24.1.5 Conversational Implicature

We have seen that conversation is a kind of joint activity, in which speakers produce turns according to a systematic framework, and that the contributions made by these turns include a presentation phase of performing a kind of action, and an acceptance phase of grounding the previous actions of the interlocutor. So far we have only talked about what might be called the ‘infrastructure’ of conversation. But we have so far said nothing about the actual information that gets communicated from speaker to hear in dialogue.

While Ch. 17 showed how we can compute meanings from sentences, it turns out that in conversation, the meaning of a contribution is often quite a bit extended from the compositional meaning that might be assigned from the words alone. This is because inference plays a crucial role in conversation. The interpretation of an utterance relies on more than just the literal meaning of the sentences. Consider the client's response $ C_{2} $ from the sample conversation in Fig. 24.4, repeated here:

A $ _{1} $: And, what day in May did you want to travel?

C2: OK uh I need to be there for a meeting that's from the 12th to the 15th.

Notice that the client does not in fact answer the question. The client merely states that he has a meeting at a certain time. The semantics for this sentence produced by a semantic interpreter will simply mention this meeting. What is it that licenses the agent to infer that the client is mentioning this meeting so as to inform the agent of the travel dates?

Now consider another utterance from the sample conversation, this one by the agent:

A $ _{4} $: ... There's three non-stops today.

Now this statement would still be true if there were seven non-stops today, since if there are seven of something, there are by definition also three. But what the agent means here is that there are three and not more than three non-stops today. How is the client to infer that the agent means only three non-stops?

These two cases have something in common; in both cases the speaker seems to expect the hearer to draw certain inferences; in other words, the speaker is communicating more information than seems to be present in the uttered words. These kind of examples were pointed out by Grice (1975, 1978) as part of his theory of conversational implicature. Implicature means a particular class of licensed inferences. Grice proposed that what enables hearers to draw these inferences is that conversation is guided by a set of maxims, general heuristics which play a guiding role in the interpretation of conversational utterances. He proposed the following four maxims:

• Maxim of Quantity: Be exactly as informative as is required:

1. Make your contribution as informative as is required (for the current purposes of the exchange).

2. Do not make your contribution more informative than is required.

QUALITY

• Maxim of Quality: Try to make your contribution one that is true:

1. Do not say what you believe to be false.

2. Do not say that for which you lack adequate evidence.

RELEVANCE

MANNER

Maxim of Relevance: Be relevant.

Maxim of Manner: Be perspicuous:

1. Avoid obscurity of expression.

2. Avoid ambiguity.

3. Be brief (avoid unnecessary prolixity).

4. Be orderly.

It is the Maxim of Quantity (specifically Quantity 1) that allows the hearer to know that three non-stops did not mean seven non-stops. This is because the hearer assumes the speaker is following the maxims, and thus if the speaker meant seven non-stops she would have said seven non-stops (“as informative as is required”). The Maxim of Relevance is what allows the agent to know that the client wants to travel by the 12th. The agent assumes the client is following the maxims, and hence would only have mentioned the meeting if it was relevant at this point in the dialogue. The most natural inference that would make the meeting relevant is the inference that the client meant the agent to understand that his departure time was before the meeting time.

24.2 BASIC DIALOGUE SYSTEMS

We’ve now seen a bit about how human dialogue works, although as we’ll see, not every aspect of human-human conversation is modeled in human-machine conversation. Let’s therefore turn now to the spoken dialogue systems used in commercial applications today.

Fig. 24.5 shows a typical architecture for a dialogue system. It has six components. The speech recognition and understanding components extract meaning from the input, while the generation and TTS components map from meaning to speech. The dialogue manager controls the whole process, along with a task manager which has knowledge about the task domain (such as air travel). We'll go through each of these components in the next sections. Then we'll explore more sophisticated research systems in following sections.