Chapter for K Tobin & B Fraser, (Eds). International Handbook of Science Education (Kluwer)



Increasingly, the data of science education research are verbal data: transcripts of classroom discourse and small group dialogues, talk-aloud protocols from reasoning and problem-solving tasks, students' written work, textbook passages and test items, curriculum documents. Researchers wish to use data of these kinds to describe patterns of classroom and small-group interaction, development and change in students' use of technical language and concepts, and similarites and differences between school and community cultures, school science and professional science, the mandated curriculum and the delivered curriculum. In a short chapter is it not possible to demonstrate actual state-of-the-art techniques of linguistic discourse analysis. My purpose here will be to formulate the issues and choices of which researchers should be aware in adopting and adapting any method of analysis of verbal data for their own work. Along the way I will cite examples from my own published work and other sources which I personally find useful. Discourse analysis is a very large subject; its principles embody a theory of meaning-making that is nearly co-extensive with a theory of human behaviour and human culture (Lemke 1995a). (For other useful introductions to discourse analysis and classroom discourse study, see Brown & Yule 1983; Cazden 1986, 1988; Coulthard 1977; Coulthard & Montgomery 1981; Edwards & Westgate 1994; Stubbs 1983; Widdowson 1979; Wilkinson 1982)

How Researchers Construct Verbal Data

The language people speak or write becomes research data only when we transpose it from the activity in which it originally functioned to the activity in which we are analysing it. This displacement depends on such processes as task-construction, interviewing, transcription, selection of materials, etc., in which the researcher's efforts shape the data. Because linguistic and cultural meaning, which is what we are ultimately trying to analyse, is always highly context-dependent, researcher-controlled selection, presentation, and recontextualisation of verbal data is a critical determinant of the information content of the data. Data is only analysable to the extent that we have made it a part of our meaning-world, and to that extent it is therefore always also data about us. Selection of discourse samples is not governed by random sampling. Discourse events do not represent a homogeneous population of isolates which can be sampled in the statistical sense. Every discourse event is unique. Discourse events are aggregated by the researcher for particular purposes and by stated criteria. There are as many possible principles of aggregation as there are culturally meaningful dimensions of meaning for the kind of discourse being studied.

The basis for aggregation, ultimately, is covariation: some change in the context or circumstances is associated with a systematic change in discourse features of interest to the study. Normally this cannot be known until the end of the study, so it is wise to collect a larger and more diverse corpus of verbal data than will ultimately be used to support the analysis. The basis of discourse analysis is comparison. If you are interested in covariation between text features and context features, you should not collect data only for the cases of interest, but also for cases you believe will stand in contrast with them. If, for example, you are interested in phenonema specific to women, to third-graders, to small-group discussions in lab settings, or to a particular curriculum topic, you should also collect potential comparison or reference data, in small amounts, for other genders, grades, settings, or topics.

Discourse analysis is also contextual. If you are interested in the language of any particular kind of event or text, you should also collect "around" it its probably relevant intertexts (see below). If you are studying how students write up their lab work, in addition to the texts they write, you will also need data on how the same topics have been discussed in whole class sessions, what the textbook says on the topic, any relevant written handouts, and perhaps also interviews with the teacher and the students. All analysis is reductive. Information from the original data is discarded in the process of foregrounding the features of interest. Wise researchers preserve the original data in a form that can be re-analysed or consulted again from a different viewpoint, posing different questions.

Spoken language is never analysed directly. It is not even often analysed directly from audio or video recordings, but from written transcriptions. The process of transcription creates a new text whose relations to the original data are problematic. What is preserved? What is lost? What is changed? Just the change of medium from speech to writing alters our expectations and perceptions of language. What sounds perfectly sensible and coherent may look in transcription (any transcription) confused and disorganised. What passes by in speech so quickly as not to be noticed, or is replaced by the listener's expectations of what should have been said, is frozen and magnified in transcription. Normal spoken language is full of hesitations, repetitions, false starts, re-starts, changes of grammatical construction in mid-utterance, non-standard forms, compressions and elisions, etc.

The tendency in transcription is to "clean it up", dismissing most of these features as irrelevant. Very often some of them turn out not to be irrelevant at all. I recommend transcribing large portions of the corpus at the "lexical" level (preserving the sequence of whole, meaningful words and meaningful non-lexical vocalisations) for survey purposes, and smaller portions at more detailed levels for more intensive analysis. The simplest transcriptions attempt to preserve information at the level of the word, but language only occasionally constructs meaning with single words. What matters is how the words are tied together, and that often includes intonation contours. Whether two phrases represent self-paraphrase or contrasting meanings often can only be determined from intonation. Transcription at the level of the word also erases information about emphasis, value-orientation, degree of certainty or doubt, attitude of surprise or expectability, irony, humor, emotional force, speaker identity, and speaker dialect or language background. Many of these features are often redundantly coded in the words as well, but some may not be. In addition, information about the timing of speech (length of pauses, simultaneous speech, sudden breaking-off of fluency, overlaps, etc.) is often important.

Written texts carry considerable visual information: handwriting forms, page layout, typography, accompanying drawings and illustrations, etc. This information, which can be very important for interpreting the meaning of verbal text, should not be lost to the analysis. Videotapes obviously contain a wealth of relevant visual information on gaze direction, facial expression, pointing and other gestures, contextual artifacts referred to in the verbal text, positional grouping, relative distances and directions, etc. Along with fieldnotes, they help us to reconstruct the social situation or cultural activity type within which some meanings of the verbal language are very much more likely than others. For useful discussions of transcription, see (Ochs 1979; Sacks et al. 1974). For the role of intonation, see (Halliday 1967, Brazil et al. 1981). On visual information in text, see (Bertin 1983, Kress & van Leeuwen 1990; Lemke in press-a; Tufte 1983).

The Contexts of Verbal Data

Language is always used as part of a complex cultural activity. Verbal data make sense only in relation to this activity context and to other social events and texts with which we normally connect them, their intertexts. Meaning is not made with language alone. In speech it is accompanied by gestural, postural, proxemic, situational and paralinguistic information; in writing by choices in the visual coding of words and other graphical information. The meaning of any text or discourse event always depends on how we connect it to some (and not other) texts and events (on general intertextuality see Lemke 1985, 1988a, 1993). What the teacher is saying now makes sense in part in relation to what she said ten minutes ago or yesterday, what we read in the book, the question you missed on the last quiz, etc. It also makes sense differently depending on whether she is reviewing or introducing new material, whether it is addressed to one student or to the whole class, whether it relates to a diagram on the board or not. What a student says may make meaning in relation to the past history of his dialogue with this teacher, the group dynamics of the class, his boredom with the topic, his personal relations with other students.

There are many schemes for systematising the probably relevant contextual factors of a text or discourse event (see for example Erickson & Shultz 1981; Hymes 1972). They all include: the participants and their social and physical relationships, material objects and semiotic representations in the immediate physical environment, the cultural definition of the activity type or situation and its roles and expectations, and the channel or medium of communication. More important than such lists are (1) the principle that the discourse itself can create a context, make a part of the environment newly relevant, or even change its meaning; and (2) that the context is itself a kind of text, it must be "read" from the viewpoint of the verbal discourse. Verbal data, including particularly written or printed texts, always makes sense in relation to (1) a context of production, the circumstances in which it was written or spoken, and (2) a context of use, those in which it is read or heard. For written texts these two can be very different (see Lemke 1989a). Texts and discourse data index or point to relevant contexts in a variety of ways (see Silverstein 1976; Wortham 1992, 1994). The simplest is through deictic forms such as this, that, the other, over there, now, as we saw before, mine etc. These forms indicate to the listener that meaning must be made jointly with the textual and the relevant contextual information. In addition to the context of situation, there is also more generally the context of culture (see Firth 1957; Halliday & Hasan 1989; Hasan 1985; Malinowski 1923, 1935) that is indexed by a text. Much of this is a presupposition of familiarity with other texts, cultural norms, genre conventions (see below), etc. in a particular community. Nonverbal signs which co-occur with spoken language, especially "body language" signs form, with speech, a single integrated meaning-making and interpersonal communication system. Very little is really known yet on how the different channels of this system modulate each other's meaning effects, but see Kendon 1990; Lemke 1987; Scheflen 1975.

The Dimensions of Verbal Meaning

Language in use always creates three interdependent kinds of social and cultural meaning. It constructs social relationships among participants and points-of-view; it creates verbal presentations of events, activities, and relationships other than itself; it construes relations of parts to wholes within its own text and between itself and its contexts.

Presentational meaning is the most familiar and most studied. This aspect of meaning is often referred to as representational, propositional, ideational, experiential or thematic content. This is the function of language for presenting states-of-affairs, for saying what is going on. It presents processes, activities, and relationships; the participants in these processes, and attendant circumstances of time, place, manner, means, etc. It defines entitites, classifies them, ascribes attributes to them, counts them. In relation to these semantic functions, its grammar has been usefully described by Halliday (e.g. 1976, 1985; see also Martin 1992). My own work on thematic patterns or formations (Lemke 1983, 1988a, 1990a, 1995b) applies Halliday's analysis to textual and intertextual patterns in discourse (see below).

Orientational meaning may be even more fundamental developmentally. This aspect of meaning, also called interpersonal or attitudinal, constructs our social, evaluative, and affective stance towards the thematic content of our discourse, towards real and potential addressees and interlocutors, and toward alternative viewpoints. It includes the language of formality/intimacy, status and power relationships, role relationships; speech acts such as promising/threatening, joking, insulting, pleading, requesting/demanding, offering, etc.; evaluative stances toward the warrantability, normality, normativity, desirability, seriousness, etc. of thematic content; construction of affective states; and construction of alliance, opposition, etc. between one theory or viewpoint about a matter and others available in the community. Useful sources on these aspects of orientational meaning include: Austin 1962; Bakhtin 1935, 1953; Halliday 1978, Hasan 1986, Hasan & Cloran 1990; Martin 1992; Lemke 1988a, 1989b, 1990b, 1992; Poynton 1989; Thibault 1991.

Organisational meaning is not always perceived in our culture as meaning, but analysis shows that it is an integral member of the team, functioning together with, and indeed enabling, the other two. Organisational meaning includes the ways in which language creates wholes and parts, how it tells us which words go with which other ones, which phrases and sentences with which others and how, and generally how a coherent text distinguishes itself from a random sequence of sentences, phrases, or words. Organisational meaning in language is generally created through simultaneous use of two complementary principles: (1) constituency structure, in which a larger meaning unit is directly made up of contiguous smaller units, and (2) cohesive structure, or "texture", in which chains of semantic relationships unite units which may be scattered through the text. Constituency structures may be interrupted and resume, and are at least in principle "completable". Cohesion chains, which have neither of these properties, are built on a variety of chain-membership principles, all of which specify a particular kind of relation of meaning among the items (e.g. synonyms, members of a common class, contrast, agent-action, action-means, attribute-item, etc.). Constituency structures (genres, genre stages, rhetorical formations, adjacency structures, clause-complexes, clauses, phrases, groups, etc.) create local meaning relationships among items which also generally belong to cohesion chains, and they provide one means for creating new bases for cohesive relations. Real texts, especially extended complex discourses, often change genre types or other constituency strategies many times, creating sub-units within a text. Cohesive relationships provide a principal means of creating semantic continuity across these segmental boundaries within a text. Note that some forms of meaning depend about equally on two of these three semantic functions, so that, for example, logical relationships (because, if ... then) normally function both presentationally and organisationally. For useful discussions of organisational meaning, see: Halliday 1978; Halliday & Hasan 1976, 1989; Hasan 1984; Lemke 1988b, 1995b; Martin 1992; Matthiessen 1992).

Semantic Content Analysis

How can we characterise what a text says about its topics, or even what its topics are, better or more concisely than the text does itself? This is possible only to the extent that the text repeats the same basic semantic patterns, makes the same basic kinds of connections among the same basic processes and entitities again and again. It happens, in our culture, and probably in most, that not only do we repeat these thematic patterns, or formations, again and again in each text, merely embroidering on the details, we also do so from one text or discourse event to another. This is especially true in the sciences and other academic subjects where there are accepted, canonical ways of talking about topics. Most textbooks will tell you pretty much the same thing about atoms, or alternating current, or Mendelian inheritance. We expect that, however they present it, what teachers say about these topics will contain this same information, and that when students reason, talk, write, or take tests, that their discourse will fit these patterns, too (at least eventually).

The common techniques of concept mapping are based on our ability to consciously abstract the essential meaning relations among key terms in scientific discourse. Discourse analysis, however, can produce the same patterns, and be more semantically explicit about their content, from free-form classroom or small-group talk, or from written materials of any kind. This means that these direct uses of scientific concepts can be directly sampled, assessed, and compared. The basic technique for doing this is described in Lemke (1990a) and its linguistic basis and extensions are discussed more fully in Lemke (1983, 1988a, 1995b). Other forms of modern semantic content analysis are statistical, corpus-based, and collocational (see e.g. Benson & Greaves 1992; Halliday & James 1993; Sinclair 1991). Given the present limitations of computer analysis of natural language texts, these analyses are based on forms rather than meanings. They can tell you the frequency distributions, and more importantly the joint distributions for pairs (or n-tuples) of words or fixed phrases, in a text. They cannot tell whether a given word is used with the relevant meaning you are interested in in any particular instance. Thematic analysis, correspondingly, must be done by hand, but it enables you to see that the same concept or relationship may be expressed by many different verbal forms and grammatical constructions, and to exclude cases where the form is right but the meaning in context is not. To do thematic analysis properly, you need to be familiar with both the subject matter content of the discourse or text, and with the semantics of at least basic lexical and grammatical relations at the level of Halliday (1985) and Hasan (1984).

Rhetorical Interaction Analysis

All language in use, whether spoken or written, is explicitly or implicity dialogical; that is, it is addressed to someone, and addresses them and its own thematic content, from some point-of-view. It does rhetorical and social work, producing role-relationships between author-speaker and reader-hearer with degrees of formality and intimacy, authority and power, discourse rights and obligations. It creates a world of value orientations, defining what is taken to be true or likely, good or desirable, important or obligatory.

Some useful questions to guide rhetorical analysis include: What are these people trying to accomplish here? What are they doing to or for one another? How is the talk ratifying or changing their relationships? How is it moving the activity along? How is it telling me what the speaker/writer's viewpoint is? What is it assuming about my viewpoint? other viewpoints? How does it situate itself in relation to these other viewpoints? What is its stance toward its own thematic content, regarding its truth or probability, its desirability, its frequency or usuality, its importance, its surprisingness, its seriousness, its naturalness or necessity? Rhetorical analysis needs to be done at each organisational level of the text. What is the function of the choice of genre as a whole (see below)? of each stage in the unfolding of the genre? of the local rhetorical formation and each move within it? of the sequencing of formations and topics? of various interruptions, digressions, and the timing of returns? Of grammatical contructions? Of word choices? Of pauses, intonations, marked pronunications?

Given a particular thematic content, there are an endless variety of grammatical (including non-standard) ways to word it. Each variation fits the need of some rhetorical situation, or helps us to construct one. While genres or common rhetorical patterns provide a definite set of expectations, they also allow or encourage considerable strategic and tactical manoevering (see examples in Lemke 1990a, chapters 1 and 3). It is not always possible to say what a particular choice or move means, but you can say what it might mean. And frequently it means more than one thing, plays a role in more than one action or strategy, even in unconscious ones. Those features of a rhetorical analysis which rely, as thematic analysis does, on patterns that are commonly found in many texts, tend to be agreed on by different analysts. But rhetorical analysis must deal with situations unique to the text at hand much more often, and these are more ambiguous and subject to different interpretations. In these cases multiple forms of evidence need to be used to support interpretations: word choice, intonation, grammatical choice, contextual information about the situation or activity. Even the participants in a discourse may disagree about the rhetorical meanings of particular features, or change their minds in retrospect or with additional information. The "intention" of the speaker, revealed in a retrospective interview, is just one more piece of data; it does not settle the question of what a feature meant for any participant at the time. Evidence of how participants followed-up on the appearance of the feature may be more persuasive.

Discourse forms do not, in and of themselves, "have" meanings; rather they have a range of potential meanings. Words, phrases, sentences are tools that we deploy in complex contexts to make more specific meanings, to narrow the potential range of possible meanings down to those reasonably or typically consistent with the rest of the context. Even in context, at a moment, an utterance or phrase may not have a completely definite meaning. It may still express a range of possible meanings, differently interpretable by different participants or readers. This is very often the case at the point where it occurs. The context needed to specify its meaning very often at least partly follows its occurrence. So it may seem to have a more definite meaning retrospectively than it has instantaneously. In fact, depending on what follows, its meaning, as participants react to it, can be changed radically by what follows (retrospective recontextualisation). Analysing a text to see what is happening to meanings moment-to-moment yields a dynamical analysis; the overall net retrospective meaning when all is said and done, yields the synoptic analysis.

For a variety of good examples of rhetorical or speech act analysis, see Gee 1990 (esp. chaps 4 & 5); Green & Harker 1988; Grimshaw 1994; Lemke 1990a; Mann & Thompson 1988, 1992; Mehan 1979; Sinclair & Coulthard 1975. For discussions of evaluative and affective meaning, see Lemke 1988a, 1989b, 1990b; Martin 1992. For viewpoint analysis see discussions of heteroglossia in Bakhtin 1935; Lemke 1988a, 1995a, 1995b; and of social voices in Wertsch 1991. For dynamic and synoptic analysis see Lemke 1984, 1988b, 1991; Martin 1985, 1992; Ventola 1987.

Structural-Textural Analysis

Verbal data has social meaningfulness only as text, not as collections of isolated words or phrases (except statistically). How does a coherent, cohesive text differ from a random collection of grammatical sentences? How are texts and discourse events unified and subdivided into wholes and parts? How can we define the boundaries of a unit or episode of a text or verbal interaction? What binds the units of a text together?

Structural analysis of texts needs to be both "top-down" and "bottom-up", that is, it needs to consistently reconcile analyses that begin from the smallest units of meaning (normally phrases and clauses) and look for how these aggregate together into larger units, with analyses that begin from the largest units (normally activities and episodes or genres and their stages) and look for how these are composed of functional constituents. The largest unit of analysis for a spoken discourse text is the socially recognised activity-type in which the discourse is playing a functional part, or the smallest episode or subunit of that activity which contains the entire discourse event. A classroom Lesson is a typical activity-type of this kind. An episode of Going-Over-Homework or Working-in-Groups may form the more immediate context. The largest unit for a written text is normally the genre of which it is an instance, or the text itself (unless it is being analysed directly in its context of production or use, in which case the activity type applies, e.g. Lemke 1989c).

A genre is a text-type specified by identifying a common structure of functional units (obligatory and optional) that is repeated again and again from text to text. A speech genre is a generally a highly-specific activity-type accomplished mainly by verbal means. The term genre is more often used for types of written texts because they are more structurally standardised in our culture. A genre has a constituency structure in which each constituent plays a functional role in the whole and has specific functional meaning relations to the other constituents on its own level. The largest units are often called stages, and they may be composed of smaller units, and these of still smaller ones, etc. Each constituent at each level of analysis should be defined in a way which is unique to the genre. A science lab report, as a written genre, might have major stages such as: Title, Author, Class, Statement of Problem, Description of Apparatus, Description of Procedures, Record of Observations, Analysis of Data, Conclusions, etc. The Description of Procedures might include a series of Procedure Statements, each saying what was done, when, and how. Each of these might not be composed of smaller genre-specific functional units, but only of grammatical units (i.e. the relationship changes from "composed of" to "realised by").

Some constituents of some genres have an intermediate level of organisation between genre-specific units and grammatical ones. These are often called rhetorical structures or formations (e.g. Lemke 1988b; Mann & Thompson 1988, 1992). They are found in essentially the same form in many different genres, but they have an internal functional or rhetorical structure in addition to the structure of their grammatical units. The most famous example in classroom discourse analysis is the IRF structure, typically realised as Teacher Question, Student Answer, Teacher Evaluation (see Lemke 1990a; Mehan 1979; Sinclair & Coulthard 1975). This could be considered genre-specific in classroom activity, but since it occurs in many different kinds of episodes (Lemke 1990a), it is more nearly a rhetorical formation. Something very similar occurs in courtroom discourse as well. More common and widespread examples include the simple Question-Answer pattern, or Examples-Generalisation, Event-Consequences, syllogisms, etc. They are, in effect, portable mini-genres.

Conversation analysis techniques often refer to particular cases as "adjacency pairs" (e.g. Sacks et al. 1974) Below the level of smallest genre-specific units and the moves within a rhetorical formation, we find the level of grammatical structure. Analysts should be aware that there are multiple simultaneous grammatical units structuring the same set of words, and that some of these may depend on intonation as well as word sequence. In Halliday's analysis, for example, a clause is simultaneously structured in terms of

Other models also have adopted the multi-structure approach (e.g. Chomsky analyses in terms of thematic roles, X-bar structures, and Logical Form). The boundaries of these different units are not necessarily the same. This brings us to the classic problem of textual structure: segmentation. Can a text be definitively divided at word boundaries into its constituent units at any level of analysis? The answer is: only sometimes. The same word can function as an element in different units, for different functions, on different scales. The boundary, particularly of a large, high-ranking unit (e.g. genre stage, rhetorical move) may be indeterminate in terms of lower level grammatical or word units because it is defined by several simultaneous criteria, each of which results in drawing the boundary in a slightly different place in the text.

As a general rule, units of meaning can have fuzzy boundaries in terms of units of form (or even in terms of units of meaning at a different level of analysis). Some texts are more rigidly structured than others. Some maintain, repeat, and complete particular genre patterns or rhetorical formations more consistently than others. Many texts frequently shift genre pattern or rhetorical strategy, with or without completion of those already started. Conversational discourse is notorious in this respect, but so are written texts by young writers who have not yet learned the genre conventions and borrow from the norms of conversational organisation. How do such texts maintain their coherence? In large part by topic continuity. More generally, by maintaining cohesion chains, whose members have no consistent structural-functional relations. If a structure looks like A-B-C-D, a chain looks like A-A-A-A.

Chains may be of many kinds. Lexical chains consist of words each of which may be the same word, have the same meaning in context, refer to the same referent, belong to the same semantic domain, etc. A short lexical chain may be accidental; a long one rarely is. Larger units than words may form chains, or strands. A structural pattern may be repeated (cf. rhetorical parallelism): A-B-C-D, A-B-C-D, A-B-C-D, etc. More commonly, and very importantly, a thematic pattern may be repeated, and varied, at different levels of abstraction (see Lemke 1995b for an extended analysis); this is very common in classroom teaching, and indeed in any text that has a lot to say about a small topic. Chains also normally interact with one another; that is, in each instance from two different chains, there is the same structural relation each time between the member of one chain and the corresponding member of the other. If we had a A-B structure, we could have an A-chain through the text and a B-chain, united by the fact that each A and B (or even just some of them) were connected in the A-B relation where they occurred in the running text (rather like a ladder). Not just chains of individual lexical items, but chains of whole thematic formations, can interact. It may take only a clause or nominal group (noun phrase) structure to tie members of two lexical chains together, but it can take much larger and more complex grammatical or rhetorical structures to do this between large thematic formations (see Lemke 1995b).

For further discussions of genre analysis see Bazerman 1988, 1994; Hasan 1984b, 1989; Martin 1989, 1992; Lemke 1988b, 1991; Propp 1928; Swales 1990; for activity-types see Lemke 1990a; Phillips 1983 (on participant structures); for rhetorical formations Lemke 1988b; Mann & Thompson 1988; for conversation analysis Sacks et al. 1974; for cohesive organisation, Halliday & Hasan 1976; Hasan 1984; Lemke 1988b, 1995b.

Case Studies and the Problem of Generalisability

How can verbal data and discourse analysis be used in studies of individual episodes and lessons, classrooms, and small groups? What is the value of such studies and how can we determine the generalisability of their findings?

Discourse analysis studies are often best when they examine a particular community in depth. Discourse analysis produces its greatest insights when rich contextual information can be factored into the analysis of each text or episode. For this reason, longitudinal designs or case studies are well suited for discourse analysis methods. Here we learn a great deal about a particular class, seeing repeated patterns within the data and a variety of strategies which create variations on those patterns. It is even possible to proceed to the level of individual psychological analysis, but in a larger sense all case studies are about "individuals", and to a much greater extent than is true for other natural systems, human communities' individuality matters for the kinds of behaviors of the system that interest us.

It is not true that science should be only about generalised properties of classes of phenomena and not about unique properties of individual instances. The balance between these two approaches must be struck differently depending on the nature of the phenomena. Electrons seem to have no individuality that matters; biological systems do, but a great deal of their structure and behaviour remains constant for a species or variety. Developmental phenomena show a wide range of individual pathways, many of them approximately "equifinal", leading to the same end result (for example, language acquisition). Human communities and cultures are often more interesting for what is unique to them than for what they all have in common. Moreover, one of the important properties of any class is precisely the specification of how the members of the class differ from one another. Many sentences have a lot in common; that is the foundation of grammar. Many texts have a little in common, hence the concept of genre. But while the resources and strategies by which texts and discourse are constructed may be common to many texts, and help to specify how they may differ from one another, what is ultimately of interest about any text is its meaning, and that is its most unique feature. Discourse analysis will not tell us a lot about how all classrooms or all science writing is alike (it will tell us a little), but it provides us with the tools to analyse and understand what exactly is going on in any discourse or text we wish to analyse. That is as much as any theory really does for us in practice.

Protocol Analysis and the Problem of Interpretation

When task activities differ significantly from normal cultural routines, how will cultural patterns of language use be distinguishable from idiosyncratic constructions? What is the object of study that we construct from such data?

One important form of verbal data is generated when researchers construct special task activities that differ significantly from normal cultural routines. This follows the traditions of the natural sciences in devising tasks meant to reveal particular aspects of phenomena, but it encounters the risk (minimal for electrons and molecules, but already significant for organisms) that behavior under task conditions differs in important, and unknown, ways from that in normal routines. The essential context-sensitivity of meaning-based phenomena (meaning is selective contextualisation) strongly suggests that if we are interested in, say, a classroom phenomenon, that we study it in situ. If we supplement this with artificial tasks, it is then necessary to establish empirically that the differences between the task context and the natural context do not alter the phenomena of interest, or to identify in exactly what ways they do alter them.

Current models of situated cognition call into question the assumption that meaning-making processes can be assumed independent of local contexts, or even that "cognition" is a process in a system limited to the organism itself (as opposed to one that includes the organism's tools and the elements of the environment with which it interacts; cf. Lave 1988; Lave & Wenger 1991; Kirshner in press; Lemke in press-b).

Discourse analysis assumes that the resources and strategies (lexis and grammar, rhetorical formations, typical cultural narratives, genres, the principles of constructing thematic formations, cohesion chains, etc.) used in producing discourse events and texts are characteristics of a community, rather than unique to an event in that community. They are part of its general cultural resources (and so differ from culture to culture and one community or subcommunity to another), but what it means to have a culture is that we preferentially deploy some of these resources in some contexts rather than others: that how we use the resources is essentially context-dependent. The analysis of the covariation between situational features and which lexical and grammatical resources are typically deployed in them is the subject of register theory (e.g. Gregory 1967; Gregory & Carroll 1978; Halliday 1977, 1978, 1991), which can also be adapted to analyse the clause-to-clause shifts in meaning that take place through a text (phasal analysis, e.g. Malcolm 1985). A similar theory for the deployment of the other resources does not yet exist (but see Kress & Threadgold 1988; Hasan 1994; Lemke 1995a; Thibault 1991). We should also recognise that while artificial activities may not be natural for the subjects asked to perform in them, they are in a sense a natural part of the culture of the researchers, which is thus mixed with that of the subjects by these procedures. There is perhaps a great deal to be learned about ourselves by analysing the nature of these tasks and their role in our own professional practice.

Comparative Studies and Cultural Bias

When we use discourse analysis and verbal data to compare males and females, middle and lower class subjects, widely differing age-groups, different cultural and linguistic groups, school practices and home, community, or professional practices, we necessarily introduce our own viewpoint, which is invariably closer to that of one of the categories compared than to the other's. Discourse data is not just sensitive to the context of immediate task and situation; it is also sensitive to the wider context of cultural norms and assumptions, knowledge, beliefs and values. The analysis of discourse data, its interpretation, is itself just more discourse, and discourse now from the point-of-view of the researcher's community.

Our research communities and their historical traditions are emphatically not equally balanced by gender, age, social class, or ethnic culture. Even studies which strive mightily for even-handedness and neutrality of description (e.g. Bernstein 1971, 1975; Hasan 1986; Heath 1983) are necessarily read by other researchers who will project their own values regarding what is better and what worse onto their descriptions of difference. In many other studies, even the questions which are asked of the data are asked from a very narrow range of human viewpoints. Discourse analysis is interpretation and it is viewpoint-dependent every bit as much as any other instance of discourse. The canonical procedures of discourse analysis which I have briefly sketched here provide a means for different analysts to systematically compare the many interdependent grounds of their respective interpretations. Whether they reach consensus or not is probably less important than the fact that the procedures be clear enough that others can enter into the discussion on common ground. These procedures, of course, are themselves the product of a narrow range of human viewpoints. We can hope that this range will widen as the field of discourse analysis, and our own society, matures toward more inclusiveness and respect for the value of diversity of viewpoints.

Curriculum Analysis and Evaluative Assessments: Ethical Issues

The methods of discourse analysis of verbal data can be used to compare curriculum documents, textbooks, and tests with classroom dialogue, teacher discourse, student writing, etc. They make possible rich descriptions of the lived curriculum, its relation to official curriculum plans, and the web of intertextuality among all the spoken and written language in which education is framed. They also make it possible to analyse how individual students use scientific language and concepts in a variety of situations, and to make this a basis for evaluative assessments.

These facts raise serious ethical questions regarding the appropriate use of discourse analysis methods in education. Our educational system operates within a larger social hierarchy of power and control. Administrative authorities seek to impose specific curricula on students, using teachers, textbooks, and the rest of the educational apparatus as the means. Their control is only as good as their means of assessing whether or not teachers teach, textbooks print, and students master what the curriculum mandates. From curriculum documents to test questions and responses, nearly all of this is in the form of text and discourse. Discourse analysis methods in principle allow far more precise ways of checking the match or mismatch of these elements than any other form of assessment or accountability.

Hopefully new assessment schemes will place more weight on practical skills and graphical modes of representation, but discourse forms will almost certainly still dominate (see Lemke in press for initial work on the analysis of visual representations and the role of mathematics in scientific discourse). For now, while automation of discourse analysis procedures remains thoroughly primitive, students and teachers who believe they have a right to control the content and directions of their own learning and teaching have little to fear from discourse analysis methods.

Meanwhile, the automation of new information technologies begins to offer at least more privileged students the opportunity to explore the universe of knowledge guided by their own evolving interests (cf. Lemke 1994, 1995c) rather than as prescribed by someone else's curriculum. Discourse analysis methods are already important in computer-based natural language systems for generating, analysing, and sorting texts. They will be even more important as components of the intelligent tutoring systems of the future, which will enable students to explore new information worlds more successfully with their help. Researchers of the next generation will help determine whether discourse analysis methods will be used to empower students in the new century, or more strictly control them.