NATURAL LANGUAGE PROCESSING BY COMPUTER AND LANGUAGE TEACHING	Vivian Cook Spelling data Writing Home SLA Home
V. J. COOK and D. FASS System 1986

Many commonly used computer teaching techniques fall into three limited categories. One category is word-guessing games in which the computer is used partly to provide a topic for discussion by the students and partly to teach some aspects of the patternings of texts. A second category is traditional techniques such as structure drills, grammatical explanation, and grammatical correction. A third is discussion stimulation, i.e. activities where the computer serves as a pretext for conversation or as an adjudicator of group decisions but is not directly teaching language. Word-guessing and traditional techniques require teachers to assume that learners benefit from conscious understanding and manipulation of linguistic rules or from mechanical practice rather than from trying to express their feelings and ideas through language, as is commonly held today. While discussion stimulation indeed forms a valuable part of teaching, this use hardly begins to tap the potential of the computer; in terms of cost effectiveness, magazine photos, train maps, or Dungeons and Dragons dice are arguably better value. It seems then that we should explore ways in which the computer can contribute more directly to modern teaching methods.

This paper argues that this can best be done by exploiting the computer’s unique ability to handle natural language. In the area of research known as Natural Language Processing (NLP), computer programs have been developed that describe the structure of sentences, that answer questions about selected subjects, and that engage in extended dialogue with humans. This capacity of computers to process human language has, however, had little influence on the use of computing in language teaching. The present article outlines some applications of existing NLP work to language teaching, looking first at syntactic parsing and then at more semantically-based processing.

Syntactic parsing is concerned with assigning grammatical structure to sentences by computer. Syntactic parsers produce representations of sentences as syntactic trees containing standard grammatical categories such as S, NP, VP, Det, N, and so on. Four basic parsing strategies can be adopted (De Roeck, 1983): (1) top-down and depth-first, (2) top-down and breadth-first, (3) bottom-up and depth-first, (4) bottom-up and breadth-first.

A top-down parser assigns structure to a sentence by starting with the symbol S and searching for rules to expand it, such as S-NP VP. It then expands NP and VP in turn, e.g. NP-+Det N; when it can expand them no further it checks to see if the terminal symbols match the input. A bottom-up parser, on the other hand, starts to work on the input sentence itself; it replaces the words by their syntactic categories, and then replaces strings of categories by other categories, e.g. Det N is replaced by NP, until it builds up the representation of the whole sentence.

Usually there are many possible routes through the parser for any sentence, although most prove to be dead ends. A depth-first parser follows one path at a time; if the path it is following ends, the parser backtracks to the last junction it met and follows an alternative path. A breadth-first parser explores all the paths at the same time; one of these is bound to be successful, so it never needs to backtrack. Perhaps the best-known syntactic parser, the Augmented Transition Network (ATN) (Woods, 1970; Bates, 1978), works top-down and depth-first. An example of the bottom-up, depth-first strategy can be found in chart parsers (Kaplan, 1970; Kay, 1976).

The choice of parsing strategy is independent of any particular grammatical theory or form of structural description; parsers have embodied syntactic theories as diverse as systemic grammar (Winograd, 1972), transformational grammar (Petrick, 1973), and Generalised Phrase Structure Grammar (Thompson, 1981). Few existing parsers, however, can deal with a large proportion of the syntax of native speakers and those that can, such as Woods (1973) or Sager (1981), have complicated grammars that are cumbersome in operation.

However, such limitations may not be important for second language (L2) teaching; second language learners use limited grammars, restricted by the natural sequence of acquisition in a second language, by the linguistic grading of language teaching courses, and by channel capacity constraints on L2 processing. The difficulty of writing a parser for L2 teaching may be considerably less than writing one to model a native speaker; Loritz’s ATN parser for language teaching shows this can be done neatly on a microcomputer (Loritz, 1984).

In a structure drill the students produce sentences that have the same syntactic structure with lexical variations produced in response to clues supplied in cue sentences (Cook, 1982). After they have answered they are told whether they are right or wrong either directly by the teacher or indirectly by comparing their actual answer with the right one. One possibility is to use a parser to vary the cue sentences; each time the drill is used the parser produces new sentences for the students to respond to. Markosian and Ager (1983) describe a system in which the teacher feeds in the format for the drill and the program itself generates the actual drill material using a parser and a lexicon. Such a use makes the drill less repetitive to the students and less of a chore for the teacher to devise. A second possibility is to involve the parser with the student’s responses. In conventional drills the teacher or the students themselves have to evaluate whether their responses differ from the model. Existing computer drills provide some correction of student syntax within a limited number of preset responses (Marty, 1982). A parser goes further by enabling the computer to detect any error within its grammar, and then using this information as a basis for feedback to the students or scoring for the teacher. If drills are seen as information processing exercises rather than as mechanical and habit-formation (Cook, 1982), a parser can extend the flexibility of computer structure drills.

The technique of grammatical explanation starts with presentation of grammatical rules, followed by practice in which the student applies them, often by filling in the blanks in sentences. A modern computer version is seen in TICCIT (Merril, 1980). A parser itself does not offer much to the rule-presentation phase except through visual displays of grammatical structure or the provision of freshly-coined examples. As with drills, the parser contributes to the practice phase by providing corrections, continuations or rephrasings of the student’s sentence. For example Pulman’s GPSG program parses the student’s sentence, marks where it is deviant, and gives a continuation from the point of deviance (Pulman, 1984).

The students type a sentence such as “William put the book on the she/A” they then ask questions, such as “Was the book put on the shelf?, ” which receives the answer “Yes. ” If they ask “What did William put?“, they are told Sentence OK up to here “What did William put . . . ” Expecting to find one of the following; preposition (in, on, etc). Examples of grammatical continuations: What did William put . . . with something. This program presents a novel and useful form of grammatical correction in which the students actively explore their knowledge of English.

A popular current activity with computers is guessing words in texts. Originally derived from the Cloze test, these programs display the text as a series of blanked out words consisting of asterisks or tildes, for example STORYBOARD (Jones and Higgins, 1984), and TRAY (Moy, 1984).

Students guess words in the text and the computer supplies them wherever they occur; various scoring schemes emphasize the games element. One program in this vein, SHANNON’S GAME (Cook, 1985a), asks the students to guess words in a text from left to right; they can ask for grammatical clues about the next word. The grammatical information is extracted from any text the teacher types in by means of a simple bottomup parser. Thus a parser brings a stronger language element into text-guessing exercises. Rather than suggesting radical alternatives, syntactic parsers tend to extend existing teaching techniques; they deepen the level of language that can be handled by the computer within conventional techniques such as drills and grammatical correction, and within more recent techniques such as word-guessing.

Many Natural Language Understanding (NLU) programs provide a semantic analysis of sentences, i.e. they capture their meaning in semantic representations. The exact form the representation takes depends on the task the NLU is to perform. Two such tasks seem particularly relevant to language teaching: query-answering and dialogue.

Query-answering (QA) systems enable users to retrieve information from, and add information to a database, by asking natural language questions; one component is usually a syntactic parser that acts as a means of access to the stored information. A well-known QA system is LADDER (Hendrix et al., 1978), which answers questions about ships such as Give me the length of the Kennedy. LADDER’s parser is more sophisticated than those outlined above in that it handles ellipsis, (Width and height? meaning What is the width and height of the Kennedy?) and pronominal reference (Who is her commander?).

Some use has already been made of QA systems in the teaching of other subjects. For instance THEATRE (Sheed, 1984) consists of a booking system for theatre tickets; pupils can find out what is on at different theatres, what tickets are left, and can make a reservation. In language teaching, the direct uses of QA systems are perhaps confined to stimulating discussion and roleplay. Databases can support classroom activities; in the context of French teaching, TELEM-NANTES (Starkey, 1984) utilizes a database about the town of Nantes based on the experimental system available to the citizens of the real town. Such an aid gives a useful reality to cultural background work and a spur to functional roleplays.

(a) Text comprehension. A text can be stored in the computer and access to it achieved through questions. A German teaching program incorporating an ATN parser by Weischedel and others presents the students with a text and then questions them about it (Weischedel et al., 1978). The example text consists of a dialogue between Fraulein Moreau, a German student, and Mr. Brown, an American. The students are asked questions such as “Wo hat Fraulein Moreau deutsch gelemt?“; they might answer “Sie hat es gelernt in der Schule;” the parser informs them of their grammatical mistake Error: past participle must be at end of clause. But the parser also analyses their sentences semantically; if it asks Wer ist Studentin? and the students answer “Fraulein Moreau ist Student,” they are told Error: look up the meaning of Student i.e. they have supplied a masculine rather than a feminine form). Finally it decides whether the answers are factually accurate; if it asks Wer spricht gut deutsch? and the students answer “Herr Brown spricht gut deutsch, ” they are told Error: false - check the dialogue again (he is in fact the American). Thus the NPL allows the computer to contribute more to the teaching of texts than in the word-guessing exercises mentioned above; rather than matching the students’ responses against present answers, it flexibly handles whatever is typed in.

(b) Communicative teaching exercises. In the current communicative approach to language teaching particular emphasis is placed on the exchange of information in the classroom. Many communicative exercises depend upon the participants solving a problem by sharing different information with which they have been supplied. MENU (Cook, 1984a) is a short example of such an information gap exercise in which the students order a restaurant meal by discovering which items on the menu are off and which are available. The addition of a NPL component makes such exercises more effective. DINNER PARTY (Cook, 1984b) for instance requires the students to design a menu for three guests they have invited to dinner. They must establish what the guests like to eat and what they are allowed to eat, i.e. whether they are vegetarians, vegans, and so on; to do this they ask questions such as “Does Christopher eat steak?, ” “What does Helen like to eat?,” or “Who likes beans?”

The program has a parser that deals with a range of question forms (technically a topdown, depth-first parser written in PROLOG), a database storing specific information about the guests, and a linking section that draws logical connections between the question and the database. A sample exchange is given in Appendix 1. DINNER PARTY shows that QA systems can be written for typical communicative exercises. The use of databases is enhanced by a natural interface with the data that practises the students’ language as well as their information retrieval abilities.

Dialogue systems provide a connected, coherent series of exchanges with the user that resembles an ordinary conversation. The earliest dialogue program, ELIZA (Weizenbaum, 1966), simulates a non-directive psychotherapist; the human user plays the part of the patient. ELIZA employs two basic strategies. One is to match the user’s input against a list of key-words; successful matches trigger preset responses; thus if the computer matches sister, it will respond Tell me more about your family. The other strategy is to feed the user’s input back in an altered form; the statement I have been very depressed lately is transformed into the response You say you have been very depressed lately. PARRY (Colby et al., 1971; Colby, 1975), which simulates a paranoid patient, works in a similar manner to ELIZA, generating responses about horseracing, bookies, and gambling, interpreted in terms of Mafia persecution. Such programs, however, only give an illusion of conversation since they have only superficial understanding, and no representation of discourse; they are dealing with language without reference to its structure or meaning.

NLU programs have, however, been developed to deal with conversation at a deeper level, in particular to handle those in which one participant has the purpose of obtaining functional information from another, as a waiter, for example, asks restaurant customers what they want to order. Simple purpose-oriented dialogue structures of this kind have been represented in NLU through the notion of a frame, “a data structure for representing a stereotyped situation” (Minsky, 1975, p. 212). A frame-based program called GUS (Bobrow et al., 1977) books customers on airline flights from Palo Alto to various destinations.

To achieve this, it has dialogue frames that it has to fill out with information. Its initial frame requires information about the client, the date, and the topic. It then explores a tripspecification frame which needs information about destination, return trip, preferred times, and so on. Finally it suggests possible flights and books a flight if the client is satisfied. In a sense it is continually exploring the gaps in its knowledge and trying to elicit the missing information, until it knows enough to make a suggestion to the customer. The advantage of ELIZA-type programs for language teaching is that they simulate some of the properties of ordinary conversation. The ELIZA-based program CHATTERBOX (Cook and Hamilton, 1984) takes students through a university admission interview; using the two ELIZA strategies, it generates appropriate responses to the student’s input; a sample exchange is given in Appendix 2. An authoring section allows the teacher to set up alternative situations by adding suitable keywords and responses, e.g. changing the interview to a dentist’s or a clothes shop. ESCAPE FROM UTOPIA (Cook, 1984c) uses ELIZA strategies within a games context dealing with situations such as the hotel, the cafe, and the station, in each of which has a mini-ELIZA program, ESCAPE then goes a stage further than CHATTERBOX in simulating the types of functional situation. It also illustrates the use of a parser within an adventure game format, familiar from commercially available programs such as THE HOBBIT (1984). Labelle and his colleagues have combined adventure games with a parser for language teaching (Dolbec, et al., 1984); in HERCULE ROBOT for example the robot detective cannot ask questions himself so that the students have to make them up from the cues provided.

The teaching programs described so far already extend the original ELIZA idea by incorporating elementary ideas of conversational sequence and purpose. Inevitably this highlights the need to take account of the structure of the conversation, as in frame-based approaches. A current project at the University of Essex is investigating the use of frames in communicative exercises that consist of a chart giving information, a sample conversation, and practice materials in which students improvise similar conversations using the information in the chart. One version is the familiar restaurant scene in which students act out exchanges between the customer and waiter; other versions are finding the way from a map, getting tickets for a theatre, and so on. STATION (Cook, 1985b) contains much of the QA system seen in DINNER PARTY, but places it in the context of a sequenced purposeful conversation: the program has to know that Which platform? refers to the train that has just been mentioned, and that How much? goes with the price of that particular journey. A sample exchange is included in Appendix 3. Essentially STATION resembles GUS adapted to language teaching, although based on computing techniques specific to PROLOG. GUS, however, has to handle the real world and real native speakers, both of which are complex and unpredictable. STATION deals with an abstract limited world, not just to make the computing work easier, but to parallel the simplified tasks that learners meet in the classroom and the simplified language they use.

A second difference from GUS is that, while normal QA systems are concerned with providing the user with information, STATION teaches the means of getting the information rather than the information itself codebreaking rather than decoding. As with syntax, the computer’s limitations in real world operation are no handicap in the classroom.

The chief use of NLP techniques is then to accommodate the computer within teaching methodologies that stress information handling and interaction. A parallel paper develops the links suggested here between NLP and communicative language teaching (V. J. Cook, forthcoming). In this paper we have seen how NLP techniques can extend the range of computer exercises. This shows one way out of the narrow constraints at present evident in computer teaching of language. The applications described here are largely potential rather than actual; the teaching programs are small-scale and experimental. But as Benjamin Franklin said “What’s the use of a new born baby?”

BATES, M. (1978) The theory and practice of Augmented Transition Network Grammars. In Natural Language Communication with Computers, L. Bolt (ed.), New York: Springer, pp. 191-260.

BOBROW, D. G., KAPLAN, R. I., KAY, M., NORMAN, D. A., THOMPSON, H., and WINOGRAD, T. (1977) GUS, a frame-driven dialog system, Artificial Intelligence, 8 (2), 135-173.

COLBY, K. M. (1975) Artificial Paranoia: A Computer Simulation of Paranoid Processes, New York: Pergamon.

COLBY, K. M. WEBER, S., and HILF, F. D. (1971) Artificial paranoia, Artificial Intelligence, 2, l-25.

COOK, V. J. (1982) Structure drills and the language learner, Canadian Modern Language Review, 38.

COOK, V. J. (1984a) Communication games with a microcomputer, World Language English, 312.

COOK, V. J. and HAMILTON, A. (1984) CHATTERBOX, program for BBC B microcomputer, The British Council.

DE ROECK, A. (1983) An underview of parsing. In Pursing Natural Language, M. King (ed.). London: Academic Press, pp. 3-17.

DOLBEC, J., LABELLE, F., and McCREESH, B. (1984) HerculeRobot and other stories: Adapting the Adventure Game to Computer Assisted Languoge Learning, mimeo, Universite de Quebec a Chicoutimi.

HENDRIX, G. C., SACERDOTI, E. D., SAGALOWICZ, D., and SLOCUM, J. (1978) Developing a natural language interface to complex data, ACM Transactions on DB Systems, J. Ass. Comput. Mech 3 (2), 105-147.

HIGGINS, J. and JOHNS, T. (1984) Computers in Languuge Learning. London: Collins.

KAPLAN, R. M. (1970) The MIND System: a Grammar Rule Language: Rand Corporation Software, Santa Monica.

KAY, M. (1976) Experiments with a powerful parser, American Journal of Computational Linguistics, microfiche No. 43.

LORITZ, D. J. (1984) Natural language processing on micros. Paper given to the Eighteenth Annual TESOL Convention, Houston.

MARKOSIAN, L. Z. and AGER, T. A. (1983) Applications of parsing theory to computer assisted instruction, System 11, 65-77.

MARTY, F. (1982) Reflections on the use of computers in second language acquisition II, System 10. l-11.

MERRIL, M. D. (1980) Learner control in computer-based learning, Computers in Education 4, 77-95.

MINSKY, M. (1975) A framework for representing knowledge. The Psychology of Computer Vision. In P. H. Winston (ed.), New York: McGraw-Hill, pp. 211-277.

PETRICK, S. R. (1973) Transformational analysis. In Natural Language Processing, R. Rustin (ed.). New York: Algorithmics Press, pp. 27-41.

PUL.MAN, S. (1984) Limited domain systems for language teaching. Paper presented at COLING.

SAGER, N. (1981) Natural Language Information Processing: A Computer Grammar of English and its Application. Addison-Wesley.

THOMPSON, H. (1981) Chart parsing and rule schemata in GPSG, Proceedings of the Nineteenth Annual Meeting of the Association for Computational Linguistics. ACL Stanford, California.

WEISCHEDEL, R. &I., VOGE, W. M., and JAMES, M. (1978) An artificial intelligence approach to language instruction, Artificial Intelligence 10, 225-210.

WEIZENBAUM, J. (1966) ELIZA - a computer program for the study of natural language communication between man and machines, Communications of the ACM 9, 36-15.

WINOGRAD, T. (1972) Understanding Natural Lunguage. Edinburgh: Edinburgh Llniversity Press.

WOODS, W. A. (1970) Transition network grammars for natural language analysis, Commons. Ass. Comput. Mach. 13, 591-606.

WOODS, W. A. (1973) An experimental parsing system for transition network grammars. In Natural Language Processing, R. Rustin (ed.), New York: Algorithmics Press, pp. 112-15-t.

Program: Could you tell me a bit more about your education to start with, Mr Smith?