Representing and Recognizing Point of View

A representation of ideological point of view is articulated and a method for detecting the point(s) of view expressed in a news story is described. A version of the method, actor-role analysis, is encoded in a computer program, SpinDoetor, which can automatically detect the point(s) of view represented in some news tories. SpinDoctor is a computer program designed to detect ideological point of viewlin news stories. To detect point of view, SpinDoctor implements a critical reading strategy called actor-role analysis (Sack, 1994a; Sack, 1994b). Actor-role analysis was developed around the following observation: one means of detecting point of view is to examine how certain people, who appear again and again in the news (i.e., news actors), are described or portrayed (i.e., are assigned roles). Thus, for example, if one is given a news story which mentions Oliver North -(in)famous for his role in the Iran-Contra affair and recent senatorial campaign -and the story assigns North the role of patriot (via the use of certain adjectives and verbs), one can quite certain that the point of view expressed in the story is significantly to the right (in the spectrum of US politics) than that expressed by another news story which assigns North the role of villain or criminal. Two aspects of actor-role analysis, as it is implemented in the SpinDoctor system, might be of especial interest to researchers concerned with textual analysis: (1) A representation for ideological point of view: Although practically all, contemporary, AI systems for NLP are capable of finding actors and roles in texts to fill in scripts, frames, or templates (e.g., Jacobs and Ran, 1993), none of these systems assign any political significance to the pairing of certain actors with certain roles (e.g., North as patriot versus North as criminal). contrast, I maintain that sets of actor-role pairs are an interesting and implementable representation for differing ideological points of view. The proposed actor-role representation of ideological point of view accords with some recent work by Lakoff (1991) and generalizes and improves upon previous AI work on representation of ideology (e.g., Abelson and Carroll, 1966; Carbonell, 1978). (2) An algorithm for anaphoric resolution: Actor-role analysis incorporates a new anaphoric resolution algorithm. It is shown how careful attention to actor-role pairings assists in the resolution of anaphoric reference. By noting, for example, that an instance of the pronoun "he" is cast in the role of victim and that, earlier in the same story, "Lieut. Rodriguez" is also cast in the role of 1 Ideological point of view characterizes the political slant of an entire story; it is different from psychological point of view (e.g., as it is used by Wiebe, 1994) which characterizes the source of given sentence or statement contained within a story. victim, SpinDoctor postulates the resolution of the instance of the pronoun "he" to the proper name "Lieut. Rodriguez." SpinDoctor uses three main data structures: (1) Actors are collections of noun phrases and pronouns hierarchically arranged into groups (e.g., "President Cristiani" is declared to be a member of actor group "government officials" which, in turn, is declared to be a part of a larger group called simply the "government" that also contains "Army officers" and others. (2) Roles are packets of verbs, adjectives, and adverbs and some associated syntactic constraints. Thus, for example, the role criminal is associated with the subject of the verb "kidnap" while the role victim is associated with the object of the verb. (3) Points of View are sets of actor-role pairings. Thus, for example, to find a news story from the late-80s/early-90s from El Salvador written from the government’s point of view, the machine looks for stories in which the government plays the role of the source of the story, the government describes its violent actions as legitimate military actions (i.e., assigns itself a military role), while, it assigns the guerrillas the role of terrorist by labeling the guerrillas’ violent actions as criminal actions. The user of the system fills the database with actor, role, and point of view definitions. However, currently, the hand-built role definitions are being replaced with definitions fromRoget’s Thesaurus and Wordnet, an on-line thesaurusci’eated by GeorgeMiller’s group at Princeton University.The analysis process followed by SpinDoctor is asfollows: (1) Given a news story SpinDoctor finds whichnounphrases (i.e., actors) play whichroles in the story. (2)Then, it determines which actors are of a similar group orare identical to other actors. To do this it does both (a)anaphoric resolution (e.g., determining who or what isbeing referenced whenpronounslike "he," "she," and "it"are used); and, (b) actor grouping(e.g., inferring that army spokesman,""Lieut. Rodriguez," and "Gen. Bustillo"are all a part of the samegroup(e.g., the "armedforces").(3) It the constructs a profile of the analyzed story whichdescribes howoften (within the given story) top-levelactors were assigned different roles. (4) Using weightedactor-role bindings allows it to distinguish, for example,stories in which the governmentis cited as a source onceand the guerrillas are cited ten times fromstories in whichthe guerrillas are cited as a source less often than thegovernment. Finally, it matches the weighted actor-rolebindings against the point of view definitions, determineswhich point(s) of view match(es) best and outputsnameof the one (or more) best-matching point(s) of view.References̄ Sack, W. (1994a) Onthe Computationof Point of View,Proceedingsof the NationalConferenceof Artificial Intelligence(AAAI94), July 31-August4, 1994, Seattle, WA.̄ Sack, W.(1994b)Actor-RoleAnalysis: Ideology, Point of Viewand the News(Teeh. Report 94-005 and MSThesis) Cambridge,MA:MITMediaLaboratory. 152From: AAAI Technical Report FS-95-03. Compilation copyright © 1995, AAAI (www.aaai.org). All rights reserved.