Drew Clary
Minds and Computers
12/9/00
Artificial Intelligence, while to some nothing more than the science of toying around with robots, fits anything but this description. AI is an interdisciplinary field involving philosophy, psychology, computer science, and engineering, among other things. Within the field of AI lay even more subfields. Two such subfields are those of language interface and learning in computers, or in many aspects, the collaboration of the two. To many, these areas and their combination are essential to the future of Artificial Intelligence. Take for instance the majority of the significant AI projects to date. These include Eliza, Parry, PDP, Hacker, GPS, Shrdlu, and many others. All of these systems, in some form or another, in varied rates of success, replicate or simulate human methods of communication and learning. Whether the programmers were attempting to simulate human processes or just trying to make a functional system that achieves similar outputs with humans is not entirely clear. In cases like Eliza, Weizenbaum was not necessarily trying to create a machine that actually used the same internal methods of language and learning as those of a human, but rather held the goal of making a system that could interact linguistically with humans on some level and give some kind of impression of language and learning. Regardless of their programmers’ intentions, however, many Natural Language Processing systems actually employ many of the methods for such processes as those innately used by humans in language and learning. While these programs were merely the first steps, they were very significant ones in that they opened
the doors for future programmers, whose goals now are to create systems that do more than just give the impression of understood learning and interaction with humans, but rather really learn and use language in the sense that humans do, and do so via the methods used by humans, of course translated in some respect into computer “language.” It is in these ideas among others that the future of Artificial Intelligence lays in.
How do people learn? Some will study a subject or topic for hours, “pounding” it into their brains. Others seem to hear it once and know it. Regardless of the apparent styles of learning, there are certain methods, whether they are acknowledged or not, that allow people to learn. “Learning strategies are methods people use to learn and retain information. They are usually effortful and under the learner’s cognitive or conscious control” (Dunn 328). There are several different methods of learning, also known as learning strategies.
Very often when students are studying for a quiz or a test, they will attempt to create mnemonic devices to help them relate words to definitions. Mnemonics are strategic devices that aid the learning and recall of otherwise difficult material. Very often these devices are based on verbal or textual mediation strategies in which two words or ideas are associated by another word or phrase, which acts as a type of verbal mediator, and links them together, at least in the mind of the person employing these devices. One example might be the device for remembering the order of the planets of the solar system: My Very Educated Mother Just Served Us Nine Pies, stands for Mercury, Venus, Earth, Mars, Jupiter, Saturn, Uranus, Neptune, and Pluto. However,
mnemonics do not have to be even close to being this apparent. As long as there is meaning behind the association between the two subjects, the mnemonic device is effective. All that is necessary is a type of internal, understood logic, and the system, the mnemonic-human system that is, can operate. Another mnemonic strategy, the method of loci, invokes visual imagery to some degree. According to Dunn, it consists of (1) identification of familiar places sequentially arranged, (2) generation of images of the item to be remembered that are then associated with the places and (3) recall of the items by revisiting the places (328-329). While in a way somewhat casual and informal, the loci method as well as the previous mnemonic methods are very mechanical and almost computational. They are devices that allow the system to recall information quickly by using types of short cuts and stored data; in many AI enthusiasts’ eyes, they are also devices that just might be replicable.
Repetition, or rehearsal, is another
common method of learning. Often a
strategy for pure memorization and encoding, rehearsal is usually used to
remember something like telephone numbers, definitions, or acting lines. The more a person rehearses a certain topic
or item in short-term memory (STM), the more likely it is to be transformed
into the permanent long-term memory (LTM).
If a person does nothing more than mentally hold the information in STM,
it is just ineffective maintenance rehearsal (Dunn 329). Another method also sounding quite
mechanical, rehearsal seems if anything the most able to be replicated in a
machine. It could be argued that this
type of learning is without whole understanding – perhaps in a human but even
more in a computer. However, it seems
that in humans and hopefully one day in computers, repetition invokes
understanding, and in some cases, even feeling or emotion. Very often once a certain topic or subject
is memorized (and only then) it is understood.
Once it can be recalled and analyzed, it becomes more and more part of
the system and its memory, or sometimes even part of its understood
memory. A not so scientific example
lies in the field of theatre. It is of
course necessary for an actor to memorize his lines. However, it is also necessary for him to “feel” his lines and
character in order to be convincing to the audience. It is consequentially not possible for him to feel and understand
his lines and character before he even knows his lines. These examples give rise to the additional
idea that often repetition is necessary for a higher understanding. In so many cases, while perhaps knowledge
and understanding is gained from mere rehearsal, this rehearsal is but the
vehicle for an understanding of something much bigger and more complex. All of these examples and defenses of
repetition are incorporated into machine learning, but often in less abstract
forms. It is clear, as seen in PDP, a
“learning” neural network, that repetition is necessary in computation – not
just in learning and understanding, but in function as well.
In a way building on the concept of rehearsal are organizational strategies, which are effective methods used to improve LTM, thus aiding the learning process. There is a tendency in humans to group materials into categories or chunks for later recall (Dunn 329). This is of course a very practical tendency, as case studies have shown, and computational machines take full advantage of the concept. Computers organize information into categories, not just as pure method, but for speed and space, and in many cases with their own kind of tendency.
Semantic elaboration, while a type of complex rehearsal, is nonetheless quite a unique learning method. In a way it displays a unique combination of bottom-up and top-down processing. It creates meaning and understanding by using new information and a person’s existing knowledge to build a unique interpretation of the new input. It is theorized by people like Rumelhart and Orleny that information is organized based on general knowledge structures of understanding and learning called schemata (qtd. in Dunn 329). Every time a person has a new experience, it becomes mapped in their memory, and further repetition of this particular subject or experience only serves to add emphasis to it. Any alteration to this pattern or scheme of experience can be seen as further leaning (Dunn 329). A kind of externalization of this method is concept mapping, which in essence shows the almost mnemonic relationships between new and old ideas and topics in an attempt to aid in identification and understanding (Dunn 329). This externalization of course hints at techniques, formats, and methods of learning and understanding eminent in computers. The mere idea of concept mapping excellently relates the learning methods of humans to those of computers and, if nothing else, initiates the idea of the possibility of learning machines.
A common thread that seems to run through these various methods of learning is the use and sometimes necessity of language. Mnemonics, for instance, almost wholeheartedly relay on language and language interface. Without words, phrases and their implications, connections using them as well as those between them would of course be impossible. Other methods of learning, while not as convincingly language-oriented, still are greatly aided by the incorporation of language. In certain situations involving these methods, word incorporation is the only plausible approach.
Machine learning, which in many ways is the next step for NLP systems, is also an area of much activity and research. According to Anzai, machine learning is “…when a system automatically generates a new data structure or program out of an old one and thus irreversibly changes itself with some purpose for a certain amount of time” (5). In a sense, could this description not be used for human learning as well? When people learn something new, not only is information added to their memory bank, but existing information often changes as well. As with humans, there are numerous methods by which computers learn, many of these methods being similar to that of humans. Quite similar to the human method of semantic elaboration is that of machine learning based on logical inference. As Anzai writes, a system for organizing knowledge by logical inference serves to generate new knowledge that had not been seen before. In addition, a system based on logic can be used for algorithms that actually generate new information by embedding other learning methods, such as generalization and specialization, into logical inference (Anzai 235). It should also be noted the aspects of logical inference evident in the manipulation and creation aspects of language. In a sense, the manipulation and organization of language structure gives rise to new forms and aspects of language, and may even explain the human’s ability to hear and understand words or a combination of words he or she might have never before heard. Another relation to methods of language lays in mnemonics and their reliance on logic. While this logic might be internal and logical to no one but the user, it is logic nevertheless. Another method of machine learning, learning by neural networks, shares many of the characteristics of rehearsal in human learning. Probably the most well known neural network system is PDP, or Parallel Distributed Processing. Key to the logic behind PDP is repetition. Essentially, PDP is a network of artificial neurons set in a certain structure, attempting to simulate the functionality of the brain. Through a kind of smoothing-out type repetition, a group of these neurons can actually learn and hold a kind of memory. Amazing as this feat is, it would not be possible without the simple concept of repetition.
“Communication is essentially a process by which the state of affairs at one place is transmitted to another place by symbolic means” (Back 266). While this general term gives rise to a fairly general definition, it is still effective in describing the ways by which people interact with one another and share information, among other things. Just the word in itself invokes a feeling of its necessity and importance in the world today. The study of communication, while fairly abstract, in many ways engages the collaboration of many different disciplines. It is the study of communication and an application of a unified communications model that would be beneficial, not only for communication itself, but for the various disciplines dependent on it. In this sense, a connection can be made to Artificial Intelligence and Language Interface Systems, for in these systems, uniformity is not only helpful, it is essential. Without a common approach, without a common language, these systems would be impossible. There is an accepted general form of analysis, proposed by Shannon and Weaver in 1949, of the process that consists of five steps: (1) the source (the original state of affairs); (2) the transmitter; (3) the channel; (4) the source of noise, or the origin of errors in transmission; and (5) the receiver, leading to the reconstructed state of affairs. The signal received is a function of the original signal and noise. The channel is only able to vary a certain kind and amount of data. Therefore the translation of the data into the form acceptable to the channel (encoding), and the reverse process used by the receiver (decoding), become key in the analysis and design of the communications system (qtd. in Back 266-267). While very general in its description, it allows some conclusions and relations to be drawn. “The scheme proposed by Shannon and Weaver lends itself nicely to mathematical analysis, giving quantitative measures of information, channel capacity, error reduction, redundancy, and efficiency of coding systems” (Back 267). Not only does the proposed process of communication lend itself to mathematics but it is nearly dependent on it. Its roots are computation and logic, for it is these concepts that allow the relation to mathematics to exist. The possibility of communication being able to be broken down into a type of mathematical computation opens many doors, not just for Natural Language Generation but for Natural Language Understanding, Language Interface, and Artificial Intelligence as a whole.
Interpersonal communication is a difficult phrase to define. Most see it as a method of giving and receiving information via signals or messages by talk, gestures, writing, as well as many other methods. As Tedeschi writes, this idea of transmission of information can be applied to genetic materials and other nonorgainc interactions. An individual might transfer information from one cognitive context to another in a form of intraindividual communication. In addition, categories representing intergroup or even international communication could be developed. Interpersonal communication refers to transfer of information by a source to a specific other target or identifiable group of members (Tedeschi 279). While usually these communications occur face-to-face, they may occur by mail, telephone, or internet. This interpersonal communication is in a way what language interface systems (NLP) are attempting to achieve. These systems’ programmers wish for the computer to receive an input – some form of information – and return a logical output. These inputs and outputs may of course be in the form of symbols, which must be decoded and then encoded. Shannon and Weaver apply electrical engineering principles to human communications. The mind of the communicator can be considered to be the source of the communications. Messages augment in the brain and are encoded for transmission to other people. The encoded message is sent as a signal to a receiver, who must decode the message (qtd. in Tedeschi 279). This system is essentially what occurs between the computer and a human, and in many cases even within the computer itself. The accuracy of the system’s interpretation of the information is also a key issue. Any interference with accurate transfer of information, or worse, may be due to ambiguous encodings, problems with channels, or faulty decoding by the target (Tedeschi 279). While this very technical description appears to be one of a computer language interface system, it is in actuality one of human communication, once again conveying the strong relation between the two.
Language, of course, is a key form of communication. Many would go as far as to argue that the two are one in the same. “Language is a means of information processing that is used to store, manipulate, create, and transmit information”(Tedeschi 280). Eminent in communication are signs, signals, and symbols. An example of a sign situation might be a hunter noticing deer tracks and interpreting them as a sign that a deer had passed recently. A signal is something that most animals can use to communicate to others – a bird chirping, a skunk spraying an intruder, or monkeys showing threatening gestures. Symbols, however, seem to be limited in their use to humans. Symbols derive their meaning from a community of users, and the use of symbols allows the development and expanding of various abstract ideas and interpretations. Furthermore, they provide the basis for the individual’s constriction of social reality, including a self-concept (Tedeschi 280). Again this description of human communication traits lends itself greatly to similarity with language interface programs. A computer’s “symbols” can be interpreted as its programming language, something that is universal and consistent throughout and that the system’s operation depends on. The use of these symbols and their occasional ambiguity, which can sometimes be detrimental, give rise to new and compounded interpretations. These new interpretations, however, might often be displayed as errors in a computer system.
As Lachman writes, during the zenith of behaviorist domination, there was a type of debate between linguists and behaviorists on the phenomenon of language that ultimately helps to define it, or at least give it a fuller description and relate it better to language interface systems and AI. On one point linguists generally felt that novelty is a compelling and central characteristic of language. Its universality allows it users to understand and produce a potentially infinite number of utterances that they had never heard before (Lachman 246). To many computer programmers, this situation is the ultimate goal: for the system to give a reply of its own creation, and not one that is simply chosen from a list of answers. Behaviorists, on the other hand, focused on the repetitive elements of language and their relation to conditioning. Linguists viewed complexity as a central characteristic of language, while behaviorists looked for simple principles of behavior. Linguists most importantly felt that language had structure, rules, and grammar, amongst its infinite capabilities. Similarly, Noam Chomsky argued the idea of this kind of structure in the brain, which in turn gave rise to language. These qualities are of course eminent in both the main processor as well as the output (language) of natural language systems. In addition, Chomsky argued that people behave as if they have learned rule systems (qtd. in Lachman 246). Another argument in a way involves the concept of innate, top-down processing. Linguists and psycholinguists pointed out that children all over the world learn different languages, begin the acquisition process at about the same time, take about the same length of time to master the essentials of their language, and go through similar steps in the process. It seems that some aspects of language must be innate. Natural Language Processors have a type of innate, top-down generation of language and it is the goal of the programmers for them to eventually hold more advanced and more human-like bottom-up processing capabilities as well. These systems, or at least these potential systems, seem to share many input processes, and output-oriented characteristics with the human brain and language systems. The theory is there, and the technology is soon to come.
Natural Language Processing (NLP) is arguably one of the most important and prospective subfields of Artificial Intelligence. In a sense what those involved with NLP wish to achieve is a system that can interact with humans on a linguistic level close to if not equal to that of humans. Most of the systems to date, such as Parry or Eliza, employ somewhat primitive methods. Basically, when a textual input is received, their processors decode it in a sense, run through a list logical response based on the input, encode an output, and display it. It is the goal of the programmers to expand on the memory (list of responses), perhaps to an infinite level, and in turn create a kind of understanding and originality of response within the machine.
By the 1970’s, NLP was still in a fairly primitive state. Programs for understanding English were typically developed by one person, who not only defined the grammar of English to be used in the system, but also wrote the parser to use this grammar during the analysis of English input; parser being a type of decoding. Very often individuals invented their own formalisms for representing syntactic and semantic information (Allport 91). In general, there was no consistency of grammar, understanding, or methods among programs, and sometimes even within individual programs. By the end of the 1980’s, NLP had at least begun to show some progress. “A significant consensus among computational linguists [had been developed] conceiving the nature of the grammatical formalisms that are most appropriate for computer natural language processing” (Allport 91). In addition, processing techniques providing efficient and principled methods for applying grammatical information during the process of language input became almost commonplace in programs of the respective caliber. Developers of new NLP systems were and are now able to draw upon a large body of knowledge and research, in turn providing the Natural Language Processing subfield of AI with a type of common framework and continuity by which to work by, and, ideally, allow large teams of researchers to contribute to the development of a single NLP system. Most present NLP systems use formalisms derived from a combination of theories from the fields of linguistics, computer science, and formal logic with the insights gained during the experiments of NLP’s earlier years (Allport 92). This integration is possible due to the constantly increasing complexity of the formalisms themselves. “Far from being a hindrance to further development, the more sophisticated formalisms for natural language processes are bringing to fruition in computational linguistics that multidisciplinary synthesis which is the essence of artificial intelligence” (Allport 92).
As noted by Allport, numerous attempts, some of them unsuccessful, have been made in furthering the field of Natural Language Processing. From these attempts that occurred during the 1970’s, it seems that the AI community realized the difficulties in using procedural formalisms for describing language, and thus began to move towards the use of declarative formalisms. Research in linguistics during the 1970’s was dominated by Chomsky’s Transformational Grammar paradigm. This approach to grammar, however, failed to provide a formalism for describing natural language, in the sense of a formal language with clearly defined syntax and semantics that could be used as input to a computer program (Allport 106-107). The beginning of the 1980’s marked a fairly significant step forward for NLP. It was around this time that ideas from the field of linguistics began to be introduced directly into the formalism used by NLP systems. Results of linguistic research would probably have been used much earlier, but there were problems inherent in the formalisms used by linguists in the 1970’s that prevented their direct use in computer programs (Allport 106).
“Any natural language processing system can be conceptually divided into three parts: grammar, dictionary (lexicon), and the programming system that holds everything together” (Heidorn 2). According to Heidorn, grammars are systems of rules that in a way connect symbols and meanings. They have a dynamic nature, and are supposed to embody generalizations that hold true for many symbols and combinations of symbols (Heidorn 2). In essence grammars aid to the general consistency and universality of the language system. Lexicons are repositories for particular units like words or phrases, and for information about these units. Lexical information is usually static and specific in nature. The comparison between grammar and dictionary provide fairly ambiguous bases for designing a real system, and leads to a question of the proper distribution between grammar rules and lexicon. The actual situation is a type of continuum of which two poles support – one of emphasis on lexical entries and simplification of rules, and the other of reliance more on grammar and rules (Heidorn 2-3). Traditional components of linguistic theory include phonetics, phonology, morphology, syntax, semantics, discourse, and pragmatics. To date, most has been in syntax and semantics, but research and progress is soon to come in the others.
PLNLP was born in the 1980’s at the IBM Thomas J. Watson Research Center, and involved many people around the globe. PLNLP, or Programming Language for Natural Language Processing, “. . . is an integrated, incremental system for broad-coverage syntactic and semantic analogies and synthesis of natural language” (Heidorn 2). PLNLP is essentially a culmination of ideas and themes into a general, almost universal Natural Language Processor. Since the PLNLP system restricts its input to typed text, it deals only with morphology, syntax, semantics, discourse, and pragmatics, all but a few of the traditional components of linguistic theory. Of these components, PLNLP has concentrated on syntax and semantics. According to Heidorn, there are six essential components to this English analysis system:
1. Syntax, consisting of the broad-coverage English sentence analysis grammar PEG (the PLNLP English Grammar), coupled with a large lexicon that is basically a list of English word stems with fairly simple associated feature information. The lexicon started with entries from the full online Webster’s Seventh New Collegiate Dictionary. Although the number of words covered is great, the amount of information per word is small, compared to what is described for many other syntactic grammars. Linguistic information is distributed much more heavily over the rules that over the lexicon in this component.
2. Corrected syntax (reassignment), which takes the output from PEG and resolves many ambiguous syntactic analyses, based on semantic information from online dictionary definitions. It recursively calls PEG to retrieve and analyze dictionary information, applying heuristic rules to that information in order to “bootstrap” its way from syntax into semantics. During this process, some word sense disambiguation falls out automatically as a result of the attachment disambiguation. Since the lexicon associated with this component actually contains entire online dictionaries, the amount of information per word is huge, much larger than what is described for other NLP systems; and the distribution of linguistic information in this component is heavily skewed toward the lexicon.
3. Derivation of logical form (PEGASUS), which takes the corrected sentence parse and produces a graph that is the basis for further semantic processing. In so doing, it determines: (a) the structure of arguments and adjuncts; (b) pronoun reference; and (c) verb phrase anaphora (the semantic structure of elided VPs). These steps are accomplished by a set of procedures that operate strictly on the output of the reassignment component, without consulting any additional lexical information.
4. Sense disambiguation, which narrows down the possible senses of verbs and nouns in the sentence. It operates on the output from the previous component, mapping target words, in their sentential context, to relevant online dictionary entries. Taking advantage of all available information – from the parsed analysis, from the dictionary, and from other sources – the most likely possible senses of words are identified through a strategy that weights various types of evidence and ranks senses according to a similarity measure. The balance of information in this component is again weighted toward the lexicon, because of the significant use that is made of online dictionary resources.
5. Normalization of semantic relations. The first step in constructing a discourse model is to refine the semantic graph, with the goal of creating a common or normalized representation for all inputs that mean the same thing. The notion “mean the same thing” is still fairly intuitive, and this component has been only partially implemented. Normalization routines are intended to inspect nodes in the graph and the relations between those nodes and identify rule-governed paraphrases across a wide variety of syntactic domains. Of course, the process of normalization has already been started by PEGASUS; for example, equivalent argument structures are produced for active and passive variants of an English sentence. Although the routines are semantically oriented, they do not lose access to the surface syntactic differences.
6. Paragraph (discourse) model. After all possible sentential normalizations have been made, the system must join sentence graphs to build a formal model of those discourse chunks which, in written text, are typically called paragraphs. Much remains tentative here, because this component has also been only partially implemented. However, at this point it seems likely that the distribution of activity between rules and lexicon will be fairly even for this component and for the preceding one.
(Heidorn 3-4)
Almost anyone who is involved in NLP knows of Joseph Weizenbaum’s Eliza. Eliza is a computer program psychotherapist. “Patients” can talk about their problems by typing to Eliza, which in turn replies with a soothing yet disquieting, ambiguous response. If given a fairly logical, correctly typed input, Eliza can often respond with an equally logical yet vague output. In a conversation between Eliza and a teenager (see Appendix), Eliza is seen at its best. The conversation even shows Eliza’s ability to relate responses to topics from earlier in the conversation, giving the impression of a kind of understanding. Of course this is a nearly perfect demonstration of Eliza’s capabilities. In most cases the program will slip up by giving an unintelligible, illogical response. Regardless of its failings, people were responding to Weizenbaum’s machine as if it were a human being. As Copeland notes, even Weizenbaum’s own secretary took to insisting that other staff leave the room so she and Eliza could talk in private. Psychoanalysts were prepared to let his machine loose on their patients, and even the Journal of Nervous and Mental Disease supported his machine as having therapeutic capabilities (13-15). If nothing else this display of human-machine interaction indicates that the concepts and ideas behind NLP systems are well founded. It should be noted that Eliza is just one of many NLP systems and that it is far from achieving any type of real thought or understanding, all with certain similarities and differences. While many advancements are needed, the idea of language interface and its progression towards the achievement of interacting, thinking machines is one of promise and to many, excitement.
If anything, the fields of language interface and learning as well as their collaboration hold much promise for the future of AI. This promise is conveyed in numerous ways, most of them relating in some way or another to the strong similarities between human methods and computer methods of language and learning. In many language interface and learning systems, the programmers did not necessarily even attempt to build a machine holding similarities of method to that of a human, but rather programmed more with the goal of functionality and some degree of human interaction capability. However, in most cases, these functional systems happened to hold great similarity to humans – not just in response, but in methods of function and to some degree even in a kind of understanding. Now of course these learning and language interface systems are nowhere near the level of human-to-human interaction, as many critics will point out. However, in relation to those of humans, the methods are well-founded, and seem to have the potential for being the foundations for great progress in Artificial Intelligence; this progress ultimately being to create machines that could do more than just give the impression of understood learning and interaction with humans, but rather really learn and use language in the sense that humans do. Even if this goal were not reached entirely, if Artificial Intelligence systems, or more specifically learning and language interface systems, were convincing enough to the point that they might pass the Turing Test, who is to say that theses accomplishments would not give rise to a higher level of understanding and “thinking” achieved by the computer? It does seem that these supposed systems that might pass the Turing Test would not reach that point via pure brute force programming, and that a higher level of language interface and learning might be needed. In a sense the two concepts build off one another: while the higher level of machine learning and language interface might give rise to machine understanding and perhaps “thinking,” this same understanding might be necessary for this level of programming. It is the goal of those involved in machine learning and Natural Language Processing systems to integrate the two in an attempt to achieve the goal shared by them. To many the idea of one accomplishment just giving rise to another is in a way unfathomable, for this type of paralleling of achievements seems impossible, at least with the mentality of many people today. However, just because something is unfathomable or seems impossible at the time hardly means that it is undoable. While quite elementary, an example of this concept lays in the years before space exploration. Before space exploration became a possibility, many probably felt that not only was it impossible, but ridiculous as well.
Allport, David. “The Changing Relationship Between AI Programming Languages and Natural Language Processing Formalisms.” Machinations: Computational Studies of Logic, Language, and Cognition. Eds. Richard Spencer-Smith and Steve Torrance. Ablex Publishing Corporation: Norwood, New Jersey, 1992. 91-123.
Anzai, Yuichiro. Pattern Recognition and Machine Learning. Academic Press, Inc.: San Diego, CA, 1989.
Back, K. W. “Communication Processes.” Encyclopedia of Psychology. Ed. Raymond J. Corsini. New York: A Wiley-Interscience Publication, 1994. 266-268.
Copeland, Jack. Artificial Intelligence: A Philosophical Introduction. Blackwell Publishers: Malden, MA, 1993.
Dunn, B. R. “Learning Strategies.” Encyclopedia of Psychology. Ed. Raymond J. Corsini. New York: A Wiley-Interscience Publication, 1994. 328-329.
Heidorn, George E. et al. Natural Language Processing: The PLNLP Approach. Kluwer Academic Publishers: Boston, 1993.
Lachman, R. “Information-Processing Theory.” Encyclopedia of Psychology. Ed. Raymond J. Corsini. New York: A Wiley-Interscience Publication, 1994. 245 -248.
Tedeschi, J. T. “Interpersonal Communication.” Encyclopedia of Psychology. Ed. Raymond J. Corsini. New York: A Wiley-Interscience Publication, 1994. 279- 281.
*Eliza’s responses are in all CAPITAL LETTERS
Men are all alike.
IN WHAT WAY
They’re always bugging us about something or other.
CAN YOU THINK OF A SPECIFIC EXAMPLE
Well my boyfriend made me come here.
YOUR BOYFRIEND MADE YOU COME HERE
He says I’m depressed much of the time.
I AM SORRY TO HEAR THAT YOU ARE DEPRESSED
It’s true that I am unhappy.
DO YOU THINK COMING HERE WILL HELP YOU NOT TO BE UNHAPPY
I need some help, that much seems certain.
WHAT WOULD IT MEAN TO YOU IF YOU GOT SOME HELP
Perhaps I could learn to get along with my mother.
TELL ME MORE ABOUT YOUR FAMILY
My mother takes care of me.
WHO ELSE INYOU FAMILY TAKES CARE OF YOU
My father.
YOUR FATHER
You are like my father in some ways.
WHAT RESEMLANCE DO YOU SEE
You are not very aggressive but I think you don’t want me to notice that.
WHAT MAKES YOU THINK I AM NOT VERY AGGRESSIVE
You don’t argue with me.
WHY DO YOU THINK I DON’T ARGUE WITH YOU
You are afraid of me.
DOES IT PLEASE YOU TO BELIEVE I AM AFRAID OF YOU
My father is afraid of everybody.
WHAT ELSE COMES TO MIND WHEN YOU THINK OF YOUR FATHER
Bullies.
DOES THAT HAVE ANYTHING TO DO WITH THE FACT THAT YOUR BOYFRIEND MADE YOU COME HERE