Waltz Questioning:Asking the Right Question September 23, 2012 James Pustejovsky David Waltz Symposium Brandeis University
Talk• Personal Connection to Dave• Waltz Questioning Algorithm• Application Instance• Results and Evaluation• 😉
Dave’s Time at Brandeis• Hired in 1984, part-time tenured Professor of CS.• Started teaching AI• Brought students to Brandeis from Illinois: – Tony Maddox, Ph.D.• Kept connection to Illinois: – Jordan Pollack, Ph.D.• Started hiring positions in AI – Rick Alterman – Me• Helped found Ph.D. program at Brandeis• Instrumental in Planning and Funding for Volen Center• Left in 1993• Left a culture of AI: we hired Jordan Pollack
Personal Connection to Dave• Ph.D. Umass Amherst Linguistics• Hired from UMass Amherst Postdoc in COINS with David McDonald• Dave helped hire me at Brandeis – Thinking Machines was a scary option• Dave wanted to establish AI at Brandeis while still working at TMI – a little security to go with the drama of a startup• Exposed me to Example-based reasoning (EBR, MBR)
Challenge as an Opportunity• Simple rule-based NLP was failing to handle real language data.• Memory-based reasoning can be used to solve: – Syntactic parsing issues (attachment, rule choice) – Word sense disambiguation – The frontier: story understanding, inference
Waltz QuestioningYou just presented the conclusions of years of work, culminatingin a solution, G. Waltz-Agent1. Assume G is true and interesting.2. Introduce a new problem, H, that is (arguably) more interesting than G.3. Establish the belief that H subsumes G, and4. Maximize confidence that you can find connection between G and H.5. Solve for H.
Verbal Event Meaning• Mary knows calculus. (state)• John jogged in the park. (activity)• Bill found a dollar. (achievement)• Mary build a house. (accomplishment)• Lexical Aspect and Aktionsarten
Changing Meaning in Context• It was then that I knew he did it. – (achievement)• John jogged to the store. – (accomplishment)• Bill found a house in Cambridge in 3 months. – (accomplishment)• Mary build houses for years. – (activity)• Aspect Calculus Event Structure Theory = G
Waltz Questioning: 1/31. Assume “aspect calculus” is G.2. Introduce a new problem, H, that is (arguably) more interesting than G: H = General Lexical ambiguity
The Problem of Lexical Ambiguity• Homonymy: unrelated senses of a word: – We sat on the bank of the river. – The bank lowered its interest rate. – Julie is the chair of the committee. – Put four chairs at each table in the room.• Solution: Memory-based Parsing
Polysemy• Conceptually related senses of a word: – Pick up the course book at the university store. – I don’t agree with his recent book at all. – We painted the door blue with white trim. – John walked through the door.
How Many Meanings?• Good – I need a good car (cities/soccer/racing/…) – good meal – good knife• Noisy – noisy car – noisy room• Fast – fast typist – fast train – fast highway
Two Types of Polysemy• Inherent polysemy: where multiple interpretations of anexpression are available by virtue of the semanticsinherent in the expression itself.• Selectional polysemy: where any novel interpretation of anexpression is available due to contextual influences,namely, the type of the selecting expression. 1. a. John bought the new Obama book. b. John doesn’t agree with the new Obama book. (inherent) 2. a. Mary left after her cigarette. (selectional) b. I’ll call you after my coffee.
Waltz Questioning 2/31. Assume “aspect calculus” is G.2. Introduce a new problem, H, lexical ambiguity.3. Establish the belief that H subsumes G.4. Maximize confidence that you can find connection between G and H. – Connections through enriching the representation: • Lexical qualia structure • Coercion and co-composition rules
Waltz Questioning: 3/31. Assume “aspect calculus” is G.2. Introduce a new problem, H, lexical ambiguity.3. Establish the belief that H subsumes G.4. Maximize confidence that you can find connection between G and H. – Connections: • Lexical qualia structure • Coercion and co-composition rules5. Solve for H. – (Generative Lexicon Theory)
Challenging the Established Doctrine• Principle of Compositionality: – The meaning of a complex expression is determined by its structure and the meanings of its constituents.• What is encoded as constituent meaning?• What is encoded as the structure?• Data, data, data!!!!!! – Let the data guide you in your modeling
Let the Data Guide You• Language models require: – Thousands of KR axioms per word – Millions of contexts per phrase/sentence – Thousands of semantic and pragmatic features influencing interpretation.
Waltz Questioning 4/3• Dave’s legacy for NLP research:• Solve for a better H with a better description of the problem.