• Nu S-Au Găsit Rezultate

•  The backbone of the story.

N/A
N/A
Protected

Academic year: 2022

Share "•  The backbone of the story. "

Copied!
32
0
0

Text complet

(1)

The Time Yards Model:

Towards Reshuffling Puzzled Fabulas in Texts

Dan Cristea

“Alexandru Ioan Cuza” University of Iași, Faculty of Computer Science Romanian Academy Iași branch, InsKtute for Computer Science

[email protected]

(2)

Fabula?...

•  The backbone of the story.

•  Every story is a menKon of a fabula, i.e., a sequence of chronologically ordered and logically connected events involving one or more actors.

J. Bruner, “Actual minds, possible worlds”. Cambridge, MASS., Harvard University Press, 1986.

M. Bal, “Narratology: introducKon to the theory of narraKve”. Trans. ChrisKne van Boheemen. 2nd ed.

Toronto: U of Toronto, 1997.

(3)

Are fabulas puzzled?

Example 1

•  In the kitchen, wai-ng for the water to boil, Margaret wondered how was it that she

jumped so naturally to embrace Adam. She could not realize what had made her feel so happy to see him again.

Tash Aw: Map of the invisible world

RAAI-2017, 18th June, Univ. Bucharest

(4)

Example 1: segmentaKon

•  [In the kitchen, wai-ng for the water to boil, Margaret wondered how was it that]

•  [she jumped so naturally to embrace Adam.]

•  [She could not realise what had made her]

•  [feel so happy to see him again.]

Margaret: in the kitchen, wondering smth

Margaret & Adam: somewhere, Margaret embraces Adam

Margaret: in the kitchen, not realising smth

Margaret & Adam: somewhere, Margaret feels happy

(5)

Example 1: our percepKon of Kme

Margaret Margaret

&Adam

RAAI-2017, 18th June, Univ. Bucharest

(6)

Puzzled fabula Example 2

•  Between the two rounds of preparing tomato sauce and a quick chat on Facebook, I

remembered that I put Teodor Baconschi's book “Facebook. Factory narcissism”,

published this year by Humanitas, among the books labelled "to necessarily read”. I could

not stop myself from reading it to the last line.

(7)

Example 2: segmentaKon

•  [Between the two rounds of preparing tomato sauce and a quick chat on Facebook, I remembered that ]

•  [I put Teodor Baconschi's book “Facebook. Factory narcissism” … among the books labelled "to

necessarily read”.]

•  [published this year by Humanitas,]

•  [I could not stop myself from reading it to the last line.]

RAAI-2017, 18th June, Univ. Bucharest

myself: somewhere1, cooking &

chaang, remembering

myself: somewhere2, cooking &

chaKng

Humanitas: somewhere3, publishing

myself: somewhere2, reading

(8)

Example 2: our percepKon of Kme

Humanitas

myself

(9)

Goals

•  A computaKonal representaKon of narraKve phenomena occurring in texts, with a special emphasis on:

– Kme development of narraKves, involvement of characters,

– changes of Kme seangs: flashbacks (remembers) and flashforwards (supposiKons, dreams, hopes),

– transiKons from one plot (theme) to another,

– perspecKves of different narrators on the same story

RAAI-2017, 18th June, Univ. Bucharest

(10)

Main ideas in this model

•  Story developments (fabulas, happenings, plots) glue together characters over Kme

•  Time is not linear in texts:

– significant Kme points: commute, rupture, join and split

•  VisualisaKon: threads that interlace and separate

– much like railways do in complex rail yards: -me yards

•  Propose notaKons for these representaKons

•  ExploitaKon?

(11)

Roots…

•  Temporal reasoning (and temporal logic: Allen, etc.)

•  Much interest:

– informaKon extracKon – quesKon answering

– textual entailment

– deciphering semanKc content of texts

See: ACL workshop on Spa-al and Temporal Reasoning (2001), LREC workshop on Annota-on Standards for Temporal Informa-on in Natural Language (2002)

RAAI-2017, 18th June, Univ. Bucharest

(12)

Related work

–  AnnotaKon convenKons/standards to cope with Kme as menKoned in language and text

•  TimeML

•  TIMEX3 – explicit temporal expressions (Kmes, dates, duraKons, etc.)

•  SIGNAL – funcKon words that indicate how temporal objects are to be related to each other (e.g on, during, when, if, etc.)

•  TLINK, SLINK, ALINK – temporal relaKonship holding between events, aspectual events, events and Kme expressions or events and signals

•  EVENT – event notaKon

–  TARSQI Toolkit capable to recognise:

•  temporal expressions (TimeEx)

•  events

•  relaKons between them

•  James Pustejovsky et al. (2003). TimeML: Robust SpecificaKon of Event and Temporal Expressions in Text, AAAI Technical Report SS-03-07

•  Verhagen & Pustejovsky: Temporal Processing with the TARSQI Toolkit, Coling 2008

(13)

We are interested in connected sequences of events

•  In NLP, on news

– TimeLine: a representaKon of events which are

chronologically ordered, mainly specific to an enKty (a character or parKcipant in an acKon, a geographical place or region)

– StoryLine: groups of interacKng TimeLines, or mergers of two or more TimeLines where the same characters or enKKes are taking part in the acKon. The text

structure obtained is not related to the flow of Kme.

Chambers and Jurafsky (2008). Unsupervised Learning of NarraKve Event Chains. In ACL.

Laparra, Aldabe and Rigau (2015). From TimeLines to StoryLines:

A preliminary proposal for evaluaKng narraKves. In ACL-IJCNLP.

(14)

MoKvaKon and goal

•  Large texts are different then news

– news: sequences of events, opinions, commentaries – novels: events are seldom presented in non-sequenKal

order, same events are interpreted from different perspecKves

•  Characters parKcipate in events

•  Each has her/his own trajectory/desKny/evoluKon

•  Although characters’ trajectories intersect in Kme,

one would like to disKnguish/separate them

(15)

Time Segments

•  A conKguous piece of text characterised by unity of:

– actors (who is involved)

– space (physical or virtual), also as a sequence of locaKons (like in journeys)

– Kme (interval) – type

•  NAR – typical narraKons; story Kme flows constantly ahead

•  SUP – supposiKons, speculaKons, dreams, hopes

•  GEN – general knowledge; no Kme involved

– perspecKve (who is the teller)

RAAI-2017, 18th June, Univ. Bucharest

(16)

Time Segments

•  A TS groups a sequence of events of one character or a stable group of characters over a period of

story Kme, which is uninterruptedly told in a span of text.

– in a TS, the story Kme evolves linearly or not at all – one TS can be narrated by only one relator

•  A book is made up of a sequence of Time Segments

(17)

Time Tracks

•  Out of TSs, Time Tracks (TTs) can be assembled.

•  A TT is a sequence of TSs filtered and ordered by two condiKons: one or more actors that connect and ascendant Kme.

•  TTs put in evidence trajectories of characters or groups in story developments (desKnies)

•  Space, most owen, change over Kme in TTs

•  A TT can be narrated by one or more relators.

RAAI-2017, 18th June, Univ. Bucharest

(18)

Unity of space versus connecKvity

•  To be in the same TT, a group of characters should be consciously connected within the same space:

– two friends that meet and talk on the street – a family sharing the same house

– colleagues in school, etc.

– two people having a conversaKon

•  face to face => same physical space

•  through a telephone line, etc. => same virtual space

(19)

Margaret in the kitchen…

•  TS1:[In the kitchen, wai-ng for the water to boil, Margaret wondered how was it that]

•  TS2:[she jumped so naturally to embrace Adam.]

•  TS3:[She could not realize what had made]

•  TS4:[her feel so happy to see him again...]”

RAAI-2017, 18th June, Univ. Bucharest

(20)

TTs and TSs

in Margaret&Adam’s episode

commute points

(21)

The Time Yard

•  A Time Yard (TY) is a representaKon (resembling a diagram) of the characters’ desKnies in a text, unfolded on the story Kme axis

•  A TY is made out of all TTs of the characters in the book

RAAI-2017, 18th June, Univ. Bucharest

(22)

Border points of TTs

•  Time tracks can be bordered by:

– start points – where they start, – end points – where they finish,

– split points – where one Kme track splits into two separate Kme tracks,

– join points – in which two different Kme tracks meat

(23)

Split point

•  TS1:[Up to climb into the truck and disappear aPer the tarpaulin fall down,]

•  TS2:[he followed them slowly, with his clumsy

gait.]

RAAI-2017, 18th June, Univ. Bucharest

(24)

Join point

•  TS1:[It was only when she came to the doorway that she realized]

•  TS2:[someone was there, lying on the front steps. He was a boy, a teenager, almost

crouched in the womb.]

(25)

Tash Aw’s “Map of the invisible world”

A bird’s eye Kme yard

RAAI-2017, 18th June, Univ. Bucharest

(26)

An XML notaKon

– in addiKon to those of TimeML

•  AutomaKc segmentaKon of the text =>

Actors, Kme segments and temporal relaKons:

<ACTOR ID ALIAS-LIST>…</ACTOR>

<TS ID NAME ACTORS TYPE LOC PER>…</TS>

<TREL ID FROM REL TO/>

•  A selecKon process =>

Time tracks:

<TT ID NAME ACTORS TS-LIST LEFT-EP RIGHT-EP/>

–  LEFT-EP: “START” | ID of a SPLIT end-point –  RIGHT-EP: “STOP” | ID a JOIN end-point

(27)

Reading the TY diagrams

•  Tracking characters’ desKnies (developments)

=> in bold in the figure: Adam’s trajectory

=> a mere selecKon operaKon in the set of TS elements

=> produce focused summaries

=> compare different opinions on the same developments…

RAAI-2017, 18th June, Univ. Bucharest

(28)

Reading the TY diagrams

•  When a book narrates about the first 16 years of Marys’ life, before moving from the house of her

parents in another town to follow the last two years of college, we find out on 100 pages many details:

•  each day Mary went to school, in the holidays she went to her grandparents house in the country side, and she enjoyed

going so many -mes to the cinema, mee-ng friends etc…

•  However, a summary should say:

–  Mary lived with her family un-l the age of 16…

=> puKng the text further away from the eyes, a

Kme track could be summarised in mulKple ways

(29)

Towards a logic of Time Tracks

If:

TT(p)↓Kme=[t1,t2] = select from TT(p) only those developments happening in the interval [t1, t2] (an ordered set of sub-TTs) TT(p)↓Kme = the Kme interval elapsed between the first event and the last event menKoned in TT(p)

TT(p)↓loc = set of locaKons of TT(p)

Then:

we say that a set of characters P={p1, …, pk} share the same space s during a Kme interval [t1,t2] if the set {TT(pj)}, with j ɛ {1,..., k} is such that s ɛ TT(pj)↓Kme=[t1,t2]loc for each j ɛ {1,..., k} and ∩TT(pj)↓loc=sKme [t1,t2]

RAAI-2017, 18th June, Univ. Bucharest

(30)

Reasoning about texts with TYM

If:

TT(p)↓dur = the duraKon of TT(p)

cood(s) = coordinates of the locaKon s S(C) = area of a country C

Then:

We say that the set of characters P={p1, …, pk} are compatriots in the country C if it is true that pj in P, s ɛ TT(pj)↓loc and

coord(s) ɛ S(C): ∑(TT(pj)↓loc=sdur) » ∑(TT(pj)↓loc=ldur), l ɛ TT(pj)↓loc but coord(l) S(C).

This is based on the following definiKon of someone living in a country: the duraKon of living in that country is much higher than that of living outside it.

(31)

Just imagine…

•  … that there is a technology that records the TT of persons with high impact in a country

•  then...

⇒ pages of history can be automaKcally wriƒen, with a proper intelligent summarisaKon agent

RAAI-2017, 18th June, Univ. Bucharest

(32)

Acknowledgements

•  To Humanitas Publishing House, for the Romanian version of Tash Aw’s novel Map of the invisible world in the translaKon of Florica Sincu

•  To the Wylie Agency, UK, for agreeing to use for reseach the original English version of Tash Aw’s novel

•  To my classes of first year master students in ComputaKonal

LinguisKcs, in both series 2015-2016 and 2016-2017, for lively evening debates during seminar hours that helped me refine details of the

model and establish annotaKon convenKons and for their parKcipaKon in group annotaKon experiments

•  To my PhD student Andreea Macovei, for annotaKng the Romanian version of the novel and co-authoring parts of this research

•  To Iuliana-Alexandra Fleșcan-Lovin-Arseni for annotaKng the English version of the book

•  To Emi Peia for implemenKng the annotator and the TY visualiser

Referințe

DOCUMENTE SIMILARE

Given a single observation sequence for training, we want to find the model (parameters) µ = (A, B, π) that best explains the observed data.. Under Maximum Likelihood Estimation,

In the following we shall stick to some sequences of linear positive oper'.ators, illustrating the generality of the method presented in thc pre- vious paragraph...

The thermal conductivity calculated using the theoretical models for the three different nanoparticles of different volume concentration was compared with the

10 “It is obligatory to take into account the cultural and political path dependencies of the countries and the secularization effects”, Detlef Pollack, Olaf Müller,

De¸si ˆın ambele cazuri de mai sus (S ¸si S ′ ) algoritmul Perceptron g˘ ase¸ste un separator liniar pentru datele de intrare, acest fapt nu este garantat ˆın gazul general,

Thus, if Don Quixote is the idealist, Casanova the adventurous seducer, Werther the suicidal hero, Wilhelm Meister the apprentice, Jesus Christ will be, in the audacious and

In [14], the authors studied all linear maps Φ on B(H ) preserving in both direc- tions semi-Fredholm operators. It has been shown that such maps Φ preserve in both directions the

Ivana et al mentioned in a similar study that no skin lesion was detected in all low voltage current accidents[24].Most accidents of electrocution reported in the above