150 likes | 163 Views
Question Answering with deep reasoning. Chitta Baral, Arizona State U. Michael Gelfond, Texas Tech U. Richard Scherl, Monmouth Univ. Text repository. Processed Text. Domain Knowledge Module 1. Domain Knowledge Module N. Question. Reasoning Module 1. Reasoning Module k.
E N D
Question Answering with deep reasoning Chitta Baral, Arizona State U. Michael Gelfond, Texas Tech U. Richard Scherl, Monmouth Univ.
Text repository Processed Text Domain Knowledge Module 1 Domain Knowledge Module N Question Reasoning Module 1 Reasoning Module k Answer, possibly with explanation
Core efforts: KR • Domain Knowledge Modules • Travel Module -- various • Intentions [10:45-11:15] AAAI’05 • Generalization : a theory of activities [2:15 – 2:45] • Support structure • Fundamental extension of AnsProlog • Consistency restoring (CR-Prolog) -- Sep 04 ASU • Probabilistic reasoning (P-log) -- Tampa • Modules • AnsProlog Modules -- Feb 05 ASU • KR Modules (AnsProlog + Prolog + CLP) [11:15-11:45] • Interfaces and methodologies • AnsProlog GUI -- Feb 05 ASU • Knowledge Modeling Language -- Feb 05 ASU • Further development of action languages – ongoing AAAI’05 (continuous actions, triggers, event ordering, etc.)
Core efforts: Reasoning • Reasoning Modules • Prediction, planning, explanation -- Baltimore • Detecting lies -- TTU thesis • Trying your best -- Feb’05 ASU • Counterfactuals -- ongoing • Reasoning with incomplete information [11:15-11:45] AAAI’05 • Explanation generation modules • Explaining an answer set -- ongoing
Bridge-building efforts – towards an integrated system • Logical Form to extracted facts [9:30-10:30] • English to Link grammar output to extracted facts [1:45-2:15] • Solving Puzzles in English (Bob Leaman’s class project) • An end-to-end system with travel module (Matt Hunsaker’s class project) • Collaborative Curation -- Feb 05 ASU • AQUAINT web site
Today’s Schedule – 5/11/05 • 9:00 - 9:20 Overview (Chitta) • 9:30 - 10:30 From Logic Forms to ASP query answering (Marcello) • 10:45 - 11:15 Reasoning about intentions (Michael) • 11:15 - 11:45 New reasoning methods (combining ASP and CSP, approximation algorithm for reasoning with incomplete information) (Michael; Ricardo) • 12:00 - 1:30 Lunch • 1:45 - 2:15 Link grammar based effort at ASU (Luis) • 2:15 – 2:45 From travel modules to a theory of activity -- initial thoughts (Chitta)
Theory of Activity: A trip • A trip has many participants. • People can join and leave the trip. • There is a start and an end of a trip. • A trip may be interrupted resulting in the trip being paused or canceled. • A paused trip might be canceled or continued. • A trip has a schedule which is often a sequence of actions intended to be performed at particular times. • Usually the schedule is followed, and most often if a particular item in the schedule is not possible then it is delayed until it becomes possible. • Some time particular items in the schedule may be canceled.
Example Qs with respect to trips • Who are part of a trip at a particular time? • What is the status of a trip at a particular time? • What are the value of fluents at a particular time? • What actions of the trip happened at what time?
Activities • Other examples: • insurrections, terrorism, games, elections, concerts, money-laundering, smuggling, strikes, tournaments, demonstrations, wars, etc. • Activities in the virtual world include work-flows. • In most of these activities the actions that are intended to happen are not necessarily scheduled as a sequence.
QA with respect to an activity narrative • Who are the participants of an activity at a particular time? • What is the value of a fluent at a particular time? • What is the status of an activity at a particular time? • Based on what we know what is a viable schedule of an activity? • What is the most likely schedule? • When did a particular action (part of the activity) happen? • Give the interval when a particular action is likely to happen? • How to prevent things from going wrong? • How to disrupt an activity?
Knowledge Modeling – objects and properties • Trip • basic static attributes – • name: a string of characters • stops: an ordered list of locations • % can be represented as a set of pairs • % { (0, locn_0), (1, locn_1) ..., (n, locn_n) } • % note : locations may be repeated in the list • % for trips that include multiple visit • % to the same location • derived static attributes – • origin: location • destination: location • intermediate_stops: set of locations • number_of_planned_stops: integer
Objects and properties -- Cont. • basic fluent attributes: • position: location union {en_route} • %encodes current location • last_planned_stop_number: integer • participants: set of persons • derived fluent attributes: • next_planned_stop: location • directly_associated actions: {depart, stop, embark, disembark} • directly_associated actions: {embark, disembark} • Person • basic static attributes – • name: a string of characters • position: location • Location • basic static attributes – • name: a string of characters
Actions • name: a string of characters • parameters: set of property classes • executability conditions: fluent formulas • conditional effects: list of pairs (phi, f) where phi is a fluent formula, and f is a fluent literal • actor: • triggers: list of pairs (phi, a) where phi is a fluent formula, and a is an action • depart is_an action • name: depart • parameter: {trip} • executability conditions: trip.position \neq enroute, trip.number_of_planned_stops < trip.last_planned_stop_number • conditional effects: { (T, trip.position = en_route) }
Actions (cont.) • stop is_an action • name: stop • parameter: {trip.location} • executability condition: trip.position = enroute • conditional effects: { (T, trip.position = trip.location) } • embark is_an action • name: embark • parameter: {trip, person} • actor: person • executability condition: trip.participant does_not_include person, trip.position = person.position, trip.position \neq en_route • conditional effects: { (T, trip.participant includes person) } • disembark is_an action • name: disembark • parameter: {trip, person} • actor: person • executability condition: trip.participant includes person, trip.position \neq en_route • conditional effects: { (T, trip.participant does_not_include person) }
Compound actions and constraints • go_on is_a compound_action • name: go_on • component actions: {embark, disembark} • CONSTRAINTS • % connect various properties. • % for example: The position of a trip and a participant of that trip are always the same.