10.2 Cohesion

Cohesion is the glue of discourse. Consider the following written piece of discourse:

(1)

George couldn't wait to get to France. However, he didn't stay there long.

"He" refers to "George," "there" refers to "France," and "however" establishes an adversative or contrastive relationship between the proposition that George was eager to go to France and that George's visit to France was short-lived. The example illustrates three types of cohesion devices. Cohesion refers to explicit linguistic devices, such as "however," "he," and "there," that help bind language into a coherent whole. Coherence, in turn, refers to the functional unity of a piece of discourse. Cohesion can be found not only in a written monolog, as in the example, but also in all dialogs.

Cohesion devices facilitate and reinforce comprehension of the whole, but they themselves are not responsible for creating meaning. Rather, they are natural cues that speakers and hearers use to signal and retrieve meanings that underlie utterances in context. The cohesion devices that are particularly relevant to prompt design are pronouns, discourse markers, and special pointer words such as "this" and "that."

10.2.1 Pronouns and Time Adverbs

A special feature of human-to-human language is that information in one unit of talk often presupposes information presented in a previous one, as exemplified in (2) through (4) (Schiffrin 1998).

(2)

I saw the cat the other day. It was still wandering around without a home.

(3)

I saw a robin the other day. It was the first one I saw this spring.

(4)

We moved here in 1982. We didn't even have jobs then.

In conversation, pronouns such as "it" and "one," as well as time adverbs such as "then," let the listener know that the referent at hand is the same as a referent uttered previously.^[2] In (2), (3), and (4), the reader or listener can easily make sense of these sentences by recovering the missing information from an earlier point in the discourse.

^[2] In writing, however, pronouns sometimes refer to something that comes up later in the text. For example, in the sentence "As soon as she got home, Pat called her lawyer," "she" and "Pat" are coreferential. In conversation, we would more likely say, "As soon as Pat got home, she called her lawyer."

Compare the stilted effect created by the repetition of the term "bookmark(s)" in the prompts listed in (5) with those in (6), rewritten with the pronoun "one."

(5)

SYSTEM:	You have five bookmarks. Here's the first bookmark. . . . Next bookmark. . . .That was the last bookmark.
CALLER:	Delete a bookmark.
SYSTEM:	Which bookmark would you like to delete?
. . .
SYSTEM:	Do you want to delete another bookmark?

(6)

SYSTEM:	You have five bookmarks. Here's the first one. . . . Next one. . . . That was the last one.
CALLER:	Delete a bookmark.
SYSTEM:	Which one would you like to delete?
. . .
SYSTEM:	Do you want to delete another one?

With the simple replacement of the noun "bookmark" with the pronoun "one," the prompts in (6) flow more naturally than those in (5). If VUI designers write prompts in context and listen to the dialog that emerges, they will be more likely to appreciate the ubiquity of pronouns in spoken language and to reap their stylistic benefits.

10.2.2 Discourse Markers

Discourse markers are defined as "sequentially dependent elements which bracket units of talk" (Schiffrin 1987). They connect utterances (or portions thereof) by succinctly relating what has just been said to what is about to be said through time, result, contrast, and other such notions. The following is a list of discourse markers, organized according to class of use (adapted from Quirk and Greenbaum 1973).

• Enumerative: first, second, third; for one thing, and for another thing, to begin with, for starters; in the first place, in the second place; one, two, three . . .; a, b, c . . .; next, then; finally, last, lastly; to conclude

• Reinforcing: also, furthermore, moreover, then, in addition, above all, what's more

• Equative: equally, likewise, similarly, in the same way

• Transitional: by the way, incidentally, now

• Summative: then, (all) in all, in conclusion, in sum, to sum up

• Apposition: namely, in other words, for example, for instance, that is, that is to say

• Result: consequently, hence, so, therefore, thus, as a result, somehow, for some reason or other

• Inferential: else, otherwise, then, in other words, in that case

• Reformulatory: better, rather, in other words

• Replacive: alternatively, rather, on the other hand

• Antithetic (or contrastive): instead (blend of antithetic with replacive), then, on the contrary, in contrast, by comparison, (on the one hand . . .) on the other hand

• Concessive: anyhow, anyway, besides (blend of reinforcing with concessive), else, however, nevertheless, still, though, yet, in any case, at any rate, in spite of that, after all, on the other hand, all the same, admittedly

• Temporal transition: meantime, meanwhile, in the meantime

• Attitudinal, commenting on truth: actually, in actuality, in (actual) fact, strictly speaking, nominally, officially, technically, theoretically

Occasionally customers resist the use of discourse markers in their VUIs on the grounds that they are perceived as "informal" or "slang." In and of themselves, however, discourse markers as a linguistic category are neither formal nor informal. The following markers, for example, are actually more formal than what we are used to hearing in everyday conversation: in addition, thus, therefore, hence, nevertheless, on the contrary, conversely, by comparison, in contrast, equally, all other things being equal, rather, and in conclusion. Markers that epitomize informal conversation are, for example, for starters, anyhow, and so.

There are significant benefits of discourse sensitivity, including the appropriate use of discourse markers, to an engineered dialog. Compare examples (7) and (8) (from Giangola 2000). Both sets of prompts aim to collect the same pieces of information to schedule an appointment in a personal information manager application. The prompts in (7) conspicuously lack discourse markers, whereas those in (8) were rewritten to project a natural, conversational style. Here we find both kinds of cohesion devices that have been discussed: pronouns and discourse markers (both in italics).

(7)

Please say the date.

Please say the start time.

Please say the duration.

Please say the subject.

(8)

First, tell me the date.

Next, I'll need the time it starts.

Thanks. <pause> Now, how long is it supposed to last?

Last of all, I just need a brief description . . .

Discourse markers add value to prompting in several ways. First, recall the central assumptions that discourse analysts make about human language. Language is a context-sensitive system that incorporates design principles specially adapted for communication. The redundant nature of discourse markers reinforces the functional relationship between two units of discourse. The discourse marker in the utterance, "Joe is a couch potato; his brother, on the other hand . . ." induces us to expect sharp contrast, giving us a pragmatic preview of what comes next. You, the reader, might have thought to complete the sentence with "is a triathlete," but you would not complete it with "hates to exercise." In (8), discourse markers such as first, next, and last of all orient listeners to their position in a sequence of questions.

Second, discourse markers suggest a humanlike awareness of how the dialog is progressing. Not only does the application recognize specific responses such as "Yes," "No," "Three p.m.," and "Operator," it seems to recognize how the user's contributions to the conversation relate to those of the system on more abstract cognitive and social levels. In these roles, discourse markers can be said to serve conversation management functions. This point is most robustly demonstrated later, in the discussion of markers such as oh, by the way, and actually, which are deceptively simple but functionally loaded.

As you can see from these examples, discourse markers serve a critical role in imparting naturalness to engineered dialogs, in addition to their essential role in projecting a familiar and favorable persona. In both robotic, context-insensitive (7) and natural, context-sensitive (8) dialogs, the information gap that initially exists between the application and the user is resolved only one piece at a time, and this has the potential to frustrate callers. Usability feedback reveals that the wording in (7) actually heightens the listener's awareness that information is being collected piecemeal. One respondent said that hearing one "Please say" after another was annoying, presumably owing to its unrelenting repetition. In general, listeners associate the dialog in (7) with such negative descriptors as "frustrating," "impatient," and "bored," whereas the dialog in (8) is associated with the positive descriptors "attentive" and "helpful."

Let's now turn to the use of specific discourse markers and see how they can enhance the natural flow of prompts in speech interfaces. These markers are now, by the way, oh, otherwise, and actually. We also look at conversational uses of okay and sorry. These are not typical discourse markers because they can stand alone but they function as discourse markers nonetheless.

Now

Now functions as a kind of paragraph-changer, a verbal indentation, as you saw earlier in "Thanks. Now, how long is it supposed to last?" (8). This discourse marker often introduces a new stage in a thought sequence, marking a progression of topics. For example, when a single prompt has more than one topic to cover, the discourse marker now subtly and gracefully ushers the caller's attention to the next topic in the sequence. In example (9), the caller has just completed a number of steps to enroll the caller's voiceprint.

(9)

SYSTEM:

That's all I need. We are done getting your voiceprint. Now, since you're new, would you like some help getting started?

Because now signals a change in topic in the discourse and does not refer to actual time, we can say things such as "Now, in the old days . . ."

By the Way

Just-in-time instruction, the instructional technique introduced in Chapter 9, lends itself to introduction with the discourse marker by the way, as exemplified in (10).

(10)

CALLER:	Call 555 465 2468.
SYSTEM:	Calling 555 465 2468. By the way, when you're done with this call, you don't have to hang up. <pause> Just say, "Come back."

As in human-to-human conversation, by the way introduces a tangential thought, often of parenthetical importance.

Oh

Such a small word, and yet volumes could be written about the discourse functions of oh. Although most dictionaries sweep oh under the carpet as a mere "hesitator" or "filler" expressing surprise, Schiffrin (1987) proposes that the central, prototypical function of the marker oh is to signal some sort of cognitive mismatch, communicating something along the lines of, "What I know or what I think does not match what you know or think."

In (11), the caller attempts to navigate to a site that is not considered secure. The interface informs that this will not be possible and tells why.

(11)

CALLER:	Go to <name of a Web site>.
SYSTEM:	Oh, I can't go to that site. It's not considered "secure."

In (12), the same interface informs callers that their personal profile cannot be bookmarked (referenced as a preferred voice site).

(12)

CALLER:	Make that a bookmark.
SYSTEM:	Oh, that can't be bookmarked . . .

This use of oh is related, by extension, to a self-editing use, as exemplified in (13) and (14), where the vertical bar represents the start of a new audio file.

(13)

CALLER:	Go to sports.
SYSTEM:	Sure, sports! | Oh, that's not available at the moment, but would you like to try something else instead?

(14)

CALLER:	Get my messages.
SYSTEM:	Okay, messages. | Oh, looks like you don't have any messages right now.

As a self-editing marker of cognitive mismatch, oh is invaluable in cases like these. The system response in (13) started with the implication that sports was available, and in (14) that messages were. The second message in each case contradicts the initial assumption. But introduced by oh, the otherwise incompatible piece of information is ushered in smoothly and gracefully, cast as a sudden change in the system's knowledge. In other words, the character played by the voice actor has only just now found out that sports or messages aren't available, so it can't be held at fault. The character was only trying to be attentive and cooperative initially but, for reasons beyond his or her control, was unable to deliver.

Be careful not to pepper the dialog with oh for reasons that are hard to justify. For example, the caller opts for "weather" at the main menu and the interface responds with the acknowledgment, "Oh, weather." Depending on how you say it, this phrase is open to a number of interpretations. One reading could be that the interface's persona has temporarily forgotten that weather existed as an option. Another reading might be that the persona is surprised that the caller has asked for weather, as opposed to some other option. Yet another possibility is that the persona is secretly wishing that the caller had not made this selection. (Along these lines, think about how the voice actor is supposed to say "Oh, weather." How would you coach the delivery?) These interpretations all resonate with Schiffrin's analysis of oh as a marker of cognitive mismatch, but this use is not appropriate here. In contrast, an appropriate context for "Oh, weather" would be to recover from a misrecognition, as in "Oh, weather," as opposed to what had been misunderstood as a request for traffic or horoscopes. In this context, the discourse marker appropriately signals cognitive mismatch: "What I heard before isn't what I'm now hearing." In any case, a more appropriate wording for the original context would be, "Sure, weather" or "Okay, weather," which communicate more neutral acknowledgments.

Because different discourse markers have different context-sensitive uses, it is critical that you choose carefully.

Actually

As with oh, there are a number of other discourse markers that suggest that an utterance may not be completely true or exactly right in reality for example, officially, technically, and theoretically. Other discourse markers, such as actually (15), really, in (actual) fact, and in reality, indicate what is considered to be the "real" truth.

(15)

CALLER:	Get my personalized weather report.
SYSTEM:	Actually, you had a list of cities for me to check on, but it's been deleted. To make a new list, go to the registration Web site. For now, just tell me what city and state you want the weather for. <pause> Or a city and a country.

One application we have worked on makes use of actually in certain error prompts, as in (16). The caller tries out the navigational command "Go forward," although he or she has not yet visited any voice sites.

(16)

CALLER:	Go forward.
SYSTEM:	Actually, I can't move ahead since we haven't gone anywhere yet. To hear a list of places you can go, just ask what's available.

Actually introduces the "real truth" and prepares the listener for some detail that he or she has probably not anticipated.

Otherwise

This is probably the most abused of discourse markers in speech interfaces. It frequently appears in error-recovery prompts, as exemplified in (17).

(17)

SYSTEM:

If you'd like to leave a message, say, "Yes"; otherwise, say, "No."

In real-life conversations, the information that follows otherwise is not strictly predictable from what came before, in spite of this marker's literal "if not" definition. The natural, informative use of otherwise is illustrated in (18) and (19).

(18)

If you're running late, call me. Otherwise, I'll come by at eight-thirty.

(19)

We'd better leave now. Otherwise, we won't get good seats.

It would be conversationally abnormal for someone to say, "If you're running late, call me. Otherwise, don't call me" or "We'd better leave now. Otherwise, we'll leave later." These "otherwise" alternatives are so predictable from their antecedents that the phrase as a whole is devoid of informational meaning. There is a conversational principle that can be roughly paraphrased as "Be informative." (This principle is a corollary to the Cooperative Principle discussed later in this chapter.) The use of otherwise in prompts such as (17) does not move the hearer into a sufficiently informative space, as would be expected in authentic discourse.

Prompts such as (17) may sound unnatural to users for another reason. "Yes" and "no" were natural responses to the state-initial prompt, which would have been something like, "Do you want to leave a message?" some two turns earlier in the conversation. What (17) is really saying is, "If you want to leave a message, answer as directly as possible my last question, because we're in the same recognition state and so the same recognition grammar is active." So the wording in (17) is designer-centered because it assumes familiarity with the concept of recognition states and active grammars.

In user-centered prompting, in contrast, otherwise would direct the caller into a more informative space, as in (20). Or you could reword the whole prompt and avoid its use altogether, as in (21).

(20)

If you want to leave a message, start talking after the beep. Otherwise, <pause> feel free to hang up.

(21)

Sorry, did you want to leave a message?

Yes or no, did you want to leave a message?

Did you want to leave a message? [parenthetically] (Just say, "Yes" or "No.")

The next two items are not discourse markers in the technical sense, but they serve a similar transitional function.

Okay

In one application we have designed, whenever the caller answers a question with no or a synonym, the ensuing prompt begins with okay, or some other acknowledgment, as exemplified in (22).

(22)

SYSTEM:	Did you want to review some more of your personal profile?
CALLER:	No.
SYSTEM:	Okay, what's next?

Okay acknowledges the caller's refusal and smoothly paves the way for the next dialog state. Notice that if we remove okay (23), not only is the result a stilted-sounding dialog, but also the dialog gives the vague impression that the system might not have heard what the caller just said.

(23)

SYSTEM:	Did you want to review some more of your personal profile?
CALLER:	No.
SYSTEM:	What's next?

In (22), okay could be replaced with all right, no problem, or not a problem, depending on the persona.

Sorry

A common criticism of the use of sorry in prompting is, "But, come on, now. How can a computer be sorry?" But sorry is useful, not so much for apology but rather as a signal that the forthcoming contribution will somehow fail to meet the listener's expectations. It can be employed as a mild bracing advisory. When your local video store clerk informs you that the movie you seek to rent week after week is yet again unavailable, there may be no hint of regret in the tone of "Sorry, it's out" or "Sorry, we don't have it."

In the recording studio, skilled voice actors can read "sorry" more as a transitional device than as a genuine expression of full-fledged, heartfelt regret. For example, in one application we have worked on, second-time recognition errors trigger the generic prefix-type prompt in (24) as well as a number of randomized paraphrases. The voice actor's delivery of "sorry" is professional and neutral-sounding, a transitional device that mirrors the conventionalized use of "sorry" in day-to-day public life.

(24)

CALLER:	My PIN is three six, um, no, it's two six four seven.
SYSTEM:	Sorry, I didn't catch that.

Avoid using discourse markers just to make the interface sound more like a real person. In our view, the primary purpose of using discourse markers in a VUI is to facilitate comprehension, thereby reducing cognitive load, and not to create the deception that the system is a human attendant. Discourse markers, after all, are a design feature of natural discourse that allows speakers to reinforce meaning relationships between utterances, in turn allowing listeners to decode them more easily. Prompt writers should take advantage of this "design feature" of natural discourse, but with care.