A Computational Lexicalization Approach

Feng-Jen Yang

Feng-Jen Yang, Computer Science Department, Prairie View A&M University

Prairie View, Texas 77446, USA

Abstract. Fine-Grained lexicalization has been treated as a post process to

refine

the machine planned discourse and make the machine generated language

more coherent and more fluent. Without this process, a system can still generate

comprehensible languages but may sound unnatural and sometimes frustrate its

users. To this end, generating coherent and natural sounding language is a major

concern in any natural language system. In this paper, a lexicalization approach

is presented to refine the machine generated language.

1 Introduction

An obvious difference between a natural language system and an information

management system is the user interface. If a user asks an information management

system the same question twice, it is very likely that the system will respond with the

same answer twice, but a natural language system hardly has this kind of dialogue.

Instead of repeating the same answer, a natural language system tends to adapt

answers according to the user’s understanding and try to reach the dialogue goal [7].

The invention of natural language systems is somehow motivated by the desire to

reform the interaction between human and computer.

As a tutoring system with a natural lang

uage interface, the CIRCSIM-Tutor tries to

simulate human tutoring sessions in the domain of baroreceptor reflex. It has been

tested to be effective and now being used as a class aid for first-year medical students

at Rush Medical College in Chicago.

The baroreceptor reflex is the mechanism i

n charge of regulating blood pressure in

the human body so that it will not go beyond the tolerable range. If something

happens to change the blood pressure, such as a transfusion, hemorrhage or

pacemaker malfunction, the baroreceptor reflex will attempt to regulate the blood

pressure in a negative feedback manner so the blood pressure will go back to a stable

state again.

While using this system the student is presented with a predefined perturbation and

en is asked to predict the qualitative changes in seven physiological variables at

three different chronological stages of the reflex cycle. These predictions are then

used as the basis of a tutoring session to remediate any misconception that the student

has revealed.

In order to simulate the dialogue of human tutors as much as possible and provide

learners

with a coherent and fluent natural language interface, this paper presents a

lexicalization approach as a post process to refine our machine planned discourse. The

Yang F. (2005).

A Computational Lexicalization Approach.

In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science, pages 128-146

DOI: 10.5220/0002558301280146

 SciTePress

discourse planner leaves a certain number of decisions open before surface sentence

generation and I choose five lexical features as the first attempt to improve the quality

of our machine dialogue. These features are chosen because they seem relatively

manageable but particularly important in our domain.

2 Domain Knowledge

The behavior of the baroreceptor reflex can be described by the qualitative influences

among seven physiological variables over three stages. The seven core variables as

they appear in the prediction table are Central Venous Pressure (CVP), Inotropic State

(IS), Stroke Volume (SV), Heart Rate (HR), Cardiac Output (CO), Total Peripheral

Resistance (TPR) and Mean Arterial Pressure (MAP). The three stages in the order of

occurrence are the Direct Response (DR) Stage, which is the time immediately after

the perturbation and before the reflex is activated, the Reflex Response (RR) Stage,

when the changes caused by the baroreceptor reflex begin to take effect, and the

Steady State (SS) Stage, the time after restabilization.

The causal relationships between these variables can be modeled by either direct or

inverse qualitative influence among variables. With a direct influence, increasing the

parameter on the cause side results in increasing the parameter on the effect side or

decreasing parameter in the cause side results in decreasing the parameter on the

effect side. For example, increasing the CO results in decreasing the CVP, but

increasing the CO results in increasing the MAP.

It is possible for a parameter to have two determinants. In such cases, learners have

to think about which determinant is stronger, since the result is based on qualitative

changes. The change in the stronger determinant will dominate the total qualitative

change, even if the other determinant has the opposite qualitative influences. For

example, the SV has two determinants, the CVP and the IS, but the IS is stronger than

the CVP. So if the CVP decreased but the IS increased, the increase in the IS is

stronger than the decrease in the CVP and the SV will still increase.

3 Why Lexicalization?

To benefit from a natural language interface, the tutoring system must be provided

with the properties that make human natural language so effective [8]. With this

concern in mind, CIRCSIM-Tutor tries to imitate the human tutor’s language as much

as possible.

Like most natural language systems, the CIRCSIM-Tutor has a discourse planner

to produce a discourse plan that specifies both the content and overall structure of a

tutoring session. In terms of determining the deep structure, knowing the content and

structure of a dialogue is enough and the discourse planner has been doing a good job.

Nevertheless, to make a dialogue fluent and coherent, knowing only the deep structure

is far from enough. There is still a considerable range of details to form the shallow

structure and feed it to the surface sentence generator. The discourse planner leaves

129

open a certain number of decisions about the surface form of the dialogue to be

generated.

3.1 A Coarse-Grained Lexicalization Example

Figure 1 is an example dialogue before lexicalization, which reveals the lack of

fluency and coherence in our machine dialogues. For example, in the second utterance

of T3, the Inotropic State should be pronominalized in a sense of maintaining the

same discourse focus. Also, the content based acknowledgements in T3 and T5 make

the machine dialogue stilted.

T1: By what mechanism is Inotropic State controlled?

S2: nervous system

T3: Correct, Inotropic State is controlled by the nervous system.

What is the value of Inotropic State in DR?

S4: decreased

T5: Nope, the value of Inotropic State is not decreased in DR.

Remember. Inotropic state is neurally controlled.

What is the value of Inotropic State in DR?

S6: not changed

T7: Correct, the value of Inotropic State is unchanged in DR.

Fig. 1. An Example Dialogue before Lexicalization

3.2 A Fine-Grained Lexicalization Example

One of the important areas of research in computational discourse is finding out what

information is contained in the sequence of utterances but goes beyond the meaning of

individual utterances themselves [4]. To this end, having better lexical usages are

absolutely essential and critical. The goal of this research is to make the machine

dialogue fluent and coherent. I, therefore, have some range of options in deciding

which lexical features to work on. The following features were chosen as the first

attempt of lexicalization, because they seem relatively manageable and particularly

important to make the dialogue in Figure 1 sound more natural and more fluent [15].

To illustrate each lexical feature and its corresponding refinement, the related

dialogue turns are chosen and improved step by step while a feature is discussed.

1. Polite Locutions

The machine tutor always asks questions directly, but human tutors tend to ask

questions politely instead of directly. They frequently use various politeness

locutions, such as

Can you tell me … or Do you know …, before the question. So,

turn T1 can be improved with a softener:

T1: Can you tell me by what mechanism Inotropic State is controlled?

130

2. Abbreviated Variable Names

The machine tutor always spells out variable names while asking questions and

giving acknowledgments, but human tutors are much more likely to use the full

name when the variable is first introduced and then abbreviate or pronominalize it

later. So, turns T3 and T5 can be improved to read:

T3: Correct,

IS is controlled by the nervous system.

What is the value of

IS in DR?

S4: decreased

T5: Nope, the value of

IS is not decreased in DR.

Remember,

IS is neurally controlled.

What is the value of IS in DR?

3. Discourse Markers

Using discourse markers can make clear the logical relationship between two

sentences or the relationship of a sentence to the tutorial goal structure. In turns T3

and T5, the two consecutive utterances belong to different topics, but the machine

tutor does not use discourse markers, while human tutors tend to use discourse

markers like so, now, but, therefore, etc., to move from one discourse focus to

another. So, turns T3 and T5 can be further improved to read:

T3: Correct, IS is controlled by the nervous system.

So, what is the value of IS in DR?

T5: Nope, the value of IS is not decreased in DR.

Remember, IS is neurally controlled.

Now, what is the value of IS in DR?

4. Acknowledgments

In turns T3 and T5, the acknowledgments are both explicit and content-based

which make the sentences sound redundant and stilted. In human dialogues,

acknowledgments following student answers are often reduced to a single word,

appended to the next sentence, or even omitted entirely. Whether an

acknowledgment can be reduced and attached to the next sentence depends on the

syntax of the next sentence, the relation of the next sentence to the answer being

acknowledged, and whether the answer is correct or not. Our human tutoring

transcripts show that usually correct answers are not repeated, but wrong answers

are repeated in a sense of pointing out the student’s misconception. So turns T3

and T5 can be even further improved to read:

T3:

Right.

So, what is the value of IS in DR?

S4: decreased

T5:

No, the value of IS is not decreased in DR.

Remember, IS is neurally controlled.

Now, what is the value of IS in DR?

A special phenomenon of acknowledging the student’s answer is that human tutors

tend to acknowledge the student’s finally correct answer more strongly than usual,

especially when the student has made some mistakes and finally got the correct

answer. So, turn T7 can improved to read:

T7:

Very good.

131

5. Pronouns

In turn T5, the intended variable name has been mentioned in the previous turn. In

this case, human tutors tend to use the pronoun it to refer to the variable

previously mentioned and stay in the same discourse focus. So, the turn T5 can be

improved to read:

T5: No, IS is not decreased in DR.

Remember,

it is neurally controlled.

Now, what is the value of IS in DR?

Generally speaking, these refinements are instances of lexical selection. This is

also an illustration of the fact that lexical variation is not random but planned and

purposeful.

Since the system is using schemata as planning operators, an efficient way of

learning the rules for lexical selection is by searching for examples of lexical usage in

transcripts marked up with tutoring schemata. I search for instances of the same

schema expressed in different ways. After further in-depth analysis of these instances,

I have established rules as a better guidance for lexical selection.

Addressing only the five lexical features discussed above, the dialogue in Figure 1

can be transformed into Figure 2.

T1: Can you tell me by what mechanism Inotropic State is controlled?

S2: nervous system

T3: Right.

So, what is the value of IS in DR?

S4: decreased

T5: No, IS is not decreased in DR.

Remember, it is neurally controlled.

Now, what is the value of IS in DR?

S6: not changed

T7: Very good.

Fig. 2. An Example Dialogue after Lexicalization

The necessity of lexicalization can be justified by comparing the quality difference

of machine generated dialogues with and without lexicalization.

4 Discourse Modeling

One of the major problems addressed in discourse research is:

How does an utterance’s context affect the meaning of the individual utterance or

part of it [4]?

That is why a major result of most discourse analysis is dividing a discourse into

discourse segments. The boundaries of segments have to be determined in a manner

much like phrases group into sentences and sentences group into paragraphs, and so

on. The meaning of a segment encompasses more than the meaning of individual parts

[4]. While segmenting the discourse, the language behavior is also modeled.

132

Many methods have been proposed for analyzing the local discourse context. The

most popular method is annotating a corpus of the type of discourse that you wish to

generate. A set of general instructions for annotating discourse segments and

identifying the purposes of discourse segments was proposed by [9]. By investigating

the relationship between reference and segmentation, Passonneau [11] designed a

protocol for coding discourse referential noun phrases and their antecedents. Other

researchers such as Allen and Core [1], Nakatani and Traum [10] and Brennan and

Clark [2] have also suggested methods for exploring lexical issues.

4.1 Discourse Coherence

A very important research resource in the CIRCSIM-Tutor project is a set of tutoring

transcripts numbered from K1 to K76. These sessions were carried out in a keyboard-

to-keyboard manner by our domain experts and their first-year students in physiology.

This research, like most of our earlier work is based on the study and analysis of these

transcripts.

Our discourse analysis is based on a fundamental discourse theory saying that a

hierarchical organization of discourse around fixed schemata can guarantee good

coherence and proper content selection [6]. When the same idea is applied to the

CIRCSIM-Tutor domain, a set of hierarchical tutoring schemata has been discovered

to model the discourse of tutoring sessions performed by our domain experts and their

students [5]. Based on these schemata, I started thinking about the approaches to

refine our machine dialogue.

If we model the behavior of lexicalization in terms of discourse trees, it deals with

integrating the leaf nodes into a coherent dialogue. This integration is related both to

discourse planning and to surface sentence generation. So, a central problem with the

lexicalization is how to make a smooth connection among the semantic representation,

the pragmatic information, and the surface linguistic phenomena. In other words, the

lexicalization has to consider the alternatives in terms of representing the content of

the participants’ utterances, performing the dialogue acts, and generating the surface

language. These alternatives not only provide a certain level of implementation

flexibility, but also introduce the possibility of optimization at some level.

Since the system is now using schemata to plan the discourse, having a coherent

movement of discourse focus is no longer a problem. The remaining work is to

produce a fine-grained lexicalization. This takes more in-depth of lexical analysis.

5 Lexical Analysis

My lexical analysis is based on the concept that a good discourse theory must be able

to account for the ordering of major discourse constituents and predict the surface

linguistic phenomena that depend on structural aspects of discourse [12]. In other

words, by knowing the structure of the discourse in progress, we should be able to

predict their corresponding surface linguistic usages. I, thus, focused my analysis on

discovering the relationship between a discourse structure and its corresponding

surface language usage. Another useful idea comes from Passonneau’s protocol,

133

especially for the problem of finding the inference relationships between different

discourse segments [11]. The draft of DAMSL [1], which uses a backward looking

function to capture how the current utterance relates to its antecedent, is also a helpful

reference.

The lexical analysis described here is focused on the semantic and pragmatic

relationships among the tutoring schemata as well as looking for special phenomena

of lexical usage in the dialogue context.

5.1 Visualization of Lexical Usage

In order to predict the surface linguistic phenomena from the structural aspects of

discourse, it is more useful to have a method that shows discourse structure and

lexical usage at the same time. This will help the analysis to take both issues into

consideration. I have developed a new representation for lexical usage that allows the

researcher to visualize lexical research. This method begins by representing the

hierarchical tutoring schemata as tables and then maps the lexical items of interest

onto those table entries according to their original positions in the schemata. In this

manner, we can visualize both the discourse structure and lexical usage

simultaneously.

Figure 3 illustrates the visualization of the variable descriptions used by our

domain experts while tutoring the variable TPR in the session K12. The discourse

structure of this dialogue is modeled by a schema called

T-corrects-variable which is

realized by two subschemata, T-introduces-variable and T-tutors-variables, and then

the

T-tutors-variable is realized by T-does-neural-DLR. The T-does-neural-DLR is

further realized by

T-tutors-mechanism, T-tutors-DR-info, and T-tutors-value, and

so on. This process keeps going until each of them is finally realized by a surface

utterance.

T-corrects-variable var=TPR

T-introduces-variable T-tutors-variable

T-does-neural-DLR

T-tutors-mechanism T-tutors-DR-info T-tutors-value

T-informs

T-elicits T-informs T-elicits

T: Now how about

TPR?

S: …

T: By what

mechanism

will

increase?

S: …

… T: So what do you think

about

TPR now?

S: …

Fig. 3. Visualization of Variable Descriptions

In this example, I used typography to indicate the lexical features that interest me.

The variable TPR is marked, along with the anaphoric references to it. The lexical

phenomena here are:

The tutor first uses the abbreviated variable name TPR to bring up this variable to

teach. In the immediately following topic, the tutor uses the pronoun it to refer to

the previous mentioned TPR. After that the tutor goes on to convey some other

134

related explanations and in the final topic the tutor uses the abbreviated variable

name

TPR again to bring back the discourse focus.

When these phenomena applied to lexicalization:

A discourse planned using the schema

T-corrects-variable will always have the

variable introduced in the first topic. So, in the second topic the machine tutor can

always use a pronoun to refer to the same variable and maintain the same

discourse focus. Also, in the sense of making a conclusion, it is appropriate to use

abbreviated variable name to bring back focus in the last topic.

Figure 4 is designed to help us visualize the usage of discourse markers while tutoring

the variable TPR in the session K10.

T-corrects-variable var=TPR

T-introduces-variable T-tutors-variable

T-does-neural-DLR

T-tutors-mechanism T-tutors-DR-info T-tutors-value

T-informs

T-elicits T-informs T-elicits

T: Take the last

one first.

T: Can you tell me

how TPR is

controlled?

S: …

And the predictions

that you are making

are for the period

before any neural

changes take place.

So what do you

think about

TPR now?

S: …

Fig. 4. Visualization of Discourse Marker Usage

The lexical phenomena in this example are:

The tutor uses the discourse marker

And to move from one topic to a semantically

continuous topic and uses the discourse marker

So to mark the final topic as an

appropriate conclusion.

When these phenomena applied to lexicalization:

A discourse planned according to the schema

T-does-neural-DLR will always

have the first two topics semantically continuous. So, it will be always appropriate

to use the discourse marker

And to connect these two topics. Also in the last topic

the tutor has to make a conclusion and the discourse marker

So is a good way to

make this conclusion.

Similarly, Figure 5 is a visualization of the way acknowledgments are used while

tutoring TPR in the session K48. The lexical phenomena in this example are:

For the first two questions, the tutor gives a hint by asking some background

knowledge and moving toward the final question. Fortunately, the student answers

these two hints right. So the tutor uses the explicit word

Right to accept these

answers. Finally, the student figured out the correct answer and the tutor

acknowledged it in a stronger manner to encourage the student and said

Great.

When these phenomena applied to lexicalization:

A discourse planned according to the schema

T-does-neural-DLR will always

have some digression before the student figures out the final correct answer. So, in

the last topic, the machine tutor can acknowledge the student’s answer more

strongly than usual to encourage the student.

135

T-corrects-variable var=TPR

T-introduces-variable T-tutors-variable

T-does-neural-DLR

T-tutors-mechanism T-tutors-DR-info T-tutors-value

T-informs

T-elicits T-informs T-elicits

T: You predicted

that TPR would

increase.

T: What

mechanism

does this?

S: Autonomic

nervous

system.

Right.

T: And during DR

what changes in

ANS activity

occur?

S: none.

Right.

T: So do you want to

change your

prediction:

S: Yes, TPR has no

change.

Great!

Fig. 5. Visualization of the Choice of Acknowledgments

5.2 Result of Lexical Analysis

The purpose of visualization is to gather together all the instances of lexical

phenomena and the contexts in which they occur. I look at two types of context, the

surrounding text and the position within the tutorial dialogue schema. Ultimately, I

have found rules, as addressed in Appendix A, which can be used to as guidelines

towards a finer-grained lexicalization in the CIRCSIM-Tutor domain.

6 Implementation

Lexicalization is a processing after discourse planning and before surface sentence

generation. To form a pipeline from discourse planning to sentence generation as

suggested by Reiter and Dale [13], the interfaces have to be clearly defined.

6.1 The Interface between Discourse Planning and Lexicalization

The discourse planner is using a set of hierarchical schemata as plan operators and the

operators currently in use are stored in a working storage. By consulting the working

storage the lexicalization module can have a copy of the discourse in progress and

apply lexical rules accordingly. Figure 6 is the lisp program template to get a copy of

the current discourse. After executing these codes the variables

w-stage, w-topic, w-

primitive

will be holding the current tutoring stage, topic and primitive, respectively.

(setq w-stage (get-value-from-KB '(w-stage-is ?x)))

(setq w-topic (get-value-from-KB '(w-topic-is ?x)))

(setq w-primitive (get-value-from-KB '(w-primitive-is ?x)))

... and so on.

Fig. 6. Retrieve the Discourse in Progress

136

6.2 The Interface between Lexicalization and Sentence Generation

The sentence generator is using a template generation approach which takes a feature

set and generating a sentence accordingly. For example, feeding the feature set

“((primitive informs) (topic mechanism) (stage dr) (var ((var-name CC))” to the

sentence generator will have the sentence

“CC is under neural control.” generated.

The major steps and their corresponding lisp codes to prepare a feature set for

sentence generation are summarized as follows:

1. Initially the feature set is empty.

(let ((features ()))

2. The feature set could be multi-level. So the program goes on to call subfeature

constructors

to construct subfeatures for all discourse operators currently in use,

such as

(primitive-feature w-primitive), (topic-feature w-topic), (stage-feature w-

stage)

, ... etc., and append them to the overall feature set.

(setq features (append features

(primitive-feature w-primitive)))

(setq features (append features

(topic-feature w-topic)))

(setq features (append features

(stage-feature w-stage)))

... and so on.

3. Each subfeature is then constructed according to each discourse plan operator

currently in use. For example, since there are only two possible values for the

primitive operator, the primitive subfeature can only be either

(primitive elicits) or

(primitive informs).

(defun primitive-feature (value)

(cond

((equal value elicits)

'((primitive elicits)))

((equal value informs)

'((primitive informs)))))

Other subfeature constructors are implemented in the same manner.

4. After all subfeatures are constructed and appended to the overall feature set, the

entire feature set is ready for a sentence generation.

7 Conclusion

The idea of lexicalization is not well-studied in natural language processing. Part of

the reason is that a fine-grained lexicalization is related to something beyond sentence

interpretation. The intentions of speakers and the understanding of listeners are the

major factors dominate the evolving discourse and lexical usage.

Many natural language research groups have found that a certain number of natural

language generation issues are beyond the consideration of discourse planning and

surface generation, but they are nonetheless important in building high-quality text

137

generation systems. A certain level of cognitive related issues has to be taken into

consideration. In this research, I focus on the task of lexical refinement to produce a

more detailed dialogue specification for the surface sentence generator to generate

more coherent and natural sounding sentences. This is a critical problem and I have

taken the first step toward it.

Acknowledgement

This work was partially supported by the Cognitive Science Program, Office of Naval

Research under Grant 00014-00-1-0660 to Stanford University as well as Grants

N00014-94-1-0338 and N00014-02-1-0442 to Illinois Institute of Technology. The

content does not reflect the position of policy of the government and no official

endorsement should be inferred. Personal communications with Professor Martha

Evens at Illinois Institute of Technology has been of great assistance to this research.

References

1. Allen, J., and Core, M. 1997. Draft of DAMSL: Dialog Act Markup in Several Layers, The

Multiparty Discourse Group at the Discourse Research Initiative (DRI) Meeting

, Schloss

Dagstuhl, Not Paged.

2. Brennan, S.E., and Clark, H.H. 1996. Conceptual Pacts and Lexical Choice in Conversation,

Journal of Experimental Psychology: Learning, Memory, and Cognition, Volume 22, Pages

1482 - 1493.

3. Grice, H.P. 1969, Utterer’s Meaning and Intentions.

Journal of Philosophical Review,

Volume 68, Number 2, Pages 147 - 177.

4. Grosz, B. 1997. Discourse and Dialogue: Overview, In Varile, G., Zampolli, A., Cole, R.,

Mariani, J., Uszkoreit, H., Zaenen, A., and Zue, V. (editors),

Survey of the State of the Art in

Human Language Technology

, Chapter 6, Cambridge, UK: Cambridge University Press.

5. Kim, J., 1998.

The SGML Markup Manual for CIRCSIM-Tutor. Technical Report, Computer

Science Department, Illinois Institute of Technology, Chicago, IL.

6. McKeown, K. 1985.

Text Generation: Using Discourse Strategies and Focus Constraints to

Generate Natural Language Text

, Cambridge, UK: Cambridge University Press.

7. McRoy, S, Ali, S., Restificar, A., and Channarukul, S. 1999. Building Intelligent Dialogue

Systems,

Journal of Intelligence, Volume 10, Number 1, Pages 14 - 23.

8. Moser, M., and Moore, J.D. 1995. Investigating Cue Selection and Placement in Tutorial

Discourse,

Proceedings of the 33rd Annual Meeting of the Association for Computational

Linguistics

, Cambridge, MA, Pages 130 - 135.

9. Nakatani, C., Grosz, B., Ahn, D., and Hirschberg, J. 1995.

Instructions for Annotating

Discourses

, Technical Report Number TR-21-95, Center for Research in Computing

Technology, Harvard University, Cambridge, MA.

10. Nakatani, C., and Traum, D. 1998. Draft: Discourse Structure Coding Manual,

The 3rd

Discourse Resource Initiative (DRI) Meeting, Chiba, Japan, Not Paged.

11. Passonneau, R. 1994.

Protocol for Coding Discourse Referential Noun Phrases and Their

Antecedents

, Revised November 1994, Technical Report, CARD Project, Department of

Computer Science, Columbia University, New York City, NY.

12. Reichman R. 1985.

Getting Computers to Talk like You and Me, Chapter 5, Cambridge, MA:

The MIT Press.

138

13. Reiter, E. and Dale, R. 2000. Building Natural Language Generation System: Studies in

Natural Language Processing

, Chapter 3, Cambridge, UK: Cambridge University Press.

14. Schiffrin, D. 1987.

Discourse Markers, Cambridge, UK: Cambridge University Press.

15. Yang, F., Kim, J., Glass, M., & Evens, M. 2000. Lexical usage in the tutoring schema of

CIRCSIM-Tutor: Analysis of variable references and discourse markers,

Proceedings of the

5th International Conference on Human Interaction with Complex System

, Urbana, IL,

Pages 27 - 31.

Appendix A Lexical Rules

Based on the analysis of lexical phenomena in the tutoring schemata, I have

developed lexical rules for

polite locutions, variable references, discourse markers,

and

acknowledgment choices. These rules along with real life tutoring examples

marked with SGML tags are listed and discussed in following sections.

A.1 Lexical Rules for Polite Locutions

Rule 1: Within the first topic of

T-does-neural-DLR, the tutor uses the locutions

Can you tell me or Do you know to bring up a question politely.

Example:

<T-does-neural-DLR>

<T-tutors-mechanism>

K10-tu-29-4: Can you tell me how TPR is controlled?

...

</T-tutors-mechanism>

...

</T-does-neural-DLR>

A.2 Lexical Rules for Variable Descriptions

Rule 1: Use abbreviated variable names

Case 1: Within the topic T-introduces-variable, the tutor uses the abbreviated name

to introduce a new variable.

Example:

<T-introduces-variable>

K11-tu-41-1: You only have TPR left.

</T-introduces-variable>

Case 2: Within the topic immediately following T-introduces-variable, the tutor

keeps using the abbreviated name of the variable to maintain the same

discourse focus.

Example:

<T-introduces-variable>

K11-tu-41-1: You only have TPR left.

</T-introduces-variable>

<T-tutors-variable>

139

<T-does-neural-DLR>

<T-tutors-mechanism>

K11-tu-49-3: How is TPR controlled?

...

</T-tutors-mechanism>

</T-does-neural-DLR>

</T-tutors-variable>

Case 3: Within the last topic of T-tutors-variable, the tutor uses the abbreviated

name of the variable to end digressions and bring back the discourse focus.

Example:

<T-tutors-variable>

...

<T-does-neural-DLR>

K10-tu-29-4: Can you tell me how TPR is controlled?

...

K10-tu-31-2: And the predictions that you are making are for

the period before any neural changes take place.

<T-tutors-value>

K10-tu-31-3: So what about TPR?

...

</T-tutors-value>

</T-does-neural-DLR>

</T-tutors-variable>

Rule 2: Use pronominal descriptions

Case 1: Within the topic immediately following T-introduces-variable, the tutor uses

it to refer to the variable and maintain the same discourse focus.

Example:

<T-introduces-variable>

K12-tu-31-1: Now how about TPR?

</T-introduces-variable>

...

<T-tutors-variable>

<T-does-neural-DLR>

<T-tutors-mechanism>

K12-tu-33-1: By what mechanism will it increase?

...

</T-tutors-mechanism>

</T-does-neural-DLR>

</T-tutors-variable>

Case 2: Within the topic immediately following T-introduces-variable, the tutor uses

this to refer to a proposition and maintain the same discourse focus.

Example:

<T-tutors-variable>

...

<T-explores-anomaly>

<T-presents-anomaly>

K26-tu-76-2: So, co decreases even though sv increases.

</T-presents-anomaly>

<T-tutors-anomaly>

K26-tu-76-3: How can you explain this?

140

</T-tutors-anomaly>

</T-explores-anomaly>

</T-tutors-variable>

Rule 3: Use definite descriptions

Case 1: Within the topic of T-introduces-variable, the tutor uses the last one or this

issue to introduce the variable.

Example:

K10-tu-29-2: Let's take a look at some of your predictions.

<T-introduces-variable>

K10-tu-29-3: Take the last one first.

</T-introduces-variable>

Case 2: Within the topic immediately following T-introduces-variable, the tutor uses

that prediction to refer to both the variable and its change and maintain the

same discourse focus.

Example:

<T-introduces-variable>

K48-tu-44-3: you predicted that TPR would increase.

</T-introduces-variable>

...

<T-tutors-variable>

<T-does-neural-DLR>

<T-tutors-mechanism>

K48-tu-44-4: Can you explain how you arrived at that

prediction?

...

</T-tutors-mechanism>

</T-does-neural-DLR>

</T-tutors-variable>

Case 3: Within the last topic of T-tutors-variable, the tutor uses your prediction to

end digressions and bring back the discourse focus.

Example:

<T-tutors-variable>

<T-does-neural-DLR>

K48-tu-44-4: Can you explain how you arrived at that

prediction?

...

K48-tu-48-2: and during DR what changes in ANS activity occur?

...

<T-tutors-value>

K48-tu-50-1: So do you want to change your prediction?

</T-tutors-value>

</T-does-neural-DLR>

</T-tutors-variable>

A.3 Lexical Rules for Discourse Markers

Rule 1: Use so and now

141

Case 1: so and now are used in T-introduces-variable to initiate a discourse focus.

This is similar to behavior observed by Schiffrin [14].

Example:

<T-introduces-variable>

K11-tu-53-2: So let me ask you, are there any other of these

variables that are primarily under neural control?

</T-introduces-variable>

Case 2: so and now are used to conclude T-tutors-variable. This is similar to the idea

of marking results discussed by Schiffrin [14].

Example:

<T-tutors-variable>

<T-does-neural-DLR>

...

<T-tutors-value>

K10-tu-31-3: So what about TPR?

...

</T-tutors-value>

</T-does-neural-DLR>

</T-tutors-variable>

Rule 2: Use first in T-introduces-variable to introduce the first topic of the first

variable being tutored.

Example:

<T-introduces-variable>

K13-tu-37-3: First, what parameter determines the value of

rap?

</T-introduces-variable>

Rule 3: Use but in T-presents-contradiction to contrast two ideas.

Example:

<T-shows-contradiction>

<T-presents-contradiction>

K10-tu-41-2: You predicted that it would go up.

...

K10-tu-43-1: But remember that we’re dealing with the period

before there can be any neural changes.

</T-presents-contradiction>

</T-shows-contradiction>

Rule 4: Use and to initiate a semantically continuous topic.

Example:

<T-does-neural-DLR>

<T-tutors-mechanism>

K10-tu-29-4: Can you tell me how TPR is controlled?

...

</T-tutors-mechanism>

<T-tutors-DR-info>

K10-tu-31-2: And the predictions that you are making are for

the period before any neural changes take place.

</T-tutors-DR-info>

...

<T-does-neural-DLR>

142

Rule 5: Use therefore to summarize T-tutors-via-deeper-concepts.

Example:

<T-tutors-via-deeper-concepts>

<T-tutors-determinant>

K27-tu-52-1: If I have a single blood vessel, what parameter

most strongly determines its resistance to flow?

...

<T-moves-to-previous-concepts>

<T-tutors-determinant>

K27-tu-54-1: And physiologically, what determines the

diameter of the blood vessels?

</T-tutors-determinant>

</T-moves-to-previous-concepts>

</T-tutors-determinant>

<T-tutors-determinant>

K27-tu-56-2: Therefore, what determines TPR?

</T-tutors-determinant>

</T-tutors-via-deeper-concepts>

A.4 Lexica Rules for Acknowledgments

Rule 1: Use a negative acknowledgment such as no or not quite to reject the

student’s first wrong answer.

Example:

K12-tu-31-1: Now how about TPR?

<T-elicits>

K12-tu-33-1: By what mechanism will it increase?

<S-ans catg=incorrect>

K12-st-34-1: If you increase pressure will you momentarily

increase resistance

</S-ans>

<T-ack type=negative>

K12-tu-35-1: No.

</T-ack>

</T-elicits>

Rule 2: Use a partial acknowledgment, such as partly correct, to partially accept the

student’s answer.

Example:

<T-elicits>

K47-tu-56-5: Can you tell me what you think that IS means?

<S-ans catg=near-miss>

K47-st-57-1: the contractility of the heart caused by preload

and sympathetic stimulation

</S-ans >

<T-ack type= partially-correct >

K47-tu-58-1: Partly correct.

</T-ack >

</T-elicits>

Rule 3: Use of positive acknowledgments

143

Case 1: Use yes or right to accept the student’s first correct answer.

Example:

K10-tu-29-2: Let's take a look at some of your predictions.

K10-tu-29-3: Take the last one first.

<T-elicits>

K10-tu-29-4: Can you tell me how TPR is controlled?

<S-ans catg=correct>

K10-st-30-1: Autonomic nervous system

</S-ans>

<T-ack type=positive>

K10-tu-31-1: Yes.

</T-ack>

</T-elicits>

Case 2: Use a strong positive acknowledgment, such as good, very good, absolutely,

exactly, or great to accept the student’s final correct answer, especially when

the student had some difficulty in reaching this goal.

Example:

<T-elicits>

K27-tu-72-2: How is this possible?

<S-ans catg=correct>

K27-st-73-1: Hr is down more than sv is up

</S-ans>

<T-ack type=positive>

K27-tu-74-1: Very good.

</T-ack>

</T-elicits>

Rule 4: Acknowledgment is omitted in some special situations, such as when the

tutor is identifying the student’s problem, or the student has a near miss

answer.

Case 1: the tutor tries to identify the student’s problem without giving any

acknowledgment.

Example:

<T-diagnoses-errors>

<T-identifies-problem>

<T-elicits>

K27-tu-50-2: Why do you think that TPR will decrease?

<S-ans catg=incorrect>

K27-st-51-1: Since HR decreased, CO will decrease and the

direct response would be decreased TPR.

</S-ans>

</T-elicits>

</T-identifies-problem>

</T-diagnoses-errors>

K27-tu-52-1: If I have a single blood vessel, what parameter

most strongly determines its resistance to flow?

(Acknowledgment omitted)

Case 2: The tutor does not give any acknowledgment when the student gives a near-

miss answer, but tries other methods to guide the student toward the correct

answer.

144

Example:

<T-tutors-via-determinants>

<T-tutors-determinant>

<T-elicits>

K25-tu-48-3: What parameter determines rap?

<S-ans catg=near-miss>

K25-st-49-1: Central venous pressure.

</S-ans>

</T-elicits>

<T-moves-toward-PT method-type=inner>

<T-tutors-determinant>

<T-elicits>

K25-tu-50-1: And what determines cvp?

(Acknowledgment omitted)

<S-ans catg=correct>

K25-st-51-1: Blood volume and "compliance" of the Venous side

of the circ.

</S-ans>

<T-ack type=positive>

K25-tu-52-1: Right.

</T-ack>

</T-elicits>

<T-moves-toward-PT>

</T-tutors-determinant>

</T- tutors-via-determinants>

145