Conversational analysis studies

[A note: I really appreciate the feedback I’ve gotten on these conversation modeling posts, both in comments and off-blog. My goal here has been to take a first stab at explaining the model of Alabaster, since I’m going to have to write up a bunch of documentation for it; and to get some responses on what people find especially confusing and what features they’d especially like. So it’s been helpful for me. To those of you who don’t find them very interesting, thanks for your patience, and I’ll be back to a more typical blend of content soon.]

Recently I had a chance to reread parts of a book I haven’t looked at in some years, Stephen Levinson’s textbook Pragmatics. Levinson talks about a number of issues to do with how meaning is understood in context and how conversational exchanges are negotiated. When I wrote Galatea, I was guided a bit by Grice’s maxims of conversational implicature picked up from a linguistics class in college; the last time I looked at Levinson’s book it was to remind myself about Grice.

This time I was more concerned with the chapter he devotes to conversational analysis: basically taking a lot of transcripts of ordinary conversations and working out, as far as possible, the principles by which people work out how to take turns, what strategies are used to mark beginnings and endings, why pauses and overlaps occur, etc.

I found reading this heartening in some ways. Many of the things Levinson talks about — the call and response structure of conversation (where conversations can often be analyzed as pairs of utterances), the importance of relevance in what people are allowed to say next, etc. — are familiar and are already built into many IF models of conversation.

So I was most interested in the things that Levinson talks about that could be modeled in IF but generally aren’t, or that raise interesting implementation questions. A few of these feel like common sense, but having them formally explained made me think about them in a new way all the same. The ones that leapt out at me:

— interruption to correct. Levinson observes that in general interruption of another person’s speech is rude, but that it is acceptable when the first speaker has said something that needs immediate correction or clarification in order for the communication to be meaningful. E.g.

M: “So when we’re in Avignon, we can–”
T: “You mean Nice?”
M: “Yes, sorry, Nice. We can get the TGV…”

(Note I don’t have the book in front of me any more, so the examples are not ones from the text but reconstructions based on my notes and memory.)

I could imagine a mechanism like this as a way to offer a disambiguation question via the NPC rather than via the parser. Might be a little odd, but in some circumstances I could imagine an exchange like

>TELL JANE ABOUT AVIGNON
“Speaking of Avignon–”

She looks confused. “Do you mean Avignon proper, or Villeneuve-lès-Avignon?”

>VILLENEUVE
“Sorry, Villeneuve. Anyway…”

— nesting of questions. Example:

S: How many tubes of paint would you like?
C: What’s the price on those?
S: $4.50.
C: Three, then.

This is a little tricky from an IF modeling perspective because it requires a concept of nested context: “Three, then” only has meaning in answer to the original question, so we have to remember that that question was pending even though it was interrupted. My quip model allows for questions to be restrictive (answer required immediately) or unrestrictive (player can change the subject as he likes without answering), but it has no built-in mechanism for this kind of semi-restrictive thing where the player can ask some related questions and then come back to the main point. It would probably be possible to hack something together, but I’ve never tried and I’m not sure whether this is something that is worth systematizing. As Levinson points out, questions can be nested multiple times and the contextual stacking can get rather deep.

— interpretation of silence in indicating dispreferred answers and/or ignorance.

This takes a little explaining, but Levinson argues that in many conversation pairs, one kind of answer is preferred and the others dispreferred. This doesn’t necessarily refer to what the people conversing want (despite the term “preferred” which might suggest personal feelings) but to what is expected. “Dispreferred” outcomes are often prefaced by pauses and words such as “well” or “actually”, and may get more elaboration than preferred outcomes. For instance:

S: Could I come by and see you tomorrow around 10?
C: Sure. [Preferred outcome; no pause, no elaboration.]

S: Could I come by and see you tomorrow around 10?
C: (pause) Well — I’ve got the washer repairman coming tomorrow and I’m not sure when he’s going to arrive and it might be awkward. [Dispreferred outcome: pause, introductory word, elaboration.]

Because of that pause, Levinson records that in many cases people are able to anticipate the dispreferred outcome, so you also get things like this:

S: Could I come by and see you tomorrow around 10?
C: (pause)
S: Or Thursday, if that would be better. [anticipates refusal, offers alternative]

It’s easiest to imagine making use of this in a real-time system, but it could still be useful as a way of reacting when the player types Z or does some other waiting, non-conversational action. Mostly I’ve had characters insistently repeat the question (and of course Levinson acknowledges that that also happens), but there might be times when it would be more realistic to have them interpret the silence in context and move the conversation in a new direction.

Along the same lines:

Mother: What time is it on the clock?
Child: (pause)
Mother: What number is the short hand pointing at?
Child: Two.

I can’t think of any interactions off the top of my head where an NPC is trying to tutor the PC like this, but it might be an interesting kind of scene to have now and then.

— the tendency of people to try to encourage offers rather than make requests, as in

A: Are you going to Peter’s birthday party?
B: Yeah, do you want a lift?

Levinson argues that this kind of exchange is generally liked better than

A: Can I have a lift to Peter’s birthday party?
B: Yes.

Or even

A: Are you going to Peter’s birthday party?
B: Yes.
A: Can I have a left?
B: Sure.

A is trying to avoid getting the dispreferred answer (“No, you can’t have a lift”) by imagining the likeliest objection (“I’m not going”) and getting it out of the way. B further helps out by anticipating the request and making an offer.

This one isn’t hard to model, with existing systems, but I thought it was interesting because it highlights a way in which IF conversation is often stilted by the overly rigorous correspondence between player command and protagonist speech. For instance, it would be very common to get a transcript like

You could ask B for a lift to Peter’s party.

>ASK B FOR A LIFT
“Could I have a lift to Peter’s party?”

“Yes.”

And there’s nothing absolutely wrong with that, but as Levinson points out it feels just a bit off from the way people typically negotiate these situations, especially with friends. On the other hand, there’s also a problem with

You could ask B if he’s going to Peter’s party.

>ASK B IF HE’S GOING
“Are you going to Peter’s party?”

“Sure — want a lift?”

…because here though the dialogue is more plausible, the option the game offers the player (“you could ask B if he’s going”) describes what the protagonist is literally about to say, it does not describe the intention and likely outcome of the act, which is not just to establish a fact but to make an arrangement. So I sometimes prefer

You could ask B for a lift to Peter’s party.

>ASK B FOR A LIFT
“Are you going to Peter’s party?”

“Sure — want a lift?”

Even though there’s a disparity between the player’s command and what comes out of the protagonist’s mouth, the question is an obvious preparation for the request, and I think most players would accept the fact that the request is never explicitly made because the NPC gets there first with an offer, not because the game somehow cheated or misinterpreted what the player wanted to do.

That negotiation between text and subtext may become even more important in cases where the characters belong to a very formal society or have conversational habits that the player might not be familiar with. I think it’s often better to offer the player an option that describes what the protagonist is trying to do, rather than exactly how the protagonist plans to say it, and then write the dialogue accordingly, unless the gameplay is partly about negotiating just this kind of conversational nuance. (See: Forever Always, Varicella.)

There was quite a lot else in the chapter: the special dynamics of telephone exchanges; noises made to hold on to one’s conversational turn (“uh”, in many English contexts) or yield it to another person (“hm” is apparently common for that); dynamics of asking permission to hold the conversational floor for a long time (“did I tell you the one about the Spanish cat?”); etc.

Some of these things would be hard, dull, or unnecessary to model in IF. Interactive fiction tends to obey some of the conventions of literature to the effect that stumbling, accidental interruptions, etc., are abstracted out of the dialogue. Moreover, the turn-based structure means that a lot of the markers of conversational exchange aren’t necessary, whereas I could imagine in a real time game letting the player say “Uh…” as a way of claiming the floor and gaining time to give a longer bit of speech.

Other of these effects can be included without requiring any explicit model, because for (say) the Spanish cat story we might tend to write a single quip that included all of the pieces: the initial question, the NPC’s consent, and then the story.

All the same, IF often has to find a middle way between emulating literature and attempting simulation of life, and for the simulation aspect, it doesn’t hurt to look at some organized data.

If people have run into other books besides Levinson’s that do a good job on topics in conversational analysis, I’d be interested to hear about them as well.

16 thoughts on “Conversational analysis studies”

  1. Relevance by Sperber and Wilson might be of interest, though it’s a little bit more of a competitor to Gricean theory than something that builds on Grice the way that Levinson chapter does. It also might not be that useful; my not-very-considered opinion was that they had a hammer and were trying to turn everything into a nail. (Basically they focus on relevance rather than the other Gricean maxims, per <a href="http://en.wikipedia.org/wiki/Relevance_theory&quot;
    wikipedia: “the human mind will instinctively react to an encoded message by considering information that it conceives to be relevant to the message. By ‘relevance’ it is meant whatever allows the most new information to be transmitted in that context on the basis of the least amount of effort required to convey it.”) On the other hand, that very characteristic might make it more useful for programming.

    Another thing that might be useful (something I’ve read about rather than read), is politeness theory as developed by Penelope Brown and Levinson. A lot of this seems like it might be useful for IF modeling.

    Oh, and if you want to remind yourself about Grice some more, his original article “Logic and Conversation” is super readable. At least by philosophers’ standards.

    I’ve really liked this series, and expect that if I ever get into writing IF in a serious way I’ll be looking at it a lot.

  2. Hi Em, I’ve been enjoying this series too, when I get the time & inclination to read something other than fluff. (Oh how the summer heat makes the brain go soft.)

    My WIP doesn’t “re-print” what the PC says after the command prompt, it just prints the NPC’s response. That puts a different spin on these ideas. Hm.

    Also, I have this list of various “elements of conversation” on my hard drive. I can’t remember where it came from, but I’m pretty sure it was a second-hand source, like Chris Crawford or some other non-linguistics person. Perhaps you’ll find it useful, as a checklist if nothing else? Here goes:

    setting –where you converse
    entrance / exit
    purpose
    topics
    formality
    who talks to whom
    turn-taking vs overlap
    cues
    appreciation
    use of humor
    how/if get to the point
    direct or indirect
    sequencing, order
    pace
    eye contact
    attitude, tone of voice
    silence
    length of each utterance, of conversation as a whole

    Also, thanks again for some code you gave me many moons ago in RAIF on how to couch a disambiguation question in an NPC’s speech. I never considered it in the context of the NPC interrupting the PC like you’ve done in this post, but again, I haven’t been re-printing the PC’s dialogue after the command prompt, either. Seems like that limits my options on how to present dialogue. Not that I’m likely to change it this late into its development, but, I wonder if other tidbits of conversation open up to me in their stead. Hm.

  3. I really appreciate the feedback I’ve gotten on these conversation modeling posts, both in comments and off-blog. My goal here has been to take a first stab at explaining the model of Alabaster, since I’m going to have to write up a bunch of documentation for it; and to get some responses on what people find especially confusing and what features they’d especially like. So it’s been helpful for me. To those of you who don’t find them very interesting, thanks for your patience, and I’ll be back to a more typical blend of content soon.

    I’ve found this a very interesting series of articles. A few random observations:

    (1) I think it would be very useful to gather all this up into a form readily accessible to users of your conversation when it’s finally released. I suppose the documentation for the extension is the obvious place to put it, though it might make the documentation rather lengthy. In any case it would be good to see this series of articles remain readily available and not just disappear into the hazy no-man’s-land of old blog posts.

    (2) I also think that much of what you’ve been talking about has application beyond the conversation system employed in Alabaster. In particular, I’ve been asking myself how far the ideas you’ve been discussing are supported by TADS 3’s conversation system. My initial impression is that many of them are, or could be with a bit of authorial tweaking, although your articles suggest that the existing conversational system in TADS 3 could be used with more nuance and sophistication than its design typically indicates (which is not intended as a criticism of the design, but, if anything, of authorial mindset).

    (3) The main problem, to which I don’t have a good answer, is that implementing conversations of any length with this degree of nuance and sophistication quickly becomes almost prohibitively labour-intensive. In TADS 3, for example, I often find myself taking an hour or more implementing a Conversation Node which the player may negotiate in ten seconds by typing YES, because I’m also trying to cover the cases in which the player tries to walk away, or issue an irrelevant response, or fiddles around with non-conversational actions (all of which TADS 3 caters for), and also provide a variety of responses to these situations so that the NPC doesn’t look too much like an automaton. I’m aware that the system you used in Alabaster tries to alleviate the problem to some extent with its conversation code generation tool, but this doesn’t remove the work of writing all the text for varied responses, and it quickly becomes a bit unwieldy for modelling the more complex situations (of the kinds you’ve been describing), where it seems to me that you’re fairly quickly driven back to coding by hand (or correcting program-created code by hand). On the one hand, I’m not sure that there’s any real substitution for hard work here, while on the other I think the amount of hard work that needs to be done is a serious limiting factor on what is likely to be attempted in IF. It’s one thing to have an NPC of the conversational complexity of Snow White in Alabaster when conversing with her is what the game is almost entirely about; it would be quite another matter to implement lots of NPCs to this depth in a game that also featured other kinds of interaction. I doubt that many of us would have the stamina to attempt it!

    (4) That said, I do think your series of articles is an important contribution to the art of implementing conversation in IF (and not just in the system you used for Alabaster). It does suggest a number of techniques we could all usefully try to employ, whatever system we’re using.

    (5) On the other hand, there do seem to me to be a number of problems in implementing IF conversation that are likely to remain intractable until we possess the technology to implement NPCs that could pass the Turing test – by which time we’d probably be writing software for holodecks rather than text-based IF! Of course I’m not saying anything here of which you’re not already well aware, but it may also be worth reflecting on the limitations of conversation in IF. In particular:

    (i) It quickly becomes very hard to allow much significant branching in the flow of a conversation, since the combinatorial explosion would quickly become overwhelming. In my own work, I often fake it by giving the appearance of conversational options that often end up doing much the same thing; otherwise it would be too easy to derail the narrative altogether.

    (ii) In human conversation, what is said can sometimes be less important than how it is said. Particularly in interpersonal interaction, intonation, facial expression and gesture may count as much as the actual words spoken. Shared cultural, social and personal members will affect how what is said is understood and responded to. It may make all the difference whether I respond to something with a joke, or my sounding sympathetic, or with a curt reply suggesting lack of interest, or an over-lengthy reply suggesting too much concern to push my own agenda. The rhetorical effectiveness and memorability of my utterances will greatly depend on my choice of words, phrasing, intonation, rhythm, use of structures and so forth. I can see no way that all these factors can be represented in a command prompt without making the command prompt impossibly cumbersome to use. To be sure we can allow the following as a command:

    >MAKE A JOKE

    And in some situations this may work reasonably well. But I suspect that seeing someone else’s canned witticism at that point will always seem a little artificial compared with the experience of exercising one’s own spur-of-the-moment wit; and so on for the other types of example.

    (iii) It may be a little easier to maintain the appearance of realistic conversation when the NPC has the conversational initiative. MAKE A JOKE is likely to work better, for example, as a specific response to something an NPC has just said than as a contextless conversational gambit by the PC.

    (6) These observations are not offered in a spirit of defeatism, but rather by way of suggesting that reflection on what IF cannot do may help focus the discussion on what it can. Can we think of any workarounds for these problems? Should we structure IF narrative so that it simply avoids the kind of situation that would require the kinds of conversational nuance that IF can’t model effectively and stick to situations it can cope with, or are there some ways of pushing the boundaries, or devising or using conventions that help disguise the limitations.

    Some of these things would be hard, dull, or unnecessary to model in IF. Interactive fiction tends to obey some of the conventions of literature to the effect that stumbling, accidental interruptions, etc., are abstracted out of the dialogue. Moreover, the turn-based structure means that a lot of the markers of conversational exchange aren’t necessary, whereas I could imagine in a real time game letting the player say “Uh…” as a way of claiming the floor and gaining time to give a longer bit of speech.

    [snip]

    All the same, IF often has to find a middle way between emulating literature and attempting simulation of life, and for the simulation aspect, it doesn’t hurt to look at some organized data.

    What this suggests to me is that we need to reflect not just on conversational realism (which is partly unattainable for the reasons I’ve indicated) but on the artistic representation of realism. Reducing speech to writing already strips out a great deal of what makes face-to-face oral communication what it is: gesture, expression, intonation, emphasis, body-language and all the rest. The written representation of speech can compensate for this to some degree by more careful word choice in what is said than spur-of-the-moment oral communication is likely to contain, and by some verbal and typographical indications of expression and stress, though these might soon become tedious if overworked.

    This can be alleviated in situations of secondary orality (texts designed for oral performance, such as plays), since the actors/narrators can supply much of these visual cues in their own performance. Even so, effective drama is not simply a transcription of everyday speech; it employs various conventions to focus and dramatize that speech, conventions that audiences come to accept as passable imitations of reality for the purpose (I daresay that must have been equally true of Athenian drama as of any of its modern descendents).

    The need to devise suitable conventions for a medium designed primarily for the eye rather than the ear (modern novels or IF) is probably even more. People do not in fact speak in visual typescript, but we have become so used to accepting typescript as a representation of oral exchange that we cease to notice the artificiality of it. The problem, then, perhaps needs to be rephrased as not so much how to make IF conversation realistic, but how to make it appear realistic, given the reading expectations of its target audience. What conventions can we persuade our audiences to accept, given the complication the command prompt creates in comparison to the print novel? What kind of tricks can we employ to distract attention from the artificiality of IF conversational exchanges and create the illusion of life-likeness? What kinds of things to we need to avoid in order not to shatter the illusion?

    If people have run into other books besides Levinson’s that do a good job on topics in conversational analysis, I’d be interested to hear about them as well.

    I don’t know the literature in this area at all. I am currently (in my academic work) researching a tangentially related area, that of oral tradition and the interface between the oral and written media (particularly in antiquity), and it’s that reading that has to some extent informed some of the foregoing comments. If you think the oral/written media difference might be an interesting area to explore, then you could do worse than start with Walter Ong’s Orality and Literacy, though it’s not at all directly relevant to the question of conversational realism, and you don’t need to read Ong to realize that viewing a set of words displayed on a computer screen is already at some remove from the experience of conducting an oral/aural conversation.

    1. Thanks for the long reply, Eric — many good points here.

      It will always be a lot of work to implement conversation deeply. I think it’s possible for the generation tool to do more of the lifting than the contributors to Alabaster were able to see, though: my own experience is that the conversation code generation tool works more efficiently and completely for an author who is able to do all the recompilation him/herself; the amount of code tweaking I have to do is relatively slight, since I’m building branches, compiling them in, adding to them, etc., in fairly rapid succession. (Obviously I realize that it may be harder for people who are less familiar with it, so we’ll see how this goes.) But I’m able to draft out complete conversations significantly faster and with less boredom than in the past; and I suspect that the latter point is also quite important because the sense of friction in implementing something tends to decrease the author’s enthusiasm for the project. At least, it always bored me senseless having to do the clerical work of remembering synonyms, coming up with unique names for each quip object, etc., and while I realize that the requirements are different for different conversation systems, most of them involve some sort of tedious clerical work or other. My hope is to lift that to a large degree and let the author focus on tweaking just the things that are actually interesting to tweak.

      Second, the task of conversation design changes a lot when the shape of the game is different; specifically, when it’s supposed to be a series of well-defined scenes instead of a single large one. Alabaster is hard because it’s so open-ended, but I have another WIP where each scene is about something fairly focused and contributes to a more firmly-shaped overall narrative — and this makes for a much more manageable writing task.

      As to not passing the Turing test — well, no, naturally that won’t happen any time soon, though it’s hard to know what the Turing test even means in a context where the human’s interaction input is tightly restricted. What I’m hoping for is not to deceive the player, but to create some effect that startles the player out of his default assumption about game characters as mechanisms, and thus helps encourage him to view them as (at least) literary entities. This is actually a bit of a moving target, since what people are “used to” seeing game characters do changes over time.

      Still, I found “Make It Good” very effective in the same way because, though the conversation wasn’t always flawless, the characters were startling in their (apparent) ability to remember, conjecture, and react, and this challenged me to regard them differently and care more about their feelings.

      Finally, re. realism: I agree with you about the use of convention and the fact that IF must present a refined and limited subset of reality. But I do think that the subset it addresses will be different from the subset addressed in a non-interactive text, and in particular Levinson’s ideas about pauses, interruptions, changes of subject, etc., may offer some guidance in how to handle and redirect non-ideal/uncooperative player input such as waiting or walking away or saying something inappropriate during a conversation.

      I don’t think one can in the end win against a player who really wants to break the illusion; but there is something I think in Aaron Reed’s philosophy as expressed in Blue Lacuna, that every input that can possibly be made to fit in as a meaningful interaction should be accepted and interpreted in that light.

      Re. MAKE A JOKE, I don’t have a lot to offer here, but I think we could play more than we have with different levels of conversational command. I’ve been toying lately with the idea of the player gaining an “inventory” of facts, stories, or even lines of philosophical argument over the course of the game, that could then be deployed in various situations to persuade NPCs. I’ve never written such a game, but it has the appeal — to me — that it opens up the possibility of more tactical agency on the player’s part; perhaps even to the extent that the player might manufacture experiences or learn facts specifically in order to be able to deploy those in conversation at a later date. And, apropos of that, it seems as though many witty comments take the form of fitting a known thing into a surprising context; so that the player might even be able to generate jokes using the unexpected application of some item in his conversational inventory.

      It would be a pretty stylized effect, I know, but I’m keen to try it out sometime, when I work my way through a few other WIP plans.

      1. As to not passing the Turing test — well, no, naturally that won’t happen any time soon, though it’s hard to know what the Turing test even means in a context where the human’s interaction input is tightly restricted. What I’m hoping for is not to deceive the player, but to create some effect that startles the player out of his default assumption about game characters as mechanisms, and thus helps encourage him to view them as (at least) literary entities. This is actually a bit of a moving target, since what people are “used to” seeing game characters do changes over time.

        I don’t think I made it at all clear what I was getting at with my reference to the Turing test, which I phrased in a slightly tongue-in-cheek way. What I was getting at is mainly the point you made in your reply about the restrictions on the human player’s input. In IF there’s a limitation on what the player can usefully type, so that the player can’t realistically choose the wording or nuance of what the NPC says. We’re nowhere near an IF system that can process interchanges like:

        “What do you think about Robert?” Jill asks.

        >SAY: WELL, HE SEEMS A NICE ENOUGH CHAP, BUT I DON’T QUITE TRUST HIM SOMEHOW. HAVE YOU NEVER NOTICED THAT HE LOOKS YOU STRAIGHT IN THE EYE? BESIDES, HE’S JUST A BIT *TOO* CHARMING — I CAN’T HELP WONDERING WHAT LIES BENEATH ALL THAT CHARM.

        Although, of course, you can have the PC say that in response to the player typing:

        >T ROBERT

        But then what the PC says may be something quite other than what the player had in mind. The point being that IF conversational system don’t give the player the kind of control over conversation as we have in real life. The challenge is then how to work round this.

        It occurred to me after making my last post, however, that the limitations of the IF command line could in some sense be as much a strength as a weakness in implementing conversation, since it in some sense helps mitigate the oral/written difference. In the course of an IF conversation the player can read the NPC’s comment or question as many times as s/he likes before framing a reply, and take as long as s/he likes before typing it. If the player could also type the exact wording of the reply, this would be even further from the experience of face-to-face oral conversation, in which we can’t usually pause to work out the exact wording of our reply and revise it as much as we like before uttering it. Paradoxically, then, typing a laconic response at the command prompt may feel more like a real conversation than typing a verbose one. A brief response like YES, LIE, T ROBERT which takes a very short amount of time to type may actually be the best way to mimic the fact that in a real conversational situation we have to commit to a reply before we have time to work out to its precise wording (particularly in situations when the NPC’s question puts the PC on the spot).

        Incidentally, I wonder if this is part of the reason why a command prompt interface seems preferable (at least to me) to a conversational menu containing a list of explicitly spelt-out quips. In real conversation we never first think out a series of fully formed sentences and then choose which one to use, so the cognitive operations involved in a conversation menu always feel highly artificial. A brief command line response diminishes this sense of cognitive peculiarity, especially since the use of the command-line prompt throughout a work of IF means that players become sufficiently habituated to it not to feel its artificiality so strongly in conversation.

        But that’s a digression. To get back to the main point, the limitation of the command line can be (and often is) used to characterize the PC, of course. The player types T ROBERT, but the NPC then speaks in his or her own voice. The problem is then how much real agency this leaves the player, or, alterntatively, how far the player can be allowed (conversational) agency that potentially undermines the characterization of the PC.

        I have in mind situations where the NPC puts the PC on the spot. Suppose the MPC asks the PC an awkward question, e.g.

        “Has Jack being seeing Angela?” Jill asks.
        {Where Jack is Jill’s husband and your best friend, and you know Jack and Angela are having an affair).

        If the player is then presented with the options LIE, TELL THE TRUTH, PREVARICATE this may actually represent the conversational situation better than allowing the player to type a long freeform answer. In a face-to-face situation someone would just have to choose one of those strategies and launch into it, so it seems quite reasonable to allow the player to choose the top-level strategy while leaving the detailed wording to the PC.

        But then it occurs to me that the options offered might help characterize the PC. If the PC is being characterized as habitually honest, maybe a LIE option shouldn’t even be offered. Perhaps the options should then be BE RUTHLESSLY HONEST, BEND THE TRUTH A LITTLE, and CHANGE THE SUBJECT.

        Again, I’ve been wondering to what extent it may sometimes better to model the social and psychological constraints on a conversational encounter by the PC’s refusal to carry out certain conversational commands rather than by the NPC’s protests. In the example that follows, sometimes (1) may work better than (2):

        (1) “Has Jack been seeing Angela?” Jill asks.

        >T WEATHER
        It’s pointless to try changing the subject; it’d be tantamount to saying yes.

        >Z
        You can’t delay your answer any longer, or Jill will make her own conclusions.

        (2) “Has Jack been seeing Angela?” Jill asks.

        >T WEATHER
        “Don’t try to change the subject, answer my question!” Jill insists, “Has Jack been seeing Angela?”

        >Z
        Time passes.

        “I asked you a question,” Jill reminds you, “Has Jack been seeing Angela?”

        Of course, in some situations it may be entirely appropriate to allow the player to hesitate or attempt to change the subject (and perhaps then for the NPC to draw her own conclusions, as in some of your illustrations upthread). But what I’m wondering here is whether if we might sometimes make conversational exchanges seem more realistic if we don’t allow socially or psychologically unrealistic options. In a particular face-to-face conversational situation there may in practice be quite tight constraints on what one can say at any particular point in the conversation (not in terms of the exact wording, but in terms of the high-level options), and it may be that the current IF command prompt actually models that better than a seemingly more expressive conversational input might. But in that case it may also be appropriate to model the internal constraints of the situation (where an honest PC can’t bring himself to lie, for example, or it would look unrealistic to break certain social or personal conventions). It may sometimes be less work to supply a single internal refusal (to change the subject) than to write half a dozen NPC protests at the attempt to do so in order to make the NPC appear a bit less robotic.

        In sum, I’m not really proposing that Turing-test enabled NPCs would necessarily be a good thing. While at first sight the ability to type in a freeform response at the command prompt and have the NPC respond appropriately might look as if it would model conversation more realistically, the reverse might actually turn out to be the case. Besides, players tend to like being able to abbreviate their commands as much as possible (few wish to type EXAMINE THE BIG OLD RED LEATHER BALL where X BALL will do), so limited conversational command expressiveness fits that tendency. Thinking about handling it creatively may help us turn its apparent weakness into a strength.

      2. Ah, okay — it sounds as though we agree more than I originally realized.

        I do think restricting the player (rather than having the NPC react) is useful some of the time; my system tries to enable both of those things. I suspect that the right choice (between your options 1 and 2 here) will depend a lot on the circumstances of the individual game, how and how much the PC is characterized, etc. (As so often…)

      3. Ah, okay — it sounds as though we agree more than I originally realized.

        I do think restricting the player (rather than having the NPC react) is useful some of the time; my system tries to enable both of those things. I suspect that the right choice (between your options 1 and 2 here) will depend a lot on the circumstances of the individual game, how and how much the PC is characterized, etc. (As so often…)

        Absolutely!

        The other difficulty I not infrequently encounter is in formulating a suitably succint yet distinctive phrasing for the player to use in a non-standard conversational response. The shorter the command the player has to type, the more it mimics the snap decision-making of face-to-face conversation, and the less inclined players are to try some variant on your wording that doesn’t work. Commands like YES, NO, LIE are ideal for that. BE EVASIVE, PREVARICATE, SYMPATHIZE, LAUGH are just about as good, and only take a little more typing. BE ABSOLUTELY HONEST, BE MORE OR LESS HONEST, TRY TO AVOID TELLING THE TRUTH WHILE PRETENDING NOT TO become increasingly problematic. Sometimes succinct phrasings suggest themselves quite naturally. Other times there’s no substitute for ASK WHERE SHE HID THE LETTER, but that’s relatively unproblematic insofar as it looks like a reasonable extension of ASK ABOUT LETTER in form, even though players tend to problematize it by trying every wording except the one you prompt them with! More difficult is trying to find a neat way to express TELL HER THAT JACK WAS ONLY SEEING ANGELA FOR PHYSIOTHERAPY SESSIONS in situations where it seems appropriate to give the player this kind of level of control.

  4. Hi Eric,

    because I’m also trying to cover the cases in which the player tries to walk away, or issue an irrelevant response, or fiddles around with non-conversational actions (all of which TADS 3 caters for), and also provide a variety of responses to these situations

    Maybe what we could use is some standardized categories for irrelevant responses/behavior? Emily, Levinson’s book there goes into great detail over what’s appropriate or at least usual in conversation, but does it spend any time going over what’s inappropriate, or how interlocutors respond to conversationally-inappropriate behavior? Does it have categories for the various kinds of inappropriateness?

    For example, sometimes a change of subject is appropriate — as in the case of an interlocutor tactfully avoiding a sensitive subject — and inappropriate — when confronted about the murder he committed, the murderer talks about his shoes, completely avoiding the question or even the significance of it. (Alabaster has a subject-changing hook, but it isn’t picky about whether or not the change was appropriate.)

    Well, let’s say we have categories for all the kinds of inappropriateness (assuming such a thing could be conceived). Say, ignoring-someone-by-doing-something-else, inappropriate subject changes or non-sequitors, ignoring the usual give-and-take quality of conversation, etc. We could write a handful of responses to each of these categories once and for all. Then we can write the game never needing to deal again with the problem.

    I’m saying this because, as it is, it seems like we’re dealing with “simply walking away from the conversation” and such inappropriateness in an ad-hoc manner, so it has to be re-solved in every new conversation that occurs in our work. Hence the time-sink.

    We can already write scenery rules like, “Instead of examining scenery, say ‘You look at [the noun]. It looks just like any other [noun] you’ve ever seen.'” So we needn’t re-write a response for each individual, irrelevant prop.

    Well, can we also in a similar manner capture and deal with all manner of irrelevant behavior, like “Instead of slighting-the-listener-by-ignoring-the-to-and-fro-of-normal-conversation, say ‘[The current interlocutor] says, ‘Um, pardon me, but did you just [explanation of inappropriateness]? Am I just some servant for your use and abuse?'”

    I don’t know what said categories of conversation-breaking behavior would be, or how many, etc., so I was wondering if Levinson had a list-of-don’ts like he seems to have a list-of-do’s.

    -R

    1. Alas, I don’t think Levinson offers a categorization of What Not To Do in the sense you mean, though he recognizes many individual instances of undesirable behavior. (Also perhaps unsurprisingly he doesn’t have any examples of recorded real-life conversation involving an unidentified murderer… at least so far as we know :)

      But I suspect what we really want is a recognition of the kinds of What Not To Do that can happen in IF, which include: failing to respond when it’s clearly one’s conversational turn; doing something else and thereby showing indifference; changing the subject abruptly; leaving the conversation entirely before it’s over.

      I think those are the big ones, really, since many of the other possible faux pas (interrupting when it’s not appropriate to do so, leaving the wrong lengths of pause between interactions, telling long stories without allowing interruption and without having established that the other person is willing to listen) would require real-time interaction to express or a more fine-grained control over conversation format.

      1. Oh, also, it occurs to me:

        (Alabaster has a subject-changing hook, but it isn’t picky about whether or not the change was appropriate.)

        Well, let’s say we have categories for all the kinds of inappropriateness (assuming such a thing could be conceived).

        For the particular situation you describe, I could imagine letting the author label some quips and/or subjects as “potentially sensitive”, so that when the player changes the subject the NPC could acknowledge (aloud, with a gesture, or only by setting an internal flag) that the subject change was appropriate and indicated some sensitivity attaching to that particular issue. Indeed, the NPC might be able to use that information to start to form an idea of what the PC was and wasn’t willing to talk about. Which could be interesting.

  5. I’ve also really been enjoying this series, so thanks!

    I’m a big fan of disparities between what the player types and the PC says, because they can aid so much in characterisation — making it easier, and more fun, to play a PC with very different knowledge or personality to the player. To my mind at least, just reprinting the PC’s exact works makes for a far more fluid conversation. I played Galatea shortly after discovering the modern IF scene, and was delighted by snippets like

    … “It was only afterward that it became strange.”

    > a strange

    “Strange how?”

    Indeed, there’s a sort of very smooth IF experience, at the core of many of my favourite works, which seems to me to bypass the idea of natural language input altogether. By this stage you have learnt the crucial commands needed to deal with the particular game you’re playing, and can go straight from the impulse in your head to reading the game’s response to it without worrying about what you’re typing — which is in fact as highly stylised as any command line, whether it’s a high-speed N E … to navigate through a large but well-known area of scenery, or, as here, a conversation. And, to return to the point, spelt-out and flavoursome conversation from the PC really helps in creating this sort of immersion — apart from anything else because you can completely ignore the text on the screen that you’ve actually typed in.

    Benji

  6. If people have run into other books besides Levinson’s that do a good job on topics in conversational analysis, I’d be interested to hear about them as well.

    I still don’t know of any books in this area, but further to my musings on the artificiality of textualized speech, I’ve just come across an article that may be of interest: Ryan Bishop, ‘There’s Nothing Natural About Natural Conversation: A look at Dialogue in Fiction and Drama’ in Oral Tradition 6/1 (1991): 58-78. Bishop is basically arguing that dialogue in fiction is quite different in form and function from conversation in real life.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s