Computational Models of Discourse Analysis Carolyn Penstein Rosé Language Technologies Institute/ Human-Computer Interaction Institute Warm-Up Can you explain speech acts from Gee’s perspective (using his terms)? i.e., what is a speech act? Can you explain speech acts from Martin and Rose’s perspective (using their terms)? Note that we can use work like Levenson’s Differently We talked about Gee’s 42 questions as being like feature extractors Martin & Rose’s systems give us a map, showing us where to look for evidence (almost the same?) Levinson’s work raises questions – shows us distinctions that need to be accounted for, potential holes in our approach Levinson helps us think in detail on a technical level about how particular words are functioning… •According to the performative hypothesis, “frankly” should be functioning the same way in all of these utterances, but it’s not. In 50 and 52 it modifies the way the message is told, in 51 it’s a warning that something negative is coming. •What does this tell us about using words as evidence in Pragmatic oriented interpretation? Which accomplishes a bet and how do you know? * What does that tell you about how to model speech acts computationally? Plan for Upcoming Unit For next time we will pass out examples from AMI and coding manual By Wed I will link in the readings for Unit 2 Next Monday I will hand out the annotated corpus and documentation for SIDE plugins Student Comment I'm not sure we can analyze a text using our current knowledge of speech acts. We can only determine if there are any speech acts and where specifically they occur. Speech act theory as we've discussed it so far doesn't include any analytical feature beyond sentence deconstruction, unlike Gee's and SFL methodologies. Can you pass the salt? It’s hot in here. Digging in to the roots… Gee: Anthropology Martin & Rose: Rhetoric and Literary Theory Levinson: Philosophy and Logic What’s the connection? Gee: Anthropology Speech acts are kinds of discourse practices, there are conventional ways of doing things with words as indicated through form-function correspondences Figured worlds set up conditionally relevant speech acts (adjacency pairs) Martin & Rose: Rhetoric and Literary Theory The Negotiation framework is a very simplified system of speech acts Student Questions I think that the main difference between speech act analysis and the other types of analyses that we have covered thus far is that it explicitly asks the analyst to consider a certain text’s many voices—its heteroglossia why would someone want to talk if they are not trying to achieve something (consciously, or unconsciously)? Please clarify… Throughout the conversation, Jim uses illocutionary force to direct the flow of conversation. From "go ahead" (2) to the series of probing questions (12, 15, 23, 33, 39, 57, 64, 82), his statements range from clearly pushing the conversation ahead to something that is closer to bordering on factual yes/no questions. Even as he gets more into informational questions, near the end of the dialogue, these are still compelling Bonnie to continue her explanations. What can we do with speech acts? Recognize what actions users are doing in a natural language based interface Sometimes mixed with domain level frames to detect what someone wanted in a task based dialogue system (scheduling a meeting, registering for a conference, making an airplane reservation) If clusters of speech acts are associated with roles, you can use them to identify roles within an interaction Current work on meeting summarization Some work on “social positioning” in the mixedinitiative dialogue community in the 90s What questions can we answer with them? You could think of them as moves in a game (like chess) Each move is part of a strategy Moves work together to accomplish intentions But each speaker has their own set of intentions – in some sense they are competing You can explain what strategies were effective or not for accomplishing any of these intentions From this analysis, you can conclude things about power, positioning, influence, etc. Why might someone be insulted when you politely explain something to them? You can talk about different social languages used to enact speech acts (e.g., direct versus indirect) What do we gain from this diversity? Gee: Anthropology Martin & Rose: Rhetoric and Literary Theory Levinson: Philosophy and Logic * In some ways our goals are different from all of these. But all of them have insights into how language works. Anthropologists know how it functions in societies. Rhetoriticians know how it functions in interpersonal relationships. And logicians have thought about how humans are able to interpret language from the available evidence. Why all the focus on truth conditions? Formal semantics is based on logic Evaluating meaning of a sentence is: translating the sentence into logic, evaluating it within a model Models define what is given and what inferences can be made Evaluating meaning in a model means evaluating whether the model makes that sentence true or not What is Pragmatics? Pragmatics is what is beyond truth conditional meaning Part of the chapter focuses on whether there is anything beyond truth conditional meaning If you accept that there is something beyond that, one thing there is is what language does, apart from what it means Formal pragmatics is about building models that allow us to compute what language does Speech act inventories Early Computational Approaches AMI Annotation Scheme Later Dialogue Act Tagging Approaches Text classification based approaches Things you should know: Using Naïve Bayes, SVM, HMMs, CRFs Note that HMMs and CRFs are statistical techniques that pick up on sequencing Conditional probability P(X|Y): If you know Y is true already, what is the probability that X is true? If probabilities are independent, you can combine them by multiplying Features: Unigrams, bigrams, POS bigrams What to think about as you read? What distinguishes the set of speech acts that are being focused on? What evidence would you as a human use to make your choice? What evidence is this computational approach taking into account? What is it missing? Can you think of examples that you think it would miss? Corpora for experimentation Unit 2: Maptask data (Negotiation coding) Possibly other chat corpora with same coding as well Unit 3: Product Reviews (Sentiment) Unit 4: Blog corpus (Age and Gender) Unit 5: AMI meeting corpus (Dialogue Acts) Other corpora Email discussion list (Social Support coding) Assignment 1 (not due til Feb2) Transcribe a scene from a favorite move, play, or TV show Select one of the methodologies we are discussing in Unit 1 (e.g., from Gee, Martin & Rose, or Levinson) Do a qualitative analysis of the data and write it up As a shortcut, you can find a script online Excerpt should be no more than one page of text Use readings from Unit 1 as a collection of models to chose from Due on Week 4 lecture 2 Turn in data, raw analysis (can be annotations added to the data), and write up (your interpretation of the analysis) Not required now!! Prepare a powerpoint presentation for class (no more than 5 minutes of material) Other Ideas: Twitter data, Google Groups, transcribe a real conversation (if your conversational partners agree…) Questions?