Speech Act Theory, Dialogue, and Information State Updates Staffan Larsson May 30, 2013
Speech Act Theory Speech acts and dialogue Plan-based dialogue models Information state update models
Outline Speech Act Theory Speech acts and dialogue Plan-based dialogue models Information state update models
Speech act theory: Language as action Utterances do not just carry information; they change things Utterances are actions Wittgenstein: Philosophical Investigations (1958) Austin: How to do things with words (1962) Searle: Speech Acts
Characteristics of performatives Performatives are not used intentionally to say anything (true or false) about states of affairs in the external world Their use constitutes an action 1. Christening: I hereby name you the Princess Elizabeth 2. Marrying: I hereby declare you husband and wife 3. Sentencing: I hereby sentence the accused to ten years in prison 4. Promising: I promise to be home on time 5. Apologising: I apologize for being late 6. Ordering: I order you to surrender immediately! Some performatives are part of a conventional or ritual behaviour supported by institutional facts (1-3 above) while others are not
Problems with the performative / constative distinction Some utterances pass the hereby test, but are used to state or assert I hereby state that John is a farmer I hereby tell you that the prime minister is not going to stand down Conclusions: constatives are a class of performatives The strict distinction cannot be maintained
Locutionary, illocutionary, perlocutionary All utterances perform specific acts via the specific communicative force of the utterance In every utterance, several acts can be distinguished Locutionary Illocutionary Perlocutionary
Locutionary act Saying something with a specific meaning Phonetic: producing certain sounds, or writing certain letters Phonetics, phonology Phatic: uttering certain words, phrases and sentences according to the rules of a language morphology/syntax Rhetic/propositional: using the product of a phatic act with a certain meaning and reference semantics (& pragmatics)
Illocutionary act Saying something with some purpose in mind The type of function that an utterance is intended to fulfil by its speaker What you do in uttering U Examples: Accusing, apologizing, blaming, congratulating, giving permission, joking, nagging, naming, promising, ordering, refusing, swearing, thanking Governed by conventions (social rules)
Illocutionary act Speech act often used for illocutionary act Consist of illocutionary force+ propositional content utterance: Shut the window! illoc. force: request Content: the act of (the hearer) closing the window In one utterance, several illocutionary acts can be simultaneously performed Can you close the window? can be both a question and a request
Perlocutionary act A perlocution is the act by which the illocution produces a certain effect (the perlocutionary effect) in the addressee Non-necessary effects; depend on context rather than linguistic conventions Perlocutionary verbs: frighten, convince, surprise Does not pass the hereby test # I hereby frighten you
Differences illocution - perlocution Intention Illocutionary act is always intentional Perlocutionary act are not always intentional Control Illocutionary act are under the speaker s control Perlocutionary acts are not under the speaker s full control Evidentiality Illocutionary act become evident as the utterance is made Perlocutionary acts are not usually evident until after the utterance is made Determinate? It is usually clear which illocutionary act has been performed It is often unclear which perlocutionary act has been performed Conventionality Illocutionary acts are more conventionally bound to linguistic forms Perlocutionary acts are less conventionally bound to linguistic forms
Searle: Conditions on speech acts Austin put forward felicity conditions on speech acts, corresponding to truth conditions for declarative sentences Searle: felicity conditions are constitutive rules rules which create the activity itself of speech acts To perform a speech act is to obey certain conventional rules which are constitutive for that type of speech act
Searle: Felicity conditions on speech acts (i) Propositional content condition specifies restrictions on propositional content of utterance promise: future act A of speaker (S) request: future act A of addressee (H) (ii) Preparatory: Real-world prerequisites promise: it is clear to S and H that H prefers A to be accomplished; A will not occur in the normal course of action request: S thinks that H can carry out A; H will not carry out A unless asked
Searle: Felicity conditions on speech acts (iii) Sincerity: must be satisfied if the speech act is to be carried out sincerely (if it is not fulfilled, it will be a case of abuse ) promise: S intends to carry out A request: S wants H to carry out A (iv) Essential S intends U to count as the speech act in question, and H recognizes this promise: S intends for U to count as an undertaking to do A (alternatively, create an obligation on S to do A) request: S intends for U to count as an attempt to get H to do A (alternatively, create an obligation on H to do A) If the essential condition is not fulfilled, the speech act has not been carried out
Searle s felicity conditions for promising (i) Propositional content: future act A by S (ii) Preparatory (a) H would prefer S s doing A to his not doing A, and S so believes (b) it is not obvious to both S and H that S will do A in the normal course of events (iii) Sincerity: S intends to do A (iv) Essential: the utterance of U counts as an undertaking to do A
Types av speech acts (Searle) Representatives / Assertives committs S to the truth of a proposition P Crick and Watson discovered the structure of DNA Directives attempt by S to get H to do some future action A Turn the TV down Commissives commit S to some future action A I ll be back in five minutes Expressives: express psychological state of S Ouch! Declarations effect changes in some immediate state of affairs President: I declare a state of emergency
Direct and indirect speech acts Basic sentence types Declarative Interrogative Imperative These are morphologically and/or syntactically distinguished in many languages They are typically associated with three basic illocutionary forces: Declarative - asserting / stating Interrogative - asking / questioning Imperative - ordering / requesting
Direct and indirect speech acts Direct speech act match between sentence type and illocutionary force Pass the salt! explicit performatives (in declarative form) I request you to pass the salt. Indirect speech act no direct relationship between sentence type and illocutionary force I will be sitting at the front can be an assertion or a promise Can you open the window? can be a request or a question
Analyses of indirect speech acts: Dual illocutionary force Searle s (1975) analysis: indirect speech acts have two illocutionary forces one literal/direct one non-literal/indirect The existence of an indirect speech act depends on relation between utterance and speech act felicity conditions Can you pass the salt Violates felicity condition on questioning (direct act) S already knows whether H can do A Queries felicity condition for requesting (indirect act) Whether H can do A (preparatory conditions) Performing and understanding indirect speech acts involves inference According to Searle, along the lines of Grice s Cooperative Principle Speech acts and implicature appear closely linked
Analyses of indirect speech acts: Short-circuited implicature However, indirect speech acts are frequently conventionalised Can you pass the salt Are you able to pass the salt? Do you have the ability to pass the salt? Indirect speech acts are short-circuited implicatures (Morgan 1978); in principle calculable, but not in practice calculated Test for conventionality of indirect request: try inserting please Can you please pass the salt?? Are you able to please pass the salt?? Do you have the ability to please pass the salt? Similar idea (Gordon & Lakoff 1975): Conversational postulates are inference rules which reduce the amount of inference needed to interpret a speech act Indirect speech acts involve both inference and convention
Analyses of indirect speech acts: Idiom model of indirect speech acts Indirect speech acts are semantically ambiguous (1) Can you pass the salt Interpretation involves no inference at all; (1) is simply recognised as a request with no question being perceived Problems: Fails to capture the fact that the meaning of a speech act can frequently be derived from the meaning of its components (compositionality) Idioms should be quite arbitrary, but indirect speech acts are quite comparable cross-linguistically An interpretation which takes into account the literal meaning of an indirect speech act is not allowed A: Can you pass the salt B: Yes, I can. [passes the salt]
Outline Speech Act Theory Speech acts and dialogue Plan-based dialogue models Information state update models
Speech acts and dialogue Dialogue is the most fundamental arena of language use Modeling dialogue = designing (formal) theories which reproduce aspects of natural conversation Theoretical purpouse: understand how language is used in interaction Practical purpouse: build dialogue systems enabling human-machine interaction using spoken dialogue
Challenges in modeling dialogue Utterances: how is dialogue segmented into basic units? Problem 1: repairs Problem 2: fragments Interaction management: How is dialogue managed by speakers and hearers in real time to allow successful and efficient communication? grounding turntaking Coherence: How do utterances in dialogue relate to each other? Adjacency pairs Dialogue acts Conversation structure
Utterances Spoken dialogue does not come pre-segmented into units in the same way as written language does Ungrammaticality due to self-repairs until you re at the le- I mean at the right-hand edge of the quarry Elliptical utterances (sub-sentential utterances, fragments) A: Who came to the party? B: Sandy.
Interaction management: Turn-taking In dialogue, participants take turn in speaking Normally accomplished very smoothly Overlaps of speech are rare (less than 5% of turns) Pauses between turns are short Taking or handing over the floor seems to be organised cooperatively between participants through the use of certain turn-taking devices, including the following: Using particular intonational contours Producing certain dialogue acts that select the next speaker (questions, requests) Making visible signs that one wants to begin speaking
Interaction management: Initiative Dialogue initiative: controlling who gets the next turn Task initiative: held by the dialogue participant driving the task forward These factors can vary independently E.g., asking a clarification question gives you dialogue, but not necessarily task initiative. There are dialogue genres where initiative is regularly distributed unevenly between the DPs (e.g., interviews, tutorials, exams) In free conversation it is normally held in equal measures by all DPs.
Interaction management: Feedback and grounding Common ground = things that are shared / mutually believed between speakers Utterances incrementally add to common ground To ground a thing... is to establish it as part of common ground well enough for current purposes. (Clark) Making sure that the participants are perceiving, understanding, and accepting each other s utterances Accomplished by simple feedback utterances such as OK and clarification requests A: I saw Peter. B: Peter? / Who? / You did what? / Pardon? Dialogue proceeds along two parallel tracks: Main business Interaction management including feedback
Coherence: Adjacency pairs Speech act theory does not say much about the relation between different speech acts However, this has been studied in the Conversation Analysis tradition (Sacks, Schegloff, Jefferson) Adjacency pairs: Sequences of two utterances which are adjacent and produced by different speakers Part 1 precedes part 2 (not necessarily directly precedes) Examples question - answer greeting - greeting offer - acceptance Uttering the first part of an adjacency pair creates an expectation for the second part
Speech acts and dialogue Speech acts as originally defined by Austin don t model key features of conversation such as coherence and interaction management In order to capture these conversational phenomena, we use an extension of speech acts called dialogue acts (or dialogue moves or conversational moves) Several dialogue act taxonomies have been proposed DAMSL, Verbmobil, Map Task,... DAMSL Dialogue Act Markup in Several Layers Although never officially published, widely used (often in modified versions) in research
Features of DAMSL dialogue acts A single utterance can be analysed as constituting more than one act allows a principled analysis of the various functions an utterance can have. Distinguishes two kinds of utterance functions: Forward- looking function (FF) Individuated with respect to the way an utterance constrains future discourse Records the effect of utterances on the public beliefs and obligations of the utterer Backward-looking function (BF) Describes how the utterance relates to previous discourse. Additional dimension of analysis: the information level at which an act operates task task-management communication management.
DAMSL forward looking functions
DAMSL backward looking acts
DAMSL example Example A: Please open the door. [FF: action-directive] B: OK. [FF: committ; BF: accept, acknowledge]
Conversation structure Subdialogues Clarification requests as discussed above can be seen as opening sub-dialogues that are inserted into the main dialogue Similarly, sub-dialogues can concern corrections, knowledge preconditions ( do you know what this is? ), etc. At an even higher level of abstraction, one might want to distinguish phases in conversations (conversation analysis) openings main business closings
Kinds of structure in dialogue Micro-structure Structure within the utterance Hesitations ( um ), feedback ( mm ), pauses, turn-taking devices ( right? ) Meso-structure Structure connecting utterances E.g. adjacency pairs, forward/backward-looking dialogue acts Macro-structure Structure connecting utterance sequences (phases) E.g. phone call: identification - greeting - small talk - main business - initiate ending - end call
Approaches to modeling dialogue Modeling what was said Modeling what to say next Modeling when to say it
Modeling what was said Identify utterance boundaries, remove disfluencies Figuring out which dialogue acts are realised by an utterance Symbolic model 1: Two-stage approach 1. A set of surface speech acts is computed through syntactic and semantic analysis of the utterance 2. The intended speech act is identified through contextual reasoning Symbolic model 2: SDRT Combines information coming from various sources, including lexical and compositional semantics, discourse structure, and cognitive states of participants Makes detailed predictions about what makes dialogues coherent
Modeling what to say next Structured dialogue models Plan-based dialogue models Information state update models Commonly, models regarded as modeling an individual dialogue participant ( self ) Models are described as specifying states and transitions between them Difffer in what they assume the states are, i.e., which kind of information they assume must be represented, and in how the transitions are specified
Structured dialogue models Common way of building dialogue systems States are atomic, i.e. they contain no information States specify what self says ( promtps ) Transitions between states are triggered by utterances (by the dialogue partner) State transition networks are either very inflexible (only utterances specified in transitions are allowed) or unmanageable (since there must be a transition for all possibile utterances in any state) Good for pre-structured tasks Strict control over the dialogue also makes them less appropriate for more dynamic and self-organised tasks
Outline Speech Act Theory Speech acts and dialogue Plan-based dialogue models Information state update models
Plan-based dialogue models A theory of illocutionary acts should be based on a general theory of action Artificial Intelligence (AI) offers such a theory Formalisation of actions state descriptions of the world mental states, attitudes how actions relate to world states and mental states
Positioned at the other end of the flexibility scale. Flow of the dialogue is based on local inferences over rich contextual representations (recording beliefs, desires and intentions of agents) Develops out of principles that are seen to be general for intentional behaviour.
Actions in AI An action A is defined by preconditions: what has to be true in order for someone to do A decomposition: how A can be divided into subactions effects: what is true after A is done intended effects - the goal of A side-effects constraints on entities involved in A (agents, propositions,...) Such a definition is called an action operator (or plan operator)
Preconditions and effects in terms of mental attitudes Ordinary actions affect the world Speech acts affect mental states So, speech act plan operators provide preconditions and effects in terms of formalised mental attitudes
The BDI model What is needed for rational behaviour? The ability to perceive the world Beliefs about the world - how things are Desires - how things should be The ability to plan and make decisions Intentions - what to do The ability to act Using the BDI-model (or similar), we can model conversational agents
An example of formal speech-act plan operators request(s,h,a) precondition: want(s, do(h, A)) achieves: bel(h, want(s, do(h, A))) cause-to-want(s,h,a) precondition: bel(h,want(s,do(h,a))) achieves: want(h, do(h,a)) do(x,a) precondition: want(x,do(x,a)) achieves: done(a) deletes: want(x,do(x,a))
Planning and recognizing speech acts Utterance generation as planning given a communicative goal G and a set of speech act plan operators, generate a plan to achieve G Speech act recognition as plan recognition given an utterance U by speaker S, find a communicative goal G and a plan P to achieve G such that U is a part of P, and S might plausibly have the goal G G determines the illocutionary act
The Intentional Structure of Dialogue The BDI approach to utterance interpretation gives rise to another view of coherence which is particularly relevant for dialogue - the intentional approach (Grosz and Sidner, 1986) What makes a dialogue coherent is its intentional structure the plan-based intentions of the speaker underlying each utterance. Intentions are instantiated discourse purpose (DP). underlying purpose of each discourse held by the person who initiates it Each discourse segment within the discourse has a corresponding purpose a discourse segment purpose (DSP) has a role in achieving the overall DP. Possible DPs/DSPs include intending that some agent intend to perform some physical task, or that some agent believe some fact
Outline Speech Act Theory Speech acts and dialogue Plan-based dialogue models Information state update models
Information State update approach View of dialogue management functions in terms of information state Key to this approach: Identifying the relevant aspects of information in dialogue How this information is updated as the dialogue proceeds
Information State The information state of a dialogue represents cumulative additions from previous actions in the dialogue and motivating future action Example: statements generally add propositional information Example: questions generally provide motivation for others to provide specific statements
Components of an ISU theory of dialogue modeling An information state A set of dialogue moves trigger the update of the information state correlated with externally performed actions such as particular natural language utterances A set of update rules govern updating of the information state given various conditions of the current information state and performed dialogue moves A set of selection rules license choosing a particular dialogue move to perform given conditions of the current information state enabling participating in a dialogue rather than just monitoring one
example IS and rule Information state type: Moves: assert(p), askif(p) Sample rule: integrate assert if in(lm, assert(p)) bel : Set(Prop) des : Set(Prop) int : Set(Action) mbel : Set(Prop) lm : Set(Move) then add(is.mbel, P) add(is.bel, P) del(lm, assert(p))
Example rule application
The QUD model IS distinguishes between information that is deemed private to the system (e.g., its plans and agenda) and such which is shared (i.e., has been grounded between the DPs). The dialogue is structured around the concept of the question under discussion (QUD)
Basic dialogue state in QUD model (simplified) private : shared : agenda : OpenQueue(Action) plan : OpenStack(PlanConstruct) bel : Set(Prop) nim : OpenQueue(Move) com : Set(Prop) qud : OpenStack(Question) pm : OpenQueue(Move) lu : [ speaker : Participant move : Set(Move) ]
Basic issue-based dialogue management Genre: Inquiry-oriented dialogue (database search) Dialogue moves: ask, answer, greet, quit Dialogue seen as, essentially, raising and addressing issues Deals with short answers. e.g. yes, no, Paris, in Dialogue plans Sample domain: travel agency
Semantics Simple First Order Logic without quantifiers, but with questions Questions Y/N-questions:?P, where P is a proposition Wh-questions:?x.p(x), where p is a predicate Here,? works much like like λ Alt-questions: {?P 1,...,?P n } Content of short answers: individual markers: paris, april,... yes, no
Semantics, cont d Q-A relations (adapted from Ginzburg) resolves(a,q): A resolves Q dest-city(paris) resolves?x.dest-city(x) relevant(a, Q): A is relevant to Q not(dest-city(paris)) is relevant to?x.dest-city(x), but does not resolve it
Sample dialogue plan goal: resolve?x.price(x) body findout(?x.dest-city(x)) findout(?x.depart-city(x)) findout(?x.transport(x)) findout(?x.dept-month(x)) findout(?x.dept-day(x)) raise({?class(economy),?class(business)}) consultdb(?x.price(x))
Sample rule: Answer integration Name: integrateanswer If then in( shared.lu.moves, answer(a) ) Q = shared.qud.top domain.relevant( A, Q ) P = domain.combine( Q, A ) shared.com.add( P ) Synopsis: Before an answer can be integrated by the system, it must be matched to a question on QUD
Basic dialogue with updates U: price information please ; raises price issue if user asks Q, push Q on QUD and push respond(q) on agenda if respond(q) on agenda and plan empty, find plan for Q and load to plan if findout(q) first on plan, ask Q S: where do you want to go? U: Paris if lu.moves contains answer(a) and A relevant to Q, add P =combines(q, A) to shared.com if P in shared.com and Q topmost on QUD and P resolves Q, pop QUD if P in shared.com and P resolves Q and findout(q) on plan, pop plan
Basics cont d... S: Do you want economy class or business class? U: economy class if consultdb(q) on plan, consult database for answer to Q; store result in private.bel if Q on QUD and P in private.bel such that P resolves Q, answer(p) S: The price is SEK 123
Typical human-human dialogue S(alesman), C(ustomer) S: hi C: flights to Paris S: when do you want to travel? C: April, as cheap as possible
Accommodation Accommodation has been applied to referents and propositions, as parts of the conversational scoreboard / information state Question accommodation If questions are part of the information state, they too can be accommodated If the latest move was an answer, and there is an action in the plan to ask a matching question, then put that question on issues (and QUD if it is a short answer) Requires that the number of possible matching questions is not too large (or can be narrowed down by asking clarification question)
Sample dialogue: accommodation S: Welcome to the travel agency. U: From London to Paris in April Not relevant to any question that has been raised, or to any current plan Look in domain knowledge for a plan (for dealing with some question Q) with matching questions Load this plan, push Q on issues Find in plan the question(s) matching the user s answer, and push them onto issues Integrate answer (requires matching question on issues) S: Alright, you want to know about price. Proceed to next plan item S: How do you want to travel? shared.issues=<?x.how(x),?x.price(x)>