Overview:

This is a collection that is primarily about developing formalizations of communication. The essays are primarily by AI scholars as well as a few notable linguists and philosophers. The entire suite of essays though does presuppose the idea that goals and plans are intrinsic to cognition and communication. There are a number of logical formalizations of communication and intention, but each of these requires a propositional model to account for knowledge in the world. This requirement is a failing from my purposes, but the approach and methods of formalization seem like fertile ground.

What I am trying to get out of this reading is a way to model knowledge and communication that is embodied (at least to the agents), situated, relational, social, and performative.

Notes:

Philip Cohen, Jerry Morgan, and Martha Pollack: Introduction

On the very first page, the editors describe a great example of communication, that is very hard to model.

Let us begin with an example. With the Wednesday advertising supplement in hand, a supermarket patron approaches the butcher and asks “where are the chuck steaks you advertised for 88 cents per pound?” to which the butcher replies, “How many do you want?” (p. 1)

This example is difficult because there is a lot of things implicit in the communication, and while the literal meanings of the responses are barely connected, the response is perfectly natural and normal. The editors follow this up with a little analysis, but I disagree with their conclusions. They suggest that this is about communication of intentions and realizing what the other agent wants and acting to satisfy this need. While I agree that the butcher in the example must know at some level what the customer wants, I think that is a poor candidate for what is going on in his head.

I would say that this is better explained by sociological methods, especially the notion of scripts. This is an instance of the “shopkeeper” script. A person walking by is attributed the role of customer, and the butcher is the shopkeeper. This attribution of roles is enforced by the environment and situation, but also more importantly by ingrained cultural experience. Everyone knows what shopkeepers and customers do. Shopkeepers and customers know what to do, they do not need to analyze each others’ intentions, but simply perform their roles.

The authors outline six questions which are to be addressed by the rest of the book:

1. Meaning: What is meaning? Is there a notion of meaning that is appropriate for all expressions?
2. Composability: How are meanings of complex expressions composed of the meanings of their parts?
3. Action: How does speech perform actions beyond the mere act of saying?
4. Indirectness: How can a sentence convey more information than the literal meaning?
5. Discourse compositionality: How does the interpretation of discourse reveal more than the sum of the meanings of the individual sentences?
6. Communication: What is communication? How are linguistic and nonlinguistic communication related? How is communication related to meaning?

These questions are important considerations for any theory of communication and would be good to address for future endeavors.

Michael Bratman: What Is Intention?

The goal here is to define and understand intention. THis concept is something that ties together mind and action. The idea is that intention is a sort of intermediary buffer between thought and action. Before perfoming actions, one might form intentions to perform higher goals, and this would direct the performance of the actions.

A challenge comes with distance, and this is explained through the rhetoric of planning. The problem is explained with a nice example: “Suppose I intend today to drive over the Golden Gate bridge tomorrow. My intention today does not reach its ghostly hand over time and control my action tomorrow; that would be action at a distance. But my intention must somehow influence my later action; otherwise, why bother today to form an intention about tomorrow?” (p. 16) Bratman poses a trilemma of issues posed by intention at a distance:

1. Distant intentions are metaphysically objectionable, because they involve action at a distance.
2. Distant intentions are rationally objectionable, because they are irrevocable.
3. Distant intentions are a waste of time.

Bratman’s solution to this mess is the idea of planning. One has a hierarchy of intentions (or goals), and this hierarchy may be revised in context of changes in state and information. I think that a similar, but intrinsically different conclusion can be drawn. Instead of plans as mental constructions, intentions become part of an intrinsic state, essentially, intentions become roles. This is the sort of approach specified by Clancey et al in Cognition and Multi-Agent Interaction.

This idea gets hinted at some more later on: Bratman relates planning to action. ” (p. 19) I have a plan to A only if it is true that I plan to A.” This distinction is subtle, and it could also be used to relate planning to identity. For example: the plan, “I am planning to graduate” relates the identity, “I am a student.” This idea is touched on when Bratman connects intention to the idea of a “pro-attitude.” The function of the attitude can be examined as “I am someone who intents to A,” or, alternately, “I am conducting myself as to A.” This can be used to bridge intentions and roles.

Bratman outlines some issues that connect intentions to beliefs. With relation to each other and general knowledge, intentions must be consistent, coherent, constrained, admissible, and stable, etc. All of these hinge on the matter of beliefs and knowledge. The constraints and rules for determining intentions resemble the way that Soar considers operators.

At the conclusion of Bratman’s paper, he is trying to address the problem of the “package deal” when actions toward intended goals can have unintended effects. He uses an example that is rather disconcerting, though. The example is a wartime situation involving two bombers who may or may not bomb a school full of children (actually, it is whether to bomb a munitions factory that is next to a school full of children which would suffer collateral damage). This example is intended to illustrate the complexities of decision making, but wholly leaves out the charged emotional element. It assumes a totally rational process, and in this case, curiously, the rational agent decides to bomb the children. This is a key example of the danger of the emphasis on rational planning.

Philip Cohen and Hector Levesque: Persistence, Intention, and Commitment

The goal of this paper is about the “rational balance” of beliefs, intentions, and actions. The focus is on an AI controlled (or modeled) rational agent in a wold with other agents. While the formulation here has what I would consider to be an undue reliance on the dogma of rationality, the authors do emphasize a social element to action, and their formalization requires the existince of other agents to work.

The authors define a seven-point theory of intention. In this framework, an intention is essentially a persistent goal.

1. Intentions pose problems for an agent. The agent must determine a way to achieve them.
2. Intentions provide a “screen of admissibility” for adopting other intentions.
3. Agents “track” the success of their attmepts to achieve their intentions.
4. To intend p: The agent must believe p is possible.
5. To intend p: The agent does not believe he will not bring about p.
6. To intend p: Under certain conditions, the agent believes he will bring about p.
7. Agents need not intend all the expected side effects of their intentions.

Outlined here is a thorough and rigorous model of a logical formalization of action, beliefs, and intentions. The manner of expression is through propositional predicates. For example: (GOAL x p), (BEL x p), (HAPPENS a), etc. Over the course of the paper, more types of propositions and logical constructs are added. The formalization here is concerned with a precise logical modeling of the world. A problem, though, is that it can be used to describe agent models (the space of beliefs held by the agent), but all of these are absolute and literal. They are independent of perception or context.

One element in this analysis is the aim to formalize the rules of intentions described by the authors. This is successful, but it exposes the flaws and weaknesses in the original model. Here, issues such as procrastination and dedication are explained by complex structuring of beliefs and intentions with respect to time. Instead, I think that procrastination and dedication are based much more strongly in situation, personality, and emotion. The awkwardness and complexity of modeling some supposedly simple emergent qualities of behavior suggests that there is a failure to consider something important within the original model.

Martha Pollack: Plans as Complex Mental Attitudes

Pollack aims in this to present an alternative approach to planning, differing from a number of original models, specifically STRIPS and NOAH, which derive from Allen. Planning frameworks, which depend on graphs, heirarchy, and decomposition flounder because of several reasons. The one that Pollack is focused on is the human disconnect. It must be possible for a planning framework to handle invalid plans, ones that would be successful given the agent’s beliefs, but cannot occur because some of those beliefs might be false.

An element here is that, based on observations, plans may be inferred and analyzed. This inference is an important topic that could (and will) see more attention. Pollack is using an example of communication wherein one agents is trying to reach a friend: A: “I want to talk to Kathy, so I need to find out the phone number for St. Eligius.” S: “St. Eligius closed last month. Kathy was at Boston General, but she’s already been discharged. You can call her at home. Her number is 555-1238.” The nature of this communication is tricky, and I would argue that it is socially defined, but it would be hard to say that it is part of a formal script. I would agrue that it is based on experience, convention, and practice, but these are hard to formalize.

To deal with these situation, Pollack describes the idea of an explanatory plan, or an “eplan”, which gives a solution to the inference problem, but is vague. Other than by direct inquiry, how does one recognize a statement or question as an explanatory plan?

Henry Kautz: A Circumscriptive Theory of Plan Recognition

This essay investigates plan recognition. The approach here depends on keyhole recognition, which assumes more complex knowledge. The analysis here is done by structruing a logical formalized representation of observed events. Observation and analysis of plans must work according to several methods, among which are entailment (deductive) and closure (inductive). These rules define how plans may be infered from observed actions.

Kautz moves into a formal representation of events and plans. Plans are broken down into components, agents (participants), constraints (temporal and equality), and preconditions. What follows are then a few theorems on how models relate to events and fact determination. Communication is expressed through this logic as a collection of predicates: indirect requests, direct requests, inquiries, etcetera. Kautz then concludes with an algorithm for the recognition of plans and intentions.

This thought is really perplexing, because it subtly suggests that human intelligence consists of performing this algorithm.

An Aside

Much of what is going on here assumes a sense of objective truth. It assumes that minds and communication occur in pure, abstract (or purely representative) methods, ignoring cultural or social context and influence.

C. Raymond Perrault: An Application of Default Logic to Speech Act Theory

This section is on modeling and understanding speech acts. There is a good review of types of speech acts derived from Austin (1962), of the types of speech acts: locutionary, illocutionary, perlocutionary. It is interesting that the author acknowledges the performative element here. Illocutionary acts are rich with variable meanings, for instance, irony, sarcasm, lies, etc. The focus in this essay is on the application of logic to illocutionary acts, and understanding how they succeed or fail.

The logical structure used is derived from predicate elements : Kxp, Bxp, Gxp, for knowledge, belief, goals. The structure outlined enables complex formalizations of reflexive social knowledge and goals. An important element of communication here is the cycling of knowledge. For instance BxByBxByp. This analysis allows for an easy formalization of certain structures: “I want x to know that I want p.” However the recursive cycles are very awkward. What follows from here is leading to a non-monotonic logic where beliefs can change with changes in knowledge.

One of the issues and challenges with this model is that it bases all speech as propositional. If the performative element of speech is analyzed, it introduces complexity and challenges. Speech operates with scripts (that may be parameterized) and rituals. Speech may also have non propositional functions, like “annoy.”

Daniel Vanderveken: On the Unification of Speech Act Theory and Formal Semantics

This essay seeks to connect the theory of illocutionary acts (from Searle and Austin) to truth semantics (Frege and Tarski). The method is to separate speech acts into several types: declarative (make assertions), imperative (give directives), interrogative (ask questions/find knowledge), exclamatory (express state), optative (express wishes). I would argue that this is the right idea, breaking speech down into its functional elements, but speech acts in conversations really can have a lot more than single functions embedded into them.

Vanderveken notes that illocutionary acts are more than propositional content, but also contain other sorts of qualities. He explains these as being a set of qualifiers (forces) on the statements. These relate to the performative aspect of speech, but he does not go as far as suggesting that qualifiers might convey other information in of themselves.

An important thing to note in this discussion of general semantics is that it assumes literal meaning. This is especially interesting in the context of Lakoff, as well as Rumelhart, who both challenge the idea of literal meaning.

Philip Cohen and Hector Levesque: Rational Interaction as the Basis for Communication

The idea here is to examine illocutionary acts as instances of actions. This is aimed to be a reaction against some existing theories. The unity of speech and action has a strong psychological precedent, especially from Vygotsky. However, this is extended from the rational action framework in the earlier chapter, whcih fails to account for other contextual elements.

The authors describe a world model that uses the propositional approach (looking at cleaning floors and opening doors). The formulation of statement predicates is interesting for further use, but seems incomplete in light of some of the underlying problems, for instance, the dependence on literal statements. Abstracting to an action model can cupercede this. I think the speech actions should be more emotionally/socially oriented. For instance, support, praise, etc.

This theory and its complex logical formulation of a helpful agent is fascinating but hugely complex (p. 243). It seems like this is overboard in explaining the simple social logic of requests. A much better, simpler, strategy would be to work from another model of interaction based on sociology or cultural anthropology.

John Searle: Collective Intentions and Actions

Searle forms a complex analysis of intention, as relates to plans and sub-plans. Intention propagates to sub-tasks and gets messy. It seems like this fails when compared with embodied interaction and knowledge. (Recall Hubert Dreyfus on phenomenology). The collective analysis here is atomic. It does not relate to extended or proximal cogition, or social practice. The world being presented here is one in which every human agent is alienated from context and must decipher precisely every intention and goal before communication or action. While Searle is a critic of AI, he ascribes here the disembodied nature of reason held by AI scholars.

Herbert Clark and Deanna Wilkes-Gibbs: Referring as a Collaborative Process

This, the last chapter in the volume, makes an interesting and very deep analysis of conversation and reference that picks up on many ideas neglected by the other authors in the book.

The essay looks at conversation as collaborative, in which meaning is co-constructed. A comparison is immediately made to sociology. The authors note how conversation is a very flexible, interactive thing, where ambiguity is made and then clarified or contextually understood. This also, notably, is very tied to the environment, not abstract propositions, but world referents.

The authors describe four assumptions of literary models that are destroyed in real conversation. These are tacit idealizations that ascribe to communication formal elements of written speech which is simply absent.

The literary model makes these tacit idealizations. (1) The reference is a proper noun (for instance, Napoleon, King George), a definite description (this year, the man with the moustache), or a pronoun (he, this, they). (2) The speaker uses the noun phrase intending to the addresse to be able to identify the referent uniquely against their common ground. (3) The speaker satisfies her intention simply by the issuing of that proper noun phrase. And (4) the course of the process is controlled by the speaker alone.

The authors reference Goffman (!) and claim that understanding is developed interactively. Communicated statements are followed with “continuers.” These enable conversation continue, and confirm that meaning is shared. This is based on practice and performance.

What is also neat about this analysis is the fact that the discourse in the experiement is entirely metaphorical. The experiment consists of having the subjects be given a grid of tangrams (making up arbitrary, but evocative shapes) and giving one subject a list of tangrams to instruct the other to recognize. The shapes are being refered to by imagery, analogy, and association. All of these are embodied and none are propositional (at least not in the sense of the other essays).

Recognition, and general communication, is a process. It involves various potential reactions, which have active knowledge changing, belief changing, and linguistic elements. These examples are: acceptance, rejection, postponement. Most notably is that each of these responses may be considered an action.