERY expensive in terms of hardware).
Another tactic is to use the same kind of information (clicking, textual), but recorded across time, to learn what might be reinforcing for particular users or groups of users given a particular location in the information service, given particular responses, particular times of day, week, month, year, etc. As users interact with such a system, the system provides increasingly dense reinforcement schedules.
A Web site could be considered an operant chamber. In many ways it resembles a pigeon box. Users click buttons, icons, whatever. Users are presented with programmed stimuli in response (web pages, text, pictures, music, video, merchandise, etc.). Users consume reinforcers (buy products, read or download articles found by search, etc.).
The situation is much more difficult than any pigeon experiment of which I am aware (even the social behavior experiments), and is difficult in many ways. One major problem is the nature of the responses of the most interest: text-emitting verbal behavior. Computer software is notoriously bad at analyzing verbal behavior and determining appropriate responses. There is software that works to a degree, but it is all far from perfect (ignoring the marketing hype of the software vendors themselves).
Ken Stephens reminded me of his company: Operant WebSites.
The total I have found to date: none. Zero. Zip. No one.
Where have I looked?
If you know of anyone who might be considered a practitioner, please send email to me at tomd@mactom.com.
Two possibles: Bill Hutchison and Ken Stephens. Would you guys consider yourselves to be the first CAVB folks? Anyone else fit here? Ken suggested looking up Chris Cherpas; here is what I have found so far: "Why I Stopped Reading Verbal Behavior (and Continue to Study Perceptual Control Theory)".
My group at AOL will almost certainly end up hiring a traditional cognitive NLP person. How long have Behavior Analysts been studying verbal behavior? How is it that cognitive folks still own such a lucrative area? How can we change this situation?
I believe the opportunities will grow with the Internet. It is fairly certain that whomever we hire as our NLP person within my search group will be a multi-millionaire within two to ten years. The same kind of situation might lead to greater wealth sooner (but with higher risk) at a smaller, newer, ".com" company. Applied folks: tell me you don't care about being independently wealthy. Academics: maybe your students who go applied and become affluent will show their appreciation via tax deductible contributions to your institutions or programs. Might just affect the health and well being of Behavior Analysis as a whole, don't you think?
One difficulty I see in "growing" CAVB folks is the multi-disciplinary nature of the work. At minimum, it involves Behavior Analysis and Computer Science. Few people are interested in both. Fewer still are competent in both, and the competencies can be difficult to develop. That has been one of the difficulties in finding even a traditional cognitive NLP person. We need someone who is a credible software engineer in addition to having credentials (i.e., degrees) or a track record in Computational Linguistics and NLP.
There is quite a bit of linguistic software out there. Some commercial software vendors:
There is also a lot of "learning" software out there, some based on neural nets, some not. Some of the more interesting systems are classified by their vendors as "knowledge management" or "personalization" software, such as those from Autonomy, Inc. and ApTex, now eHNC. It all works, to a degree.
Linguistic software: To be credible, a CAVB person would have to know the traditional software inside out: the algorithms, techniques, companies, schools, research areas, the successes and failures.
Software Engineering: To be credible and useful in our environment, the person would also need computer related experience that included several of the following: object oriented analysis and design; C/C++; Java; Unix; the usual range of Unix development tools (compilers, debuggers, profilers, memory checkers, code management, editors, etc.); client/server development with an emphasis on server development; distributed multi-tier software; software lifecycle; multithreading; design patterns; performance and scaling techniques; strong communication and interpersonal skills; and so on. Few, if any, software engineers have all the skills we want, so there is a lot of learning on the job, classes, etc. My group only hires folks who either have five or more years full time software development experience, or a Ph.D. in a relevant area. (Other groups within the Search Divlet are not quite as demanding. HR says we are hard to shop for.)
Take a look at the most recent sections of my work history for more of an idea what we have been doing.
I'd love to see a paper describing, in practical terms, what differences a behavioral approach might make in the areas currently dominated by linguistic software and natural language processing software. How would the software differ from what is currently on the market? How would its usage differ? How would functionality differ? Performance? Accuracy? Would it "learn"? In what areas might Behavior Analysis make the largest, and most marketable, difference?
Ken Stephens provided some very useful feedback. Thanks, Ken. I appreciate it. Being the bull-headed person I am, I of course feel free to use what I like and ignore the rest (for now).
I apologize for misusing the Verbal Behavior jargon. It has been nearly twenty years since I read Skinner's "Verbal Behavior", and I am certain that I did not really understand it then. In fact, my interest at the time was in social behavior, in general, and I took verbal behavior as a rather specialized form of social behavior (on the other hand, maybe all social behavior should be considered verbal behavior). Had some fun disagreements with Don Hake and Cloyd Hyten over what constituted social behavior, and how "direct" the social stimuli could be before the behavior could no longer be termed social. Of course, both ultimately came around to my point of view. Surely do wish Don had not died so young.
In particular, I do not know how to talk about the textual form of verbal behavior that we deal with. Even if the user is speaking, what we get is text. The only part of the software/hardware system that comes into contact with the vocal behavior is the speech recognition system on the front end (get speech recognition for your Mac from the online Apple store for $70; I think speech recognition is even available for Microsoft PC's; handheld wireless internet devices with speech recognition are on the market, or will be in a very short time). Regardless of the mode of verbal behavior of the user, we get the textual form.
Not sure what to call this textual format. Is not "transcriptive"; nothing is being copied or dictated. There is actually no way to determine whether the user is speaking, typing, using a pointing device, retinal or eye position reader, alpha wave reader, a program that is watching the user and issuing queries on the user's behalf, or what. All we know is that we get text that is supposedly in a human language, and directly or indirectly indicates something that the user wants us to do for the user. It is our task to "understand" the user input and behave appropriately. If we succeed, the user stays with us and we a paycheck. If we fail, the user goes to another service, and ultimately our business fails and we all start panhandling on the streets of Northern Virginia and DC.
As to whether there is a human listener, it depends. Humans are always in the loop more or less directly. The goal in using NLP software, of course, is to make it less direct (I'd like to be a LOT less directly involved). How direct does it have to be? There really is no such thing as "direct" contact with a human speaker or listener. Even face-to-face, there is indirection in that air molecules must be vibrated in such as way as to convey those vibrations without too much signal degradation to the listener inches away, and such that they impinge on the listener in such a way as to have the desired effects; oh, and there is a time lag (speed of sound is not infinite). How about telephone conversations? Is that verbal behavior? The already indirect sound vibrations are converted to analog electrical signals, which are then probably converted to digital signals, which are then translated via any number of thousands of devices, media, and miles, with measurable delays. Suppose there are longer delays, as with phone conversations relayed via higher orbit satellites? Half duplex radio conversations? Lagardly conversations via voice mail? Is are computerized "chat" and "instant messaging" conversations verbal behavior? How about the same conversations with a few lengthy delays inserted, such as email? Suppose your speech is converted to text which is then automatically translated to text in other languages which is then spoken to a listener days later in another country, and the listener turns around and responds in a similar manner? Is that verbal behavior? Suppose the translation program works by extracting "concepts", which it then uses to generate text containing the same "concepts" in another language, and the result is sent to the listener? Is that verbal behavior? How many times can the stimuli be transformed among media, shapes, colors, etc., and the behavior of the humans involved still be verbal behavior? How many delays and of what durations are permissible?
And on and on and on. Where do you draw the line? Why? Is it of practical significance? What is the function of the line? And is it a Gordian line that needs to be hacked up a bit?
Back to my Behavior Analysis page
Last modified on: