20 Questions on Dialogue Act Taxonomies

There is currently a broad interest in dialogue acts and dialogue act taxonomies, and new uses, taxonomies, and standardization efforts continue to be proposed. This paper presents a discussion of issues that are important to be addressed, if taxonomies are meant to be shared and understood the same way by proposers and others. The discussion is framed in terms of 20 questions, the answers to which will help make the meanings of taxonomy elements more clear to disjoint communities of users. Introduction There is currently a very wide range of theories and taxonomies of dialogue acts1 available for a researcher to choose from. Moreover, specific deficits in any theory lead researchers to continue to develop new taxonomies to suit their particular purposes. To some degree, this is to be expected; dialogue act taxonomies can be seen as a kind of language for describing communicative events, and there is certainly no deficit in the continued creation and development of new formal languages (e.g., programming languages like Java) or (at a slower pace) natural languages. On the other hand, in both natural and artificial languages, the use of similar signs for different concepts can cause confusion and misunderstanding, often with serious undesirable consequences (e.g., in programming languages, the use of = as an assignment rather than equality operator in a boolean context; or the firing of an American city official for using the word niggardly (of independent Scandinavian origin) because it sounded too similar to an offensive racial epithet euphemistically referred to as “the N word”2. Similar confusions often occur when one researcher tries to interpret the dialogue act taxonomy of another. For example, when encountering an act labeled as inform: which subset of the constraints in (1) can said to be claimed (or perhaps none of them, depending on some other formulation entirely). This kind of confusion has led some (e.g., (External Interfaces Working Group, 1993; Discourse Resource Initiative, 1997; FIPA, 1997)) to propose standard theories that could be well-defined and understood and used across groups while others (e.g. (Allwood, 1977; Cohen and Levesque, 1990)) prefer to treat dialogue act (i.e., illocutionary force) identification as of only secondary importance, as a derived concept within a more general theory of rational interaction, using other primitives. (1) a. declarative mood was used b. propositional information was expressed c. new information was expressed d. the addressee came to believe what was expressed e. what was expressed is actually believed by the speaker f. what was expressed is actually true It is hard to dispute that dialogue acts are a useful concept, given the wide variety of uses to which they are put. Some of these uses include3: representations of the pragmatic meaning of utterances in dialogue theories (Vanderveken, 1991; Bunt, 1996; Poesio and Traum, 1997; Poesio and Traum, 1998), building blocks for grammars of dialogue (Winograd and Flores, 1986; Bilange, 1991), labels for corpus annotation (Carletta et al., 1997; Alexandersson et al., 1997), agent communication languages (External Interfaces Working Group, 1993; Sidner, 1994; FIPA, 1997; Singh, 1998), object of analysis in dialogue systems (Allen et al., 1996; Bretier and Sadek, 1996), and element of a logical theory of rational interaction (Sadek, 1991). Despite this popularity of the concept, there are still a number of issues that present significant challenges for creating a taxonomy of dialogue acts that can be understood and used by other groups. Here, I will briefly raise some of the issues that have often caused confusion when interpreting one taxonomy of dialogue acts within the viewpoint of another. These issues must be addressed in order to have a clearer idea of what one means by saying a dialogue act occurred, whether the dialogue act taxonomy is meant for labeling a naturally occurring corpus, as part of a formal theory of action, or as a system-internal representation of the dialogue. Although there are many such issues and variations on the ones listed below, I focus here on 20, in the form of questions, in homage to the “dialogue game” named in the title. These questions are grouped, for convenience into sections of related questions. Defining Dialogue Acts 1: Which is most important: fit to intuitions or formal rigor? This question has implications beyond just dialogue act definitions, but is applicable for any attempt to provide a formal theory of commonsense notions. Very often it is difficult to precisely formulate complex intuitions using available formal techniques. The question then arises as to which goal to sacrifice for the time being: should one formalize a simpler notion that does not have the same properties of the intuitive concept (e.g., for belief, an undesirable property of logical omniscience, in a normal modal logic), or sacrifice some desirable formal properties, such as a model-theoretic semantics, necessary and sufficient conditions for categories, or soundness and completeness of an inference system. The answer will depend on the purposes to which the concept is to be put: if the primary goal is to discover and prove properties of the system, formal properties are not easily sacrificed. On the other hand, if the goals are more empirically motivated, a formal system with undesirable properties may not be close enough to be useful, while a concept without some of these properties may suffice for the task at hand (corpus labelling or use in a computer program). There should also be a place for intermediate points, that make some sacrifices at each side, while striving for maximum utility in a given purpose. In particular, with respect to dialogue acts, it can be relatively easy to state precise definitional conditions of occurrence within a formal logic of action, however a problem may arise when these conditions diverge from a more intuitive (and intuitively useful) notion of action that empirical analysts and dialogue system designers would actually like to use. 2: Is the definition of a dialogue act an issue of Lexical Semantics or Ontology of Action? There are different tasks one might be attempting when defining the meaning of a dialogue act. Is it to provide an account of when someone might be justified in describing an occurrence using a sentence headed with a particular verb (e.g., inform, request), or to provide a technical vocabulary to compactly describe various types of occurrences in convenient ways for analyzing other aspects of interaction. As (Allwood, 1977) warns, these endeavors should be clearly separated, even if one might want to use similar categories to describe each (as is done in (Allwood, 1980)), or maintain a position of identity of semantic and conceptual structure (Jackendoff, 1983). Intuitions, or annotation by naive coders without instructions to the contrary may tend to focus on the former enterprise, which may have undesirable consequences for the way in which the taxonomy is to be used. The key question is how much weight, if any, should be given to linguistic intuitions about when it is true or appropriate to use a particular sentence to describe an occurrence. For lexical semantics, this is the paramount question (barring issues of polysemy), while it might not really be a factor for an ontology which might diverge from language classifications for independently motivated reasons. 3: Under what conditions may an action be said to have occurred? There are a number of different criteria that are being used to decide whether or not an action occurs in a given situation. (Allwood, 1980) uses four criteria, shown in (2), each of which can be a sufficient condition for ascribing that an action has occurred, while none is necessarily present. (2) a. intention of performer b. form of the behavior (e.g., linguistic form)