Teaching a monolingual student to respond to Spanish when that student is electronic and has «artificial intelligence» is a challenging and enlightening experience. The subject in the experiment is Intelligent Assistant, the «software genie» of Q&A, Symantec's file management software that, together with its integrated word processing program, is exceptionally easy to use. Q&A provides clear conventional menus for users to query databases, but what really distinguishes it from other products is its friendly Intelligent Assistant, the natural-language interface that manages the database.

Natural language refers to «any language that humans learn from their environment and use to communicate with each other», as opposed to artificial language, «created by humans to communicate with their technology» (Harris 3). The term «interface» refers to the connection or means through which a user communicates with the computer's software. Being able to write requests in «plain English», rather than in code or computer language, is a desideratum for many people who want immediate productivity from software without investing a lot of time and effort to learn special commands.

As a teacher of Spanish, I could not help but feel disappointed that such software was available only in English, and at the same time intrigued by the pedagogical challenge of trying to teach the monolingual Intelligent Assistant to understand «plain Spanish». Could Symantec's electronic «English Language Interpreter» learn to respond to a foreign language? To what extent would programmed English hinder reception in Spanish? Would the electronic student's learning processes be similar to those of college students? This article describes some of the considerations, procedures, results, and implications of an experiment in teaching Spanish to the Intelligent Assistant, as well as suggestions for practical applications for both teachers and language students.

Since the Intelligent Assistant works with databases, it was essential to provide a suitable learning environment. The assigned task was for the Intelligent Assistant to work with a database of 20th-century Spanish fiction designed to help graduate students prepare for comprehensive exams and find ideas for class discussions, papers, and dissertations. The title of the database, «Proyecto compartir», reflects the fact that the students themselves serve as principal collaborators by recommending books, contributing abstracts and bibliographical information, and suggesting keywords for major topics. Each entry in the Compartir database consists of three screens. The first contains the field names autor, obra, fecha, abstracto, colaborador (the student contributor), and categorías (keywords). The second is reserved for bibliografía, and the third, continuación, for any additional bibliography or annotations. This database would provide a learning environment for my students to learn literature, and, at the same time, for my electronic student to prepare for its role as a bilingual research assistant.

A good teacher should be aware of any limitations and idiosyncracies that might affect a student's learning, and this is equally true when dealing with the Intelligent Assistant. Q&A's instructional manual functions much like an academic record in that it helps the teacher form an idea of what the student already knows. The documentation also describes the electronic student's capacity for learning and for performing tasks. It explains, for example, that the Intelligent Assistant, whom Symantec calls IA for short, will create, fill in, find, sort, examine, edit, update, and delete forms, as well as prepare report, and make numerical calculations. IA will supply information about the current date and time, recognize interrogative words, accept follow-up

––––––––   285   ––––––––

questions, compare different parts of the database, and interrelate multiple records. It relies on a built-in vocabulary of 400 words, but automatically knows the words featured in the database. It accepts synonyms, likes to be told when it is learning adjectives and verbs, and responds best to clear, precise instructions. IA's body of knowledge is, of course, restricted to its field of expertise, which is the assigned database.

In order to teach Spanish to IA, the instructor also should know something about how the student learns, which in this case means having some understanding about natural-language processing -an emerging, complex field that includes translation, generation of text for specific use, and communication between humans and machines. Natural-language programs take into account many factors:

Approaches to natural language vary and may be pattern-matched, grammar-based, knowledge-based, semantic, or «connectionist» (Obermeier 225-32). In practice, programmers often integrate syntactic, semantic, and contextual knowledge strategies; they also strive to disambiguate language in order to achieve accurate interpretations.

A question pertinent to the pedagogical experiment is the extent to which the thought and learning processes of natural-language interfaces like IA resemble those of human learners. Cognitive psychologists, who study how humans think, approach language from a different perspective than linguists and logicians (Harris), and their theories have greatly influenced the field of natural-language processing. Some cognitive psychologists describe humans as general information processing systems that use receptors to obtain signals from the external world, a processor (which some consider multichannel, like recent computer systems) to manipulate the information received, a memory to store information, and effectors, to convey signals to the external world (Harris 7).

Another direction that exploits the analogies between human and artificial processing is neutral-network parsing, a recent approach that strives to imitate linguistic processing of information, based on neurological evidence in humans (Obermeier). While it is is impossible to determine the approaches that creators of a particular software package favor, we may assume the influence of at least some of these ideas in the development of IA. In any event, teaching Spanish to IA inevitably invites speculation on the similarities and differences between human students and their electronic counterpart.

The first step in teaching IA Spanish was to rename IA, who now responds to the name of «Ayudante Inteligente», which also appears on the main menu as soon as the program begins. The next step was to introduce IA to the database it would be managing in order to familiarize it with the field names (i. e., autor, obra) and the subject (obras). IA was then ready to begin working while continuing to learn.

The procedure for entering commands and questions is simply to place them in a box on the screen. Each word appears highlighted as IA examines it and proceeds through the query. When it encounters unfamiliar items, it reports that it doesn't know the highlighted word and offers options in order to teach it a new word, see or change its vocabulary, or go ahead («the word doesn't matter»). The user can instruct IA in formal «teach me» lessons or enter synonyms and alter vocabulary as needs develop.

Just as human students bring to the classroom a certain store of knowledge, IA has its own built-in vocabulary in English and knowledge about the database itself. IA learns all the field names on its own, not by semantic meaning, but rather by their function, ascertaining whether they refer to names, locations, values, or dates. It can only learn synonyms that refer to its built-in vocabulary, illustrating the well-known pedagogical practice of building on what a student already knows.

After learning less than 50 new expressions, IA could respond to a variety of queries or commands about the books in the «Proyecto compartir» database, such as:

Figs. 1 and 2 show the results of some simple requests and the responses of IA in its acquired language. IA could also handle more complex queries involving multiple fields, either/or searches (called Boolean logic), combinations, and formulas for exclusion of data, such as «Busca libros entre 1942 y 1985 con protagonista femenino y primera persona, pero no intertextualidad» (which produced the report shown in fig. 3), and «Enséñame libros antes de 1960 acerca de la guerra con adolescentes pero no con protagonista femenino» (shown in progress, with a request for verification in fig. 4).

The teacher can use ambiguous or subjective terms, and even expletives and slang, but must define them or instruct IA to ignore them. «Busca información acerca de Delibes» is too ambiguous for IA to process unless the instructor explains that información means «obras, años y abstractos».

Experience with IA shows that clusters of synonyms, such as quiero ver, busca, and enséñame for the built-in word «find», facilitate teaching. By asking IA to define, one can enter a single synonym in the query box or up to three at a time on a synonym screen.

IA has no trouble with phrases or idiomatic expressions, as long as it knows whether to identify them with a field, ignore them, assign a value, or relate them to previous vocabulary. For example, it can understand a command as idiomatic and colloquial as ponme al corriente de novelas del año de Matusalén in the following way: «Define "ponme al corriente" as "find"» and «Define "año de Matusalén" as <1800». The mathematical signs «<» and «>» enable IA to interpret antes de and después de when searching for comparative dates.

The teacher may adopt alternate strategies, such as designating the preposition de as a synonym of «by», as a word to be ignored, or as part of an expression. IA will accept more than one interpretation but will ask which one to choose for an assigned task. Ignoring or associating a word is faster than synonyms, which cause IA to request confirmation. IA ignored por favor with no trouble, but when asked to ignore articles, thought it necessary to include the corresponding nouns in reports. Combining words into expressions is more efficient than teaching individual items.

The organization of adjectives provides a fascinating lesson in how IA learns language. IA cannot transfer to Spanish its knowledge of morphological changes that make adjectives comparative and superlative in English, but this is unnecessary because it can learn that más raises value. Also, by assigning a low value to an adjective, like viejo and a high value to nuevo related to the number field año, IA can sort or compare novels by years. Fortunately -since Spanish syntax can differ markedly from English- IA is less sensitive to syntax and relies principally on semantics and context. So long as the words associated with the fields are present, IA «understands» different syntactical structures. Its «instinctive» morphological capacities include recognizing plurals formed by adding an «s», which it «mentally» eliminates.

One can watch IA's actual «thought» processes on the computer display, which provides graphic illustration of the stages that characterize natural language systems: the paraphrasing mode, in which input is converted into meaning representation (MREP); the inference mode, which draws all possible inferences about the input and a related part of the knowledge base; and the question-answering mode that converts and analyzes the input question and accesses the knowledge base for an answer (Harris). This is fascinating in itself, but in addition, enables the teacher to see how IA works with two languages. In the paraphrase mode, IA translates each synonym acquired in its learning mode back to English, producing temporarily on the screen a sort of Spanglish, together with a request for verification in its own terms, as in figure 4. IA presents the output of the inference mode in its native language, but, happily, when the report is ready, the input sentence appears once more, completely in Spanish.

In view of IA's inability to think entirely in the acquired language, the display of its thought processes is very revealing and highly suggestive: IA «mentally» translates new synonyms back to English, but not those it has learned by direct association or context. For example, when the teacher instructs IA to associate a verb, phrase, or query with the author field, it understands that when it encounters that expression, the author's name is the answer. The expression «primes» IA for a response by inducing a corresponding context. This is evidently the result of programming based on packets of entities, which

––––––––   287   ––––––––

are any objects (including persons) that can be described or about which statements can be made (Harris 318). On the other hand, IA can only relate synonyms to its immutable English lexicon. The fact that IA translates lexical items but not contextual ones makes it clear that for our electronic student, situational learning discourages translation.

Good pedagogical practice determines specifically what goals one wishes to accomplish and the best strategies to achieve them. In order to discourage IAs habit of translating into English, directing its learning process toward context rather than synonyms proved more effective. The teaching lessons for association that the manual suggested, however, were for adjectives and verbs, while I was using a variety of structures that were neither. IA learned that con meant «about», but did not know how to use it, i. e., that it should examine the keyword field. In such cases, the strategy of association to stimulate contextual learning, previously described, proved much more effective than synonyms. By teaching «Alternate Field Names», IA learned to identify the word cuándo with año and con with categorías, and that quién should elicit autor. Terms like se publicó, which are not innately associated with año, could nonetheless relate to that field heading in this particular database. Fig. 5 shows the six interchangeable components chosen for the field name año, which still left spaces for others. Redundancy creates no problem; IA accepts two alternate fields, like cuándo and se publicó, simultaneously.

My electronic student was in some ways intractable, for translation would continue to occur, although less frequently. By adjusting my teaching techniques by trial and error, however, I achieved what I considered the best performance possible, given the innate abilities of the student: IA would accept instructions in Spanish and carry them out quickly and dutifully.

Working with IA in Spanish can provide a valuable experience in classes for training teachers. The instructor can give students a common database with the original monolingual IA and ask them to teach it to manipulate the database. Students can demonstrate the prowess of their IAs, and explain the goals, methods and techniques employed. IA and the assigned database provide a controlled situation in which the electronic student's capacity is a given and results depend solely on the resourcefulness of the teacher, who cannot resort to gestures, tone of voice, or acting out words, as in the classroom.

Teaching or working with IA can also provide a profitable experience for students who are learning Spanish. They can observe in IA some of their own difficulties and practices. They should feel reassured by realizing that sometimes unknown words can be ignored without interfering with communication. Indeed, as Mark Larsen has noted, encountering unfamiliar words that force students to rely upon context can be excellent training for oral use of the language, since «we often engage in conversations in which some word escapes our comprehension» (943). Students can experiment to see what types of words IA can ignore without experiencing difficulty with comprehension.

If all goes well, students will learn content from the databases IA manages because of increased motivation. My example was a literature database, but another possibility is a database that contains information about important historical events in Spanish history, with their dates and a brief description.

By «programming» IA to learn Spanish, even experienced teachers can gain a better understanding about cognitive processes involved in setting up natural-language systems and the implications of such systems for teaching human students. For example, we can appreciate the arbitrary nature of language and the need for clarifying meaning (i. e. defining reciente as after 1940), formulating questions and ordering tasks in clear, precise language. The experience is also a lesson in respecting students' possible limitations. It is a challenge to try different strategies to see which one make it easier for IA to learn or reduce its requests for clarification. Teachers might also enjoy comparing IA with human learners and engaging in the polemics of cognitive scientists and psychologists of how closely machine thought approximates human thought or vice versa.

IA proved quite adept at learning Spanish. Devising actual teaching procedures was enjoyable and the results rewarding. Much of the credit for the successes belongs to the

––––––––   288   ––––––––

instructions and suggestions of Q&A's manual, which deserves an award for clarity.

As we have seen, experience in teaching IA shows that its reactions in many respects resemble those of beginning students: it tends to translate into already familiar language and, if it has doubts about the assigned task, will ask questions for further guidance and request verification in its native English. On the other hand, there are some notable differences: once it understands what to do, it responds immediately by producing a report on the screen that is always as correct as the information in the database. Another source of satisfaction is that it never forgets, unless you want it to, by deleting learned vocabulary from its lexicon. While it can learn to understand Spanish, it expresses options and requests for verification only in English, unless Symantec in the future should program these responses in Spanish.

As for retention, an important pedagogical goal, one might wonder whether the IA remembers what it has learned about a particular database or whether it is necessary to reprogram it for a new one. Fortunately, Symantec has provided a copy feature that makes it possible to reproduce the structure of the database, together with the bilingual IA. It is easy then to change the names of the fields and reintroduce IA to the database so it can learn the new names.

For language teachers and students, the Intelligent Assistant can be more than just an effective tool for database management; it can be a laboratory to test their skills and to explore the nature of language learning in an artificial intelligence environment.

Harris, Mary Dee. Introduction to Natural Language Processing. Reston, VA: Reston (Prentice-Hall), 1985.

Larsen, Mark D. «Obstacles of Integrating Computer Assisted Instruction with Oral Proficiency Goals». Hispania 70 (Dec. 1987): 936-44.

Obermeier, Klaus K. «Natural-Language Processing». BYTE Dec. 1987: 225-32.

Pollack, Jordan and David L. Waltz. «Interpretation of Natural Language». BYTE Feb. 1986: 189-98.

Q&A. Computer software. Cupertino, CA: Symantec, 1988: Version 3.0.

Hispania [Publicaciones periódicas]. Volume 73, Number 1, March 1990