Semantics from Narrative: State of the Art and Future Prospects

Transcription

1 SLDS 2009 Semantics from Narrative: State of the Art and Future Prospects Fionn Murtagh Science Foundation Ireland, and Royal Holloway, University of London

2 Challenges Addressed Great masses of data, textual and otherwise, need to be exploited - decisions need to be made Correspondence Analysis handles multivariate numerical and symbolic data with ease Structures and interrelationships evolve in time We must consider complex web of relationships We need to address all these issues from data and data flows

3 We will look at how this works, using Casablanca film script Then return to the data mining approach used

4 Interaction and decision making - Casablanca (1942) Script half completed when production began Dialog for some scenes written while shooting in progress My work on Casablanca is joint with Adam Ganz and Stewart McKie, Dept of Media Arts, RHUL

5 Casablanca Based on unpublished 1940 screenplay by Murray Burnett and Joan Alison, Everybody comes to Rick s Script by JJ Epstein, PG Epstein and H Koch Film directed by Michael Curtiz and produced by Hal B Wallis and Jack L Warner Shot by Warner Bros between May and August 1942

6 Casablanca script has 77 successive scenes 6710 words in these scenes We use (later) all words, ignoring punctuation and taking all in lower case We analyze frequencies of occurrence of words in scenes, so the input is a matrix crossing scenes by words

7 Illustrative example: Casablanca (1942) A first data set had 77 successive scenes crossed by attributes - Int, Ext, Day, Night, Rick, Ilsa, Renault, Strasser, Laszlo, Other (ie minor character), and 29 locations Many locations were met with just once; Rick s Café was the location of 36 scenes (We did not distinguish between Main room, Office, Balcony, etc)

8 12 attributes displayed; 77 scenes displayed as dots Factor 2, 15% of inertia Day NotRicks Other Ilsa Strasser Renault Laszlo Rick Int Night RicksCafe Ext Factor 1, 34% of inertia Approx = 49% of all information displayed Can study interrelationships between characters, other attributes, scenes, etc

9 Some underlying principles 1/2 Cross-tabulation data, scenes x attributes Embedding scenes, attributes in a metric space We are probing the geometry of information

10 Triangular inequality holds for metrics Vertical x y z Example: Euclidean or as the crow flies distance d(x, z) d(x, y)+d(y, z) Horizontal

11 Some underlying principles 2/2 Axes are the principal axes of momentum Identical principles used as in classical mechanics Scenes are located as weighted averages of all associated attributes; and vice versa

12 Christiaan Huyghens ( ) Huyghens theorem relates to decomposition of inertia of a cloud of points This is the basis of Correspondence Analysis

13 And now: the topology of information Euclidean embedding provides a very good starting point to look at hierarchical relationships An innovation in this work: the hierarchy takes sequence, eg timeline, into account This captures novelty, anomaly, change

14 Property Property 1

15 Property Property 1

16 Property Property 1

17 Property Property 1

18 Property Property 1

19 Property Property 1

20 Property Property 1

21 Property Property 1

22 Property Property 1

23 Property Isosceles triangle: approx equal long sides Property 1

24 Strong triangular inequality, or ultrametric inequality, holds for tree distances d(x, z) Height max{d(x, y),d(y, z)} d(x, z) =35 x y z d(x, y) =35 d(y, z) =10 Closest common ancestor distance is an ultrametric

25 77 scenes clustered Shows up 9 to 10, and progressing from 39, to 40 and 41, as major changes

26 A look under the hood Correspondence analysis supports following: analysis of multivariate, mixed numerical/ symbolic data web of interrelationships evolution of relationships over time

27 Correspondence Analysis is A Tale of Three Metrics - Chi squared metric - appropriate for profiles of frequencies of occurrence - Euclidean metric, for visualization, and for static context - Ultrametric, for hierarchic relations and for dynamic context

28

29 Analysis of semantics: 1 Context - the collection of all interrelationships Euclidean distance makes a lot of sense when the population is homogeneous All interrelationships together provide context, relativities - and meaning

30 Analysis of semantics: 2 Hierarchy tracks anomaly and change Euclidean distance makes a lot of sense when the population is homogeneous Ultrametric distance makes a lot of sense when the observables are heterogeneous, discontinuous Latter is especially useful for determining: anomalous, atypical, innovative cases

31 Back to a deeper look at Casablanca We have taken comprehensive but qualitative discussion by McKee and sought qualitative and algorithmic implementation

32 McKee, Methuen, 1999 Casablance is based on a range of miniplots McKee: its composition is virtually perfect Text is the sensory surface of the underlying semantics

33 Analysis of Casablanca s Mid-Act Climax, Scene 43 subdivided into 11 beats (subscenes) McKee divides this scene, relating to Ilsa and Rick seeking black market exit visas, into 11 beats Beat 1 is Rick finding Ilsa in the market Beats 2, 3, 4 are rejections of him by Ilsa Beats 5, 6 express rapprochement by both Beat 7 is guilt-tripping by each in turn Beat 8 is a jump in content: Ilsa says she will leave Casablanca soon In beat 9, Rick calls her a coward, and Ilsa calls him a fool In beat 10, Rick propositions her In beat 11, the climax, all goes to rack and ruin: Ilsa says she was married to Laszlo all along Rick is stunned

34 MPrincipal plane of 11 beats in scene 43 m-10 M20 m-20 M-20 M-15 M-10 M-05 M00 M05 M10 M15 MFactor 1, 2, 126% 122% of inertia Principal plane of 11 beats in scene Factor 2, 122% of inertia Factor 1, 126% of inertia 210 words used in these 11 beats or subscenes

35 MPrincipal plane of 11 beats in scene 43 m-10 M20 m-20 M-20 M-15 M-10 M-05 M00 M05 M10 M15 MFactor 1, 2, 126% 122% of inertia Principal plane of 11 beats in scene Factor 2, 122% of inertia Factor 1, 126% of inertia

36 MPrincipal plane of 11 beats in scene 43 m-10 M20 m-20 M-20 M-15 M-10 M-05 M00 M05 M10 M15 MFactor 1, 2, 126% 122% of inertia Principal plane of 11 beats in scene 43 Factor 2, 122% of inertia Repulsion 00 Attraction Beat 8: Lisa to leave Casablanca! Beat 11: Lisa married to Laszlo all along! Factor 1, 126% of inertia

37 McKee s guidelines applied to Scene 43 Lengths of beat get shorter leading up to climax: word counts of final five beats in scene 43 are: The planar representation seen accounts for approx = 248% of the inertia, and hence the information We will look at the evolution of this scene using hierarchical clustering - but based on the relative orientations, or correlations with factors

38 MHierarchical clustering of 11 beats, using their orientations m00 M00 M02 M04 M06 M08 M10 1M 2M 3M 4M 5M 6M 7M 8M 9M Hierarchical clustering of 11 beats, using their orientations Full dimensionality analysis Note caesura in moving from beat 7 to 8, and back to 9 Less so in moving from 4 to 5 but still quite pronounced

39 Style analysis of scene 43 based on McKee Monte Carlo tested against 999 uniformly randomized sets of the beats In the great majority of cases (against 83% and more of the randomized alternatives) we find the style in scene 43 to be characterized by: small variability of movement from one beat to the next greater tempo of beats high mean rhythm

40 Our way of analyzing semantics We discern story semantics arising out of the orientation of narrative This is based on the web of interrelationships We examined caesuras and breakpoints in the flow of narrative

41 Work of J Eliashberg, Wharton, U Penn Use features characterizing scripts to predict boxoffice success

42 Having tracked various aspects of semantics in filmscript Can we apply similar principles to the research literature? Objective 1: to evaluate funding proposals, and allocation of funding Objective 2: to evaluate trends and evolution in fields and subfields of research For planning and resource allocation Personal, institutional, national, disciplinebased

43 Take 5 articles on neuro-imaging studies of visual awareness and cognitive alternatives in early blind humans

44 Methodology Consider sections: resp in the five articles there are 7, 6, 6, 6, 7 sections Consider paragraphs within sections: resp in the five articles there are: 51, 38, 60, 23, 24 We analyze sections x words in each article Words are 2 or more characters in length Numbers of words (and unique words) in the five articles: 8067 (1534), 6776 (1408), 8247 (1534), 3891 (999) and 5167 (1255) We also used for each article: abstract, bibliography

45 Issues assessed at individual article level Which sections contribute most strongly to the factors Which terms, including cited works, contribute or are correlated most with factors - hence which are most important or most salient Which technical terms are most

46 We find abstracts to be good proxies for the articles Abstracts projected into the plane Bold italics: 5 articles Factor 2, 22% of inertia Factor 1, 36% of inertia

47 And we find bibliographies to be good proxies also - Possible implications for bibliometrics Reference sections projected into the plane Bold italic: 5 articles Factor 2, 22% of inertia Factor 1, 36% of inertia

48 Conclusions Here bibliography (in each of the five articles) was the set of all cited references, including author names, titles, journal titles and other details Caveat: citing cultures differ across disciplines Nonetheless: Perhaps complementing networks of citing articles, as commonly used in bibliometrics can sematic analysis based on Correspondence Analysis - as pursued here - better capture the narrative and hence trends?

49