Learning Grice’s maxims as a machine so we can break them like humans
Asking the right question is not always a piece of cake, but answering appropriately can be even more tricky. Using words in a grammatically correct way is just the starting point of language use. In communication, speaker skills go far beyond linguistic competence. Hymes (1972) refers to this as communicative competence, the ability speakers have to use language in a specific context of use appropriately, taking into account what to say, when to say it, how to say it and with whom. In this regard, a set of unspoken cooperative rules can be of help in guiding the flow of conversation. These norms, elucidated by philosopher H.P. Grice, are part of the socalled ‘Cooperative Principle’.
Grice’s Cooperative Principle (Grice, 1975) serves as a fundamental precept for achieving cooperative and effective communication. The Cooperative Principle states that, in communication, people tend to assume that others are cooperating to contribute meaningfully to the conversation.
This principle is underpinned by four maxims, useful for both formulating questions and providing answers.
Maxim of quantity
As far as answering is concerned, the maxim of quantity suggests that responders should offer enough information to adequately address the question, avoiding both excessive verbosity and insufficient detail. This means providing sufficient context and relevant facts without overwhelming the recipient with unnecessary information.
In the following example, Sheldon Cooper from The Big Bang Theory gives too much information to his friend Penny by getting lost in unsolicited details and technicalities:
Penny: What’s the difference?
Sheldon: What’s the difference?!
Leonard: Here we go…
Sheldon: In the winter that seat is close enough to the radiator to remain warm, and yet not so close as to cause perspiration. In the summer it’s directly in the path of a cross breeze created by open windows there, and there. It faces the television at an angle that is neither direct, thus discouraging conversation, nor so far wide to create a parallax distortion
Maxim of quality
The maxim of quality underscores the importance of truthfulness in answers, urging responders to provide accurate and reliable information.
Sometimes, a well-intentioned lie can save a friend’s sanity, as in the case of Leonard Hofstadter. Although, as in the following case, Sheldon’s entire statement is invalidated by this.
Sheldon: I trust Penny will adhere to the Official California Restaurant Workers Solemn Oath of Ethics and Cleanliness.
Amy: I don’t believe there’s any such thing.
Sheldon (To Leonard): You lied to me?
Maxim of relation
The maxim of relation highlights the necessity of preserve relevance: responders should ensure that their answers directly address the query at hand.
In the next example, Sheldon tries to explain Leonard’s work to Penny, but instead of getting straight to the point, he starts off a little more widely:
Sheldon: It’s a warm summer evening, circa 600 BC. You’ve finished your shopping at the local market, or agora… and you look up at the night sky. There you notice some of the stars seem to move, so you name them planetes or wanderer.
Penny: (Raises hand)
Sheldon: Yes, Penny?
Penny: Um, does this have anything to do with Leonard’s work?
Sheldon: This is the beginning of a 2,600-year journey we’re going to take together from the ancient Greeks through Isaac Newton to Niels Bohr to Erwin Schrodinger to the Dutch researchers that Leonard is currently ripping off.
Maxim of manner
Finally, the maxim of manner pertains to ensuring clarity: answers must be conveyed in a manner that is easily understandable to the questioner.
Below, Sheldon does not demonstrate great communicative competence as he does not adapt to his interlocutor with whom he does not share the specialised common ground of physics. Consequently, he gets lost in a sea of obscure and technical words:
Sheldon: Now, remember, Newton realized that Aristotle was wrong and force was not necessary to maintain motion. So let’s plug in our 9.8 meters per second squared as A and we get force, Earth gravity, equals mass times 9.8 meters per second per second. So we can see that MA equals MG and what do we know from this?
Penny: Uh, we know that… Newton was a really smart cookie.
You can find some other fun examples of Sheldon’s communicative incompetence on YouTube!
Contrary to what we saw Sheldon usually does, by adhering to these maxims, communicators can usually facilitate question answering, fostering meaningful dialogue and comprehension between interlocutors.
Despite all that, interlocutors can sometimes intentionally violate these maxims in conversation. While Gricean Maxims serve as the bedrock of effective communication, they are not always adhered to, especially in nuanced social interactions. Nevertheless, we understand each other as we share a common ground. Intentional violation of these maxims can depend on various communicative purposes, such as irony, sarcasm, implicatures, and innuendo. As an example, while Leonard is capable of using false statement for sarcastic purposes, his friend Sheldon is entirely unable to understand sarcasm. It’s as if he adheres strictly to the maxim of quality in everyday conversation, wherein everything one says must be true, as follows:
Leodard: God Sheldon, do I have to hold up a sarcasm sign to my mouth every time I speak?
Sheldon: You have a sarcasm sign?
The violation of such maxims is also explained with the Relevance Theory by Wilson and Sperber (2006). According to this theory, relevance is not merely a normative principle but a cognitive mechanism guiding how individuals allocate attention and process information. It suggests that utterances are processed based on their potential for cognitive effects, with relevance being a crucial factor in determining interpretive efforts. This is important to explain how and why individuals interpret utterances in particular ways even if answers appear to be irrelevant in a Gricean way.
Differently from what happens in human-human communication, conversational agents, on the other hand, operate within programmed parameters aimed at optimizing task efficiency and user satisfaction, usually prioritizing adherence to Gricean Maxims to ensure clarity, accuracy, and relevance in responses. On the one hand, this strinct adherence may, however, sometimes lead to answers lacking the nuanced understanding or flexibility typical of human conversation. On the other hand, the possible violation is not usually intentionally driven by specific communicative purposes, but a result of a minor communicative competence in machines (Miehling et al., 2024).
Furthermore, according to Miehling et al. (2024), besides the abovementioned Gricean maxims, other two rules might as well be important in human-machine interactions: i.e., benevolence and transparency.
Benevolence deals with the moral responsibility of a response and has the following requirements:
(1) The response should avoid being insensitive, rude, or harmful.
(2) The response should refrain from engaging with or endorsing requests that are harmful or unethical.
Transparency outlines the following three requirements:
(1) The response should recognize the speaker’s knowledge boundaries, making clear any limitations in expertise, evidence, experience, or context.
(2) The response should recognize the speaker’s operational capabilities, highlighting the nature of actions that can or cannot be performed.
(3) The response should be forthright about the speaker’s willingness to engage with specific subjects or heed relevant advice.
To sum up, although quantifying effective human-like conversational AI interaction faces numerous inherent challenges, especially in answer generation, delving into the pragma-linguistic and cognitive mechanisms driving the production of relevant responses could be crucial in improving its quality. This suggests that a deeper understanding of these processes and how to break them is essential for advancing the capabilities of conversational AI systems.
References
(Grice, 1975) Grice, H. P. (1975). Logic and conversation. In Speech acts (pp. 41–58). Brill.
(Hymes, 1972) Hymes, D. (1972). On communicative competence. Sociolinguistics, 269293, 269–293.
(Miehling et al., 2024) Miehling, E., Nagireddy, M., Sattigeri, P., Daly, E. M., Piorkowski, D., & Richards, J. T. (2024). Language Models in Dialogue: Conversational Maxims for Human-AI Interactions. arXiv preprint arXiv:2403.15115.
(Wilson & Sperber, 2006) Wilson, D., & Sperber, D. (2006). Relevance theory. The handbook of pragmatics, 606–632.
