New Approaches to Story Modeling for Understanding, Generation and Summarization David K. Elson December 6, 2004 What kind of a talk is this? • About modeling – In principle • About stories – Theory and practice • About stories and NLP – History – New news summarizer • Initial results • Not about generation background 2 The Plot 1. 2. 3. 4. 5. 6. 7. A Word on Models Narrative for News? Related Work The Big Idea Waxing Philosophic Back to News Wrapping Up 3 The Plot 1. A Word on Models 2. 3. 4. 5. 6. 7. Narrative for News? Related Work The Big Idea Waxing Philosophic Back to News Wrapping Up 4 What is a model? • A discrete set of conceptual categories meant to quantize the range of expected values of a natural phenomenon. • A machine or human can apply the model to an instance of the phenomenon by assigning a category to each token to a category (“parsing”). – An adaptive system updates its model to fit new patterns in new data. 5 Example of a model Phenomenon Model ABCDEFGHIJKLMNOPQRSTUVWXYZ Phenomenon, parsed according to model DATE DID SOMETIMES WITHOUT PROBABLY AFTER PROBABLY SOON AMONG ITSELF ONCE SENT OUT HOUSE GOING SAD UNDER UPON 6 More on models • We as humans work by giving inductive conclusions to the categories in our models based on statistical patterns from experience. – What is a tree? What is blue? • For example, we have seen enough houses with bedrooms to expect a bedroom when we discuss a house. – This is an expectation model. 7 How many distinct categories? • Too few, and our deductions are wrong. – “Does the House of Representatives have a bedroom? • Too many, and we can’t make inductions. – “Is there going to be a room for me to sleep in at Grandma’s house?” 8 Equivalence in NL modeling • Statistical language modeling – Unigrams D(w1) • Enough data to see patterns, but • So coarse that the data are heterogeneous, and the inductions from patterns are wrong. – Trigrams D(w3|w1,w2) • Specific enough to for patterns to be useful (high information content), but • Too many categories, too little data to see patterns. 9 More on expectation models • What am I saying here? • Parse using Model 1 (English, millions of categories) – Versatile, but not very useful • Parse using Model 2 (fifty categories) – Model doesn’t fit data: neither versatile nor useful • Parse using Model 3 (three categories) – Useful, but not very versatile 10 Bigger problem • Humans are good at creating and maintaining useful models subconsciously to make knowledge. • They then try to pass on that knowledge to the computers, by “arbitrarily” defining categories. – Language is messy. – Every model “compiles,” whether or not it’s right. 11 Modeling nature with machine learning It’s C2! ML Features F1, F2 … F2000 Model: this is C1, C2, or C3 12 We tend to throw ML at a problem. I think it’s C1… ML Features F1, F2 … F2000 13 Is This A Bad Idea? • If we were classifying rocks, then no. – Phenomenon is static and well-understood. – Emphasis should be on new feature detection. • But we are modeling human language. – Incredibly complex • emotion, culture, upbringing, a wide range of intentions. • We are more comfortable and skilled with computers than with philosophy. 14 How would you model these NLP problems? • • • • • • Fact or opinion? Relevant sentence or irrelevant? News or editorial? How to generate text with style? How to detect tone? How to detect emotion? 15 The good news • There are many people in other fields whose whole problem is your first step: finding a model. – Theories in journalism, linguistics, psychology • You can sometimes even ask questions to the people who created your corpus! – “Poems are made by fools like me, but only God can make a tree.” 16 The main point • Look closely at your data. • Pick practical, yet useful distinctions for your categories that fit the data. – Your final research goal, without grey areas, may be an impractical model to train on. • Look at the non-technical literature. • But don’t fall into a rabbit hole of modeling! – Remember, the model problem will never be solved! 17 The Plot 1. A Word on Models 2. Narrative for News? 3. 4. 5. 6. 7. Related Work The Big Idea Waxing Philosophic Back to News Wrapping Up 18 My research • A model for stories and storytelling. • Abstract, but potentially applicable to many NL fields. • I’m also applying it now to news summarization. 19 Untapped assumptions for news? Text Narratives Narrative News Articles News 20 Story Form for News Summarization? • Many news articles are narrative in form – Structural restrictions • Frequently use quoted opinions, statements of history, character descriptions… – Content restrictions • Every news article happens because an editor thought the story would be interesting. How did they decide? • If stories can be modeled, so can news – News summaries are minimal retellings of the news story – If we can tell a story, we can write a summary 21 News is Newsworthy When It’s a Good Story JERUSALEM (CNN) -- The Israeli military has embarked on a wide-ranging raid in Jenin after getting tips that the Palestinian militant group Islamic Jihad was planning to carry out terror attacks against Israel from the northern part of the West Bank, the Israel Defense Forces said Thursday. • Syntactic: description of causally related events – Consequence followed by cause • Semantic: Narrative characters and goals – “Avoidance of terror” goal, “preemptive strike” plot 22 The big questions • What is a story? • Is there a model for stories that we use as humans and journalists? – Is there a small set of possible stories? – Do we have a story expectation model that fills in the gaps? • Can the model be encoded? – Can a computer understand a story or tell a story? – Can it do it well enough to be useful for news summarization or story generation? 23 The Plot 1. A Word on Models 2. Narrative for News? 3. Related Work 4. 5. 6. 7. The Big Idea Waxing Philosophic Back to News Wrapping Up 24 Everyone wants a model Story Generation Story Understandin g Creative Hack Theory Psychology Model of Stories Literary Theory Structuralist, Constructivist and Semiotic Theory 25 Many Models, Many Dimensions It’s a GRAMMAR! It’s about FRAMES! SEMIOTICS! PLOT FRAGMENTS! It’s about FABULA! CHARACTER GOALS! 26 An essential distinction to start with Text structure: Hook, elaboration, background (news) Rising/falling action, climax, coda (fiction) Discourse Narrative semantics: Heroes, villians, goals, obstacles, plot arcs, themes, morals Story Raw timeline: Actions, events, characters Fabula Observed Implied Most models are models of one of the top two layers. Some conflate them all. 27 For example Discourse Story Fabula 28 Fabula John and Joe’s Story: Fabula 0 • • • FABULA • • TIME • LINE • • • • John and Joe’s wife, Sue, have an affair. John moves to Vermont. John gets a job in broadcasting. Joe finds out about the affair and divorces Sue. John becomes a well known anchor. Joe becomes depressed. John and Sue get married and have kids. John is stalked by a mysterious person. John hires a PI and a bodyguard. Joe is arrested for stalking John. 29 Layers of narrative Discourse Story Fabula 30 Story • The fabula is perceived by a focalizer – Perceived by either (point of view) • An omniscient being • A character in the fabula • Some nested combination (e.g., Sherlock Holmes) – Events can be dismissed or emphasized – Events may be perceived out of order 31 Story • Narrative semantics are applied a fabula to make it palatable and interesting – Invoke our own models of what a story is • Based on our life exposure to stories • We need someone to “root for,” a conflict to care about – My research: What is that model? • What are the classes of plots, characters, themes that we expect? • Natural to writers, known to all readers, but hard to identify 32 Story John and Joe’s Story: Focalization #1 “Famous Newcaster Threatened by Stalker” – John is a funny, easygoing guy and a popular newscaster with a loving wife and kids. – One day John is heckled while on air; later he is followed home. The perpetrator escapes. – John’s pet cat is abducted. He feels very frightened. – John hires a PI and a bodyguard. This creates conflict with his family. 33 Story John and Joe’s Story: Focalization #2 “Betrayed Friend Seeks Revenge Against Arrogant Famous Newscaster” – A disheveled drifter, Joe, arrives in Vermont. – Joe describes his mission to a waitress: – John stole his wife from him. – John unapologetically left and got famous. – Joe demands compensation from John. 34 A brief sidestep: what narrative semantics did we just discover? • • • • • • • • The personal revenge plot Love as a device for battle The sympathetic, comfortable hero The sympathetic, downtrodden hero The arrogant, selfish, rich villain The deranged, threatening villain The stalking and theft of property plots The pet cat as beloved property Is there a complete list? How is it organized? 35 Layers of narrative Discourse Story Fabula 36 Discourse • The surface depiction of the story • Many “genres” of narrative discourse: – – – – – – – Written prose Stage drama Cinema Poetry Hieroglyphs Shadow puppets News article • Each has its own structural characteristics 37 Discourse The narrative agent’s tools in prose • • • • • • • • • Descriptions of actions Quotations of speech Selection of details Comments and descriptions Imagery Foreshadowing Grammar and word usage Ideological pontifications Historical digressions 38 Many Models, Many Dimensions It’s a GRAMMAR! It’s about FRAMES! SEMIOTICS! PLOT FRAGMENTS! It’s about FABULA! CHARACTER GOALS! 39 Literary Theory Campbell Plot Fragments • 1949: Mythologist Joseph Campbell publishes Hero with a Thousand Faces – Societal and psychological importance of myth – Myth as cultural heritage – Myth archplots: • Call to adventure and refusal of call • Threshold crossing, dragon-battle, dismemberment, crucifixion, abduction, wonder journey, whale’s belly • Return, rescue, resurrection, threshold struggle • Sacred marriage, father atonement, apotheosis • Still not formal or complete enough to be useful 40 Literary Theory Propp Plot Fragments • 1966: Russian Formalist Vladimir Propp looks at the Morphology of a Folk Tale • Most Russian folk tales draw from a pool of 31 plot fragments – – – – – – The hero leaves home The villain attempts to deceive the victim The hero and villain join in direct combat The false hero or villain is exposed The villain is exposed The hero marries and ascends to the throne • A grammar dictates how the fragments fit together 41 Story Understanding Early Software Plot Fragments • SAM: Script Applier Mechanism (Cullingford, 1978) – Reads NL text into causal chains – Rigid knowledge base and inference engine – Small input tolerance (only handles trivial stories) • PAM: Plan Applier Mechanism (Wilensky 1978) – Models around character goals, infers motivations – Still heavy reliance on inference engine, knowledge base – Still can’t handle real stories • TALE-SPIN: first generator (Meehan 1979) – On-the-fly problem solving – Stories are often not interesting, little conflict 42 Story Understanding • Mid-80s Marker Passing FAUSTUS (Norvig 1987) and ATLAST (Eiselt 1987) are marker passing systems – – – Less bound by rigid scripts and frames than SAM and PAM Markers are “passed” through the semantic network until they run out of “energy” Strong inferences are drawn, weak ones dismissed 43 Story Understanding • • Rumelhart Grammars 1975: David E. Rumelhart Syntactic rules for story structure: – Rule 1: Story = Setting + Episode – Rule 6: Internal Response = (Emotion | Desire) – Rule 10: Preaction = Subgoal + (Attempt)* • Summarization rewrite rules – • Summary(CAUSE[X,Y]) = “Instrument(X) caused (Y)” Grammar is not powerful, but very influential – – Is a grammar the right approach? Large debate (e.g., Wilensky, Black) 44 Story Generation MINSTREL Frames • Scott Turner 1992 • Story generation and surface rendering • Models the human process of storytelling – Consistency, suspense, fulfillment of themes • Appears “creative” with case-base reasoning – Recalls, adapts, and applies previous solutions • Domain: King Arthur’s Court – Goals include obtaining wives, slaying dragons 45 “Today”: Rebirth • Away from the symbolic, toward the practical • 1999: AAAI Symposium on Narrative Intelligence – – – – – MIT: Narrative as retold in social networks (Usenet) MIT Media Lab: Converting travelogues to narrative form Swedish Institute of CS: Browsing the Web in narrative Brenda Laurel: interfaces with characters, metaphors IIT (Canada): Narrative with autonomous agents – New definitions, new models • Based on conflict between characters (Szilas) • Based on the reader’s reaction (Bailey) • Based on new grammars (Lang) 46 Story Generation BRUTUS.1 Grammars • 2000: Selmer Bringsjord and David Ferrucci (RPI) • Story generation and discourse rendering • Formal representation of dramatic themes – Betrayal, self-deception • Renders with both grammars and simulations – Literary techniques (irony, metaphor) • Domain: a student’s dissertation defense – Very small! – Don’t know how “serious” this is, but a lot of press! • Optional New York Times reading 47 My Take On This History • Use of narrative semantics conflated with knowledgebased AI for news – Taboo to even cite! • I believe we can separate them – Practical partial understanding with statistically trained story models. • Plot-fragments: too few categories to be versatile – No internal consistency • Debate over story grammars saw them as purely linguistic (surface discourse) – But texts and fabulas can be modeled independently. 48 The Plot 1. A Word on Models 2. Narrative for News? 3. Related Work 4. The Big Idea 5. Waxing Philosophic 6. Back to News 7. Wrapping Up 49 All The World’s A Stage • Let’s step away from the computer for a while. – Goal is model of storytelling. Tool is computer AI. • Let’s focus on the middle layer (“Story”) and model the ways we can perceive and portray events in the fabula. • An interesting story has: – Characters with whom we identify; – Goals for the characters that we want them to obtain; – Difficulties that the characters confront in pursuing their goals. 50 Let’s just list the wants! • • • • • • • • • • • Life Physical health Mental health To be physically home Provision of a want for family The wants of clan (e.g., ethnic) Freedom of movement Freedom of behavior (e.g., political freedom) Companionship Privacy Etc. 51 A Rhetorical Grammar • • • • • STORY: (STATE, ACTION)+, WORLD STATE: (CHARACTER, GOAL+, ENVIRONMENT)+ CHARACTER: (CHARACTER+ or ) GOAL: WANT+, subgoal(GOAL) WANT: (life) or (physical health) or (provision(WANT, CLAN)) or (protection(WANT)) or (physical journey) or (privacy) or (acquisition(symbol(WANT)))… • CLAN: family, ethnic group, gang 52 A Rhetorical Grammar • The abstract plot of Finding Nemo with the particulars (names, places) as terminals in the parse tree: – State: [Marlin] wants [provision (health (Nemo))] with SUBGOAL of [physical journey (ocean)] – [Nemo] wants [to be physically home] and [companionship] and [freedom of behavior from [Marlin]]. 53 The horoscope effect • Like seeing your life in a horoscope, the model is not “proven” because some story fits. • This is where the number of categories point matters – Like any model, tradeoff between versatility and specificity • Impossible to document the model that can write the full script of Finding Nemo – Too much real world knowledge needed • But we can make this model progressively more detailed until we hit the knowledge barrier – The building blocks of stories that we (or a computer) can combine in novel ways that still work 54 Value Systems • Each character assigns a comparable value to all her wants. • This captures many situations: – Plots: Sacrifice of one want for another • Suicide bombers, whistle-blowers – Characters: Unusual or “evil” value systems • We are surprised at a choice of sacrifice – Stories of total commitment to a hobby, isolated living, etc – Ruthless villains value others’ lives lower than their wants • A character changes from a “bad” to “good” system – Comes to prioritize family over career 55 Clans • Any subset of the population with a common bond forms a clan. – Ethnic group, socio-economic group, special interest group, family, gang, gender… • This captures many situations: – Plots: love across clan boundaries, struggling to join or leave a clan, war between clans, protecting or saving a clan member from danger – Characters: conflicted on a clan boundary, the leader of a clan, going against the clan’s traditions – Environment: the clan in mortal danger 56 Categories of Human Flaws • You internalize your emotions and don't let anyone help you. • You judge and/or stereotype others too quickly based on too little information. • You hold grudges long beyond their useful lifetime. • You fear what you don't know or understand. • You defend other people's ideas at the expense of your own. • You don't know your self-worth. • You get upset with yourself too often and for too trivial matters. • You are incapable of asserting and defending yourself. • You believe you have no flaws. • You believe it is your duty to help others overcome their faults. • Etc. 57 Categories of Personality Contrasts • • • • • • • • • • • • Organized, logical introvert vs. free-associating extrovert Private and closed off vs. connected and generous Type-A (driven) vs. type-B (patient) Highly educated and analytical vs. ignorant and impulsive Looking toward the future vs. longing for the past Cynicism of age vs. romanticism of youth Risk taker vs. play-it-safer Popular (or famous) vs. unpopular Talented vs. hack Talented vs. gifted Pragmatist vs. true believer Etc. 58 We already use story models • We are documenting a mental expectation model • How would you close these stories? – The beautiful princess, locked away in her tower so that she would never know love, called out for the handsome prince. – Harry looked for the keys under the car. • On art vs. entertainment… 59 Hypothetical stack trace • [Harry] does PLOT[ [physical search] for [symbol[ability[freedom of movement]]] ] – Why? – Plot common for WANT[freedom of movement] – [freedom of movement] a common REQUIREMENT for WANT[be at point X] – Therefore, this is subgoal[WANT] where WANT requires Joe at point X • If this is the human trace, can the computer mimic? – Narrative inference, statistically trained – Too much inference needed to parse from text directly. • Car keys or house keys? Car keys are always a symbol? 60 Differences from prior work? • Isolation to only middle (“Story”) layer – More versatile, complete, powerful than plot fragment models like Campbell • Underlies stories in all genres – No conflation of story and linguistic telling of story – Smaller, more abstract and portable model than hard AI – Not practically implementable as a free-text parser • “If person A takes person B’s car with consent, it is a friendly favor; otherwise, it is a theft that illustrates these flaws and gives B a reclamation of property want…” – Not knowledgeable enough for generation 61 So what did we accomplish? • Much work in creating abstract model – As written notes • Model itself may be a research contribution – Formal language of storytelling – Not sure of what community, cognitive science? • Computational story-logic engine in progress – Might apply and specialize to various NLP problems • News • Question-answering • Other specialized generation 62 The Plot 1. 2. 3. 4. A Word on Models Narrative for News? Related Work The Big Idea 5. Waxing Philosophic 6. Back to News 7. Wrapping Up 63 On creativity • Creative writers hate these projects. – – – – “You can’t put creativity in a box!” “How can one list serve every kind of story?” “You’re just finding cliches!” “Making computers creative will destroy a part of humanity!” • Think of this as a dictionary and grammar of stories. – Every English word has been discovered and defined, but we can still write novel sentences. – Billions of stories possible with model; only a few work. – Creativity lies in selection, application to a setting and execution (telling the story well) – This project has vastly improved my own creative writing. • Building with raw wood and concrete instead of preset rooms. 64 On creativity • Every story is formulaic, for sufficiently small formulas. – The best writers read, experience the most. • A a mental sense plots, characters, dynamics that work. • They use this model when they write. – To document this model does not destroy it. • Any more than documenting English language does. • The model is flexible enough to create bad stories, too. – Computers will never be able to tell stories as well as humans. 65 The almighty story • Narrative semantics used everywhere! – Journalists’ creed: “tell a good story” • Convey complex issues with personal stories – Nonfiction authors: “find the story” in the subject • Missions of people involved – Historians: spin compelling stories from facts • Find symbolism, personality flaws and contrasts, etc. – Education: use a story analogy – Law: two competing stories on the same fabula • Each side portrays a different victim – – – – Politics: vilify the other side Medical ethics: understand the patient’s story Coincidences, conspiracy theories, police investigations Oh… and art & entertainment 66 The almighty story • Cognitive science: We even use story form in thinking about and describing our own lives. – We are all in a state of pursuing our wants – Mental health “needs” are the same as our “wants” • Explanations for actions such as joining gangs – Diaries, gossip express selves through story – Marketers appeal to our wants and promise a happy ending 67 Rule of Law • We strongly fear a savage world – – – – Loss of protection from civilization Murder committed with impunity We can be killed for no crime of our own No deterrent of accountability • Any threat to our rule of law is a villain, and any protector is a hero we love and respect. 68 Rule of Law • Backbone of entire genres of fiction: – Western • Sheriffs vs. outlaws in society on cusp of rule of law – Post-apocalyptic • Bandits, protectors of the peace – Thrillers, spy movies • Supervillains want to destroy order and accountability • Serial killers impose fear by killing innocents – Mafia • Subculture with no rule of law – War 69 Example: Rule of Law • Common story in the news – Remote places where there is no rule of law – Interviews with villagers living in fear – Terrorism QuickTime™ and a MPEG-4 Video decompressor are needed to see this picture. • Invoked for persuasion – Aggrandize the heroes who will protect us – Cast the facts to support the story as purely as possible 70 Modeling the world as stories • We spread world views by casting people into narrative roles. – – – – Freedom fighter or terrorist? Unreasonable nut or hero of reform? Underdog or wannabe? Careless doctor or victim of theft? • Allegories and jokes as minimal models – Small stories for abstracting and casting – Powerful, but easily abused 71 Why wax philosophic? • The more we look, the more we will see stories, and the more we will see that how they overlap. • The wider we look now, the better the model will be at finding the core abstractions. • We can’t explain the whole world. – We can always condense and constrain to apply to some field. 72 The Plot 1. 2. 3. 4. 5. A Word on Models Narrative for News? Related Work The Big Idea Waxing Philosophic 6. Back to News 7. Wrapping Up 73 Content selection • Essential step for summarization – Identify the main point of a document – Separate from different types of details • Sentence extraction with IR and other statistical – – – – Inderjeet Mani TF/IDF, word clustering, etc. Radev, Teufel etc. Probabilistic content models (Barzilay, Lee) • Content selection techniques in generation – – – – RST (Mann, Thompson, etc) Discourse modeling (Marcu, Berstein, etc.) Plans (Hovy, Moore, Paris, etc.) Schemas (McKeown) 74 What is a story? Text structure: Pitch, elaboration, background (news) Rising/falling action, climax, coda (fiction) Discourse Narrative semantics: Heroes, villians, goals, obstacles, plot arcs, themes, morals Story Raw timeline: Actions, events, characters Fabula Observed Implied By modeling the top two layers, we might better isolate the real-world events, and know which are the most important for a summary. 75 Best place to look: The Pitch JERUSALEM (CNN) -- The Israeli military has embarked on a wide-ranging raid in Jenin after getting tips that the Palestinian militant group Islamic Jihad was planning to carry out terror attacks against Israel from the northern part of the West Bank, the Israel Defense Forces said Thursday. • Minimal summary of article • Richest source for narrative insight – Security theme, evacuation plot – Can guide further content selection • But… nontrivial to find – Not always the lead sentence 76 Architecture 1. Discourse Modeler – – – 2. Story Modeler – – 3. Categorize each paragraph by function Find the pitch Separate the narrative news from non-narrative Only so many stories can be told From the pitch, find the “narrative backbone” Plan Generator – – Story-specific content selection Story-specific generation plans or schemas 77 Architecture Input Text 1. Summary Discourse Modeler Discourse – Categorize each paragraph by function modeler Existing – Find the pitch Extraction, – Separate the narrative news from non-narrative Content Planning Pitch Lexicalization 2. Story Modeler tools – Only so many storiesStory can be told – From the pitch, findModeler the “narrative backbone” Plan Narrative categories 3. Plan Generator – – Plan Story-specific content selection Story-specific generation plans Generator or schemas 78 Discourse modeler • Reasonably well-defined structure for news articles – Each paragraph serves a purpose – Ordering is important • Categorize each paragraph by function Draft list: • Pitch • Elaboration • Background • Analysis • MOS Reaction • Character Description • Testimonial • Supportive • Dissenting • Expository • Illustrative Anecdote 79 Simple declarative at beginning BAGHDAD, Iraq (CNN) -- Saddam Hussein stepped into an Iraqi court on Thursday Pitch: and entered“New a new chapter in Iraq's history, in hearing preliminary charges against him Development Story of Conflict” that included the gassing of Kurds and the invasion of Kuwait. He listened to seven preliminary charges outlined in his arrest warrant -- the invasion of Kuwait in 1990, suppression of the Kurdish and Shiite uprising in 1991, political killings, religious killings, the killingBackground of a big Kurdish family, and the gassing of the Kurds in Halabja. List with dates These are not the formal charges against Saddam, which will be worked out and detailed in an indictment over the next few months, beginning the investigative phase Analysis of the case. The spectacle, meanwhile, was surprising for a generation of Iraqis who came to fear MOS Reaction the sight of him. Saddam had bags under his eyes and looked gaunt. He had a neatly trimmed beard and was thinner than he appeared in December, when he was seized in a hideout near Character Description Tikrit. • Different story types, include different types of paragraphs, in different orders: 80 Quotations at beginning "Mr. G.," aka chess instructor Douglas Goraleski, stood in front of a large cloth chessboard hanging in the classroom and said to the 25 fifth-graders smiling back at him, "Is it okay to cheat?" Illustrative Anecdote "You have to be a little sneaky," he confirmed, adding that playing chess also requires patience, concentration, planning and strict adherence to the rules. And at risk of having their game forfeited, the students did as they were told during their weekly hour of chess instruction. Oyster is one of hundreds of schools across the country where chess is being taught -not just as an after-school activity for students naturally drawn to the game, but during school hours to youngsters of all abilities. Pitch: “New Trend” Story In several U.S. cities and dozens of countries, the game is being used more than ever as a powerful educational tool. It helps develop skills ranging from critical thinking and time management to sportsmanship and self-esteem, proponents have said. "Everything that happens on a chessboard happens in real life," said Marley Kaplan, president of Chess-in-the-Schools, Testimonial a nonprofit program based in New York that teaches chess to 38,000 inner-city elementary and middle school students. Quotation after pitch •Different stories, different structures, different priorities for summarization 81 Experiment: Methods • Manual analysis to identify discriminatory features • Hand-crafted decision tree – Lexical, syntactic clues – No notion of context • Training/testing data: 350 lead article sentences 82 Training/Testing Data • Labeled among pitch, background, analysis, testimonial 200 180 160 140 • Baseline: to assume pitch = 52% accuracy • Parsed (Collins 2000) 120 100 80 60 40 20 0 Pitch Background Analysis Testimonial 83 Experiment: Algorithm • Check for lexical and syntactic “blacklist” patterns – Last, When, Recently = Background – Question, noun phrase “IS” noun phrase = Analysis – “X, said Y”; “X, Y told Z”; “Y told Z X” = Testimonial • Match against patterns for simple declaratives – No subject and past-tense action = Analysis • Check for more patterns within the action clause – Did, Going = Analysis – Used to = Background • Default = Pitch 84 Experiment: Results Automatic Analysis Background Pitch Testimonial Analysis 52 10 12 0 Background 2 17 2 1 Pitch 30 0 145 1 Testimonial 0 0 10 56 Truth • 80% accuracy • Can recurse on statements within testimonials – Caveat: not yet labeled 85 Next step: Use the pitch • Classify the entire article – – • Narrative news: stories of conflict, new trends, revelations Non-narrative news: analyses, editorials, reviews, etc. If narrative, categorize the “backbone” – – Applicable plots, characters, and themes Statistically trained and classified 86 From “backbone” to summary • Ordering and selection constraints – “Always begin with the pitch” – “Avoid anecdotes” – “Elaboration before Testimonials” • Different emphasis for different stories – “History more important for Justice than Disaster, Testimonials less important” • New sentence generation – Story types imply specific details (e.g., statistics) – Lexicalization rules for expressing story types – Story structure grammar 87 Next steps • Apply other tools to discourse modeler – Markov chains and HMMs for contextual clues – Bayesian networks – Decision tree learners • Collect pitches and pick best narrative distinctions – Experiment with methods for interpretation • Finalize a summarization approach 88 The Plot 1. 2. 3. 4. 5. 6. A Word on Models Narrative for News? Related Work The Big Idea Waxing Philosophic Back to News 7. Wrapping Up 89 Denouement • Model the real world with care • The rabbit hole of story modeling • Practical use for extractive summarization – New categories to aid content selection • Narrative news vs. non-narrative, etc. – Discourse modeling, specialized for news • Find the pitch, find the story – Classification of narrative backbone • Useful, yet practical (not full understanding) 90 Thank you! 91