Reasoning About Actions in Narrative Understanding

Reasoning about actions has been a focus of in terest in Al f rom the beginning and continues to receive a t tent ion. Rut the range of s i tuat ions considered has been rather narrow and falls well short of what is needed for understanding natura l language. Language understanding requires sophisticated reasoning about actions and events and the world 's languages employ a variety of g rammat ica l and lexical devices to construe, direct attention and focus on, and control inferences about actions and events. We implemented a neural ly inspired computa t iona l model that is able to reason about, l inguist ic act ion and event descript ions, such as those found in news stories. The system uses an active. event representation tha t also seems to provide natural and cognit iveIy mot ivated solutions to classical problems in logical theories of reasoning about actions. For logical approaches to reasoning about actions, we suggest, that looking at story understanding sets up fair ly strong desiderata both in terms of the f ine-grained event and action dist inct ions and the kinds of realt ime inferences required. 1 I n t r o d u c t i o n Formal approaches to model reasoning about changing environments have a long t rad i t ion in A l . This research area was in i t ia ted by McCar thy [McCar thy , 19(H)], who claimed that reasoning about actions plays a fundamental role in common sense. T r y i ng to bui ld language understanding programs not only underscores the importance of reasoning about actions, but also suggests that the the set of si tuat ions and the kinds of inferential processes required are richer than has been t rad i t iona l ly studied in formal approaches. Language understanding requires sophisticated reasoning about actions and events. The wor ld 's languages have a variety of grammat ica l and lexical devices to construe, direct attention and focus, and control inferences about, actions and events. Consider the meaning of stumbling in the fo l lowing newspaper headline " Ind ian Government s tumb l ing in imp lement ing L iberal izat ion Plan' ' Clearly, the speaker intends to specify that the l ibera l izat ion plan is experiencing some di f f icul ty. Moreover, the g rammat ica l fo rm is + VP-ing suggests that the di f f icul ty facing the plan is ongoing and the final outcome of the plan is indeterminate. Compare this to the subtle meaning differences w i th g rammat ica l and lexical modif iers on the same root verb such as has stumbled or starting to stumble. Most readers are l ikely to infer after reading this sentence that the government 's l iberal izat ion pol icy is l ikely to fa i l , but this is only a default causal inference that is made in the absence of in format ion to the contrary. Final ly, how does stumble, whose basic meaning is related to spat ial mot ion and obstacles get interpreted in a narrat ive about in ternat ional economic policies? We have implemented a computa t iona l model that is able to reason about act ion and event descriptions f rom discourse fragments such as the one above. The system uses an active event representation that also seems to provide natura l and cognitive!}mot iva ted solutions to classical problems in logical theories of reasoning about, actions. We first present the main features of our representation and show that if provides a computa t iona l model for exist ing formal isms for reasoning about act ions. We then suggest how look ing at story understanding sets up fa i r ly strong desiderata for logical approaches to reasoning about actions both in terms of the finegrained event and act ion dist inct ions and the kinds of realt ime inferences required. 2 T h e A c t i o n M o d e l Our act ion theory comprises of two central components; 1) an e x e c u t i n g representation of actions (called xschemas ) based on extensions to Petr i Nets and 2) a Bel ief Net model of state tha t captures and reasons about, complex dependencies between state variables. 350 COGNITIVE MODELING 2.1 An E x e c u t i n g Semant ics o f Actions We represent actions as a modif ied class of high-level Petri Nets [Reisig, 1985] called x-schemas. The most relevant features of Petr i nets for our purposes are their abi l i ty to model events and states in a d ist r ibuted system and the clean manner in which they capture sequentiality, concurrency and event-based asynchronous control . D e f i n i t i o n 1 . T h e bas ic x s c h e m a : An x s c h e m a consists of places( P ) and Transitions ( 7 ) connected by weighted directed arcs Each arc has weight In­ put Arcs connect Inpu t Places to Transi t ions. Ou tpu t Arcs 7* ) connect Transit ions to Ou tpu t Places. Arcs are typed as enable arcs , inhibitory arcs I, or resource arcs 7v . X-schemas have a well specified realt ime execution se­ mantics where the n e x t s t a t e funct ion is specified by the f i r i n g r u l e . In order to s imulate the dynamic be­ havior of a system, a M a r k i n g (d is t r ibu t ion of t o k e n s in places (depicted as dark circles or numbers)) of the x-schema is changed according to the fo l lowing f i r i n g r u l e . D e f i n i t i o n 2 . E x e c u t i o n S e m a n t i c s o f t h e bas ic x s c h e i n a A t rans i t ion T is said to be e n a b l e d if no inhibitory arc lias a m a r k e d source place a n d all sources of enable arcs are m a r k e d and all input arcs have at, least wpt tokens at their source place, where wpt. is the weight of the arc f rom The f i r i n g o f an e n a b l e d t ransi t ion T , removes wpf tokens f rom the source of each noninh ib i tory , non-enabled input arc V and places wrp tokens in each output place of T . X-schemas cleanly capture sequent ia l ly , concurrency and event-based asynchronous contro l ; w i th our extensions they also model hierarchy, stochasticity and />aranietenzation ( runt ime bindings). Besides typed arcs (Def in i t ion 1), the fo l lowing two extensions to the basic Petri net are designed to allow us to model hierarchical action sets w i th variables and parameters: First., tokens carry in fo rmat ion (i.e. they are ind iv id ­ uated and typed) and transi t ions are augmented w i th predicates which select tokens f rom input places based on the token type, as well as relate the. type of the tokens produced by the t i r ing to the types of tokens removed f rom the input . Second, transit ions are typed. Figure 1 shows the four types of x-schema transi t ions, namely stochastic, durative, instantaneous and hierarchical t ransit ions. An in­ stantaneous t ransi f ion(shown as dark rectangles) fins as soon as it is enabled. A t imed t rans i t ion (shown as rect­ angles) fires after a fixed delay or at an exponential ly d ist r ibuted rate. Hierarchical t ransi t ions (depicted as hexagons), act ivate a subnet., wai t for i ts re turn, or time­ out. Figure 1: Basic types of transit ions. T h e o r e m 1 . [Narayanan, 1997] An j-schema is for­ mally equivalent to bounded High Level Generalized Stochastic Petri Net (HLGSPN). The reachability graph of a marked x-schema is isomorphic to a semi-markwv process. 2.2 A Belief Net Model of States Our representation of states must, be capable of model­ ing causal knowledge and be able to support both belief u p d a t e s and r e v i s i o n s in comput ing the global impact of new observations and evidence both f rom direct ob­ servations and f rom action effects. Our implementat ion of the agent's state uses Belief Networks [Jensen, 1996J. Belief networks consist of a net of variables and a set, of directed l inks. Each variable lias a f ini te set of mu­ tual ly exclusive states. The variables together form a DAG (Directed Acyclic Oraph) . To each variable A wi th parents B\ . . . Bn there is attached a condit ional probabi l i ty table For a Belief Net , the fo l lowing theorem allows us to calculate the j o i n t probabi l i ty P(V) f rom the condi­ t ional probabi l i t ies in the network. T h e o r e m 2 . T h e C h a i n Rule [Jensen, 1996] Let be a BN over I . Then the joint proba­ bility distribution P(U) is the product of all conditional probabilities specified in