Improving String Processing for Temporal Relations Tim Fernando - - PowerPoint PPT Presentation

▶

Aug 20, 2023 128 likes •881 views

Improving String Processing for Temporal Relations Improving String Processing for Temporal Relations Tim Fernando David Woods ADAPT Centre Computational Linguistics Group Trinity Centre for Computing and Language Studies School of Computer

SLIDE 1

Improving String Processing for Temporal Relations

David Woods Tim Fernando

ADAPT Centre Computational Linguistics Group Trinity Centre for Computing and Language Studies School of Computer Science and Statistics Trinity College Dublin, Ireland

ISA-14, August 25th 2018

SLIDE 2

Improving String Processing for Temporal Relations Introduction

Introduction

Q. What’s this talk about?

SLIDE 3

Improving String Processing for Temporal Relations Introduction

Introduction

Q. What’s this talk about?
A. Representing temporal information compactly for reasoning

and processing.

SLIDE 4

Improving String Processing for Temporal Relations Introduction

Introduction

Q. What’s this talk about?
A. Representing temporal information compactly for reasoning

and processing. Example “John slept through the fire alarm last Tuesday.” lt lt, js lt, js, fa lt, js lt

SLIDE 5

Improving String Processing for Temporal Relations Introduction

Motivation

Ordering events and times is crucial in natural language understanding.

SLIDE 6

Improving String Processing for Temporal Relations Introduction

Motivation

Ordering events and times is crucial in natural language understanding. Providing a useful way to both viualise a document’s temporal structure, and perform inference with the same framework.

SLIDE 7

Improving String Processing for Temporal Relations Introduction

Motivation

Ordering events and times is crucial in natural language understanding. Providing a useful way to both viualise a document’s temporal structure, and perform inference with the same framework. Assisted annotation for TimeML or similar schemas.

SLIDE 8

Improving String Processing for Temporal Relations Introduction TimeML

ISO-TimeML

What is TimeML?

SLIDE 9

Improving String Processing for Temporal Relations Introduction TimeML

ISO-TimeML

What is ISO-TimeML?

SLIDE 10

Improving String Processing for Temporal Relations Introduction TimeML

ISO-TimeML

What is ISO-TimeML? An ISO standard markup language for annotating temporal information in texts.

SLIDE 11

Improving String Processing for Temporal Relations Introduction TimeML

Why TimeML?

Widely known, and ISO standard. The TLINK elements map well to Allen’s interval relations. The TimeBank corpus (183 documents manually annotated with TimeML).

SLIDE 12

Improving String Processing for Temporal Relations Introduction TimeML

TimeML Example

Example John taught 20 minutes every Monday.

Example from http://www.timeml.org/publications/timeMLdocs/timeml_1.2.1.html

SLIDE 13

Improving String Processing for Temporal Relations Introduction TimeML

TimeML Example

Example John taught 20 minutes every Monday.

Example from http://www.timeml.org/publications/timeMLdocs/timeml_1.2.1.html

SLIDE 14

Improving String Processing for Temporal Relations Introduction TimeML

TLINKs Example

Example John taught 20 minutes every Monday. <TLINK timeID="t1" relatedToTime="t2" relType="IS INCLUDED"/> <TLINK eventInstanceID="ei1" relatedToTime="t1" relType="DURING"/>

SLIDE 15

Improving String Processing for Temporal Relations Introduction TimeML

TLINKs Example

Example John taught 20 minutes every Monday. <TLINK timeID="t1" relatedToTime="t2" relType="IS INCLUDED"/> <TLINK eventInstanceID="ei1" relatedToTime="t1" relType="DURING"/> Several ways to represent this information.

SLIDE 16

Improving String Processing for Temporal Relations Introduction TimeML

TLINKs in a Directed Graph

Example John taught 20 minutes every Monday.

SLIDE 17

Improving String Processing for Temporal Relations Introduction TimeML

TLINKs using T-BOX

Example John taught 20 minutes every Monday.

SLIDE 18

Improving String Processing for Temporal Relations Introduction TimeML

TLINKs as Strings

Example John taught 20 minutes every Monday. t2 t1, t2 ei1, t1, t2 t1, t2 t2

SLIDE 19

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

Sequences of Sets-as-Symbols

Strings of n sets α1α2 · · · αn are used to represent n moments

f time.

SLIDE 20

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

Sequences of Sets-as-Symbols

Strings of n sets α1α2 · · · αn are used to represent n moments

f time.

Each set αi contains exactly those fluents (temporal propositions, treated as intervals) which are true at moment i. e.g. {a}{a, b}{b}

SLIDE 21

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

Sequences of Sets-as-Symbols

Strings of n sets α1α2 · · · αn are used to represent n moments

f time.

Each set αi contains exactly those fluents (temporal propositions, treated as intervals) which are true at moment i. e.g. {a}{a, b}{b} Sets are drawn as boxes, so strings can be read like comic strips, snapshots of film, or timelines. e.g. a a, b b

SLIDE 22

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

Comic Strips

Image from The National Archives UK

SLIDE 23

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

An Intertial World

Note! A fluent occurring in multiple boxes does not imply a longer duration.

SLIDE 24

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

An Intertial World

Note! A fluent occurring in multiple boxes does not imply a longer duration. For example, a is equivalent in interpretation to a a a , though the latter string is said to feature stutter.

SLIDE 25

Improving String Processing for Temporal Relations Introduction Strings for Temporal Data

An Intertial World

Note! A fluent occurring in multiple boxes does not imply a longer duration. For example, a is equivalent in interpretation to a a a , though the latter string is said to feature stutter. We can remove stutter through a block-compression

peration:

b c(s) :=    s if length(s) ≤ 1 b c(αs′) if s = ααs′ α b c(α′s′) if s = αα′s′ with α = α′

SLIDE 26

Improving String Processing for Temporal Relations Introduction Allen’s Interval Relations

Allen Relations

Allen treats intervals as primitive for events, not start/end points. Allen’s Interval Relations form the basis of TimeML’s TLINK relation types. These are easily transformable to strings.

SLIDE 27

Improving String Processing for Temporal Relations Introduction Allen’s Interval Relations

Allen Relations as Strings

R a R a′ SR (a, a′) R−1 a R−1 a′ SR−1 (a, a′) < a before a′ a a′ > a after a′ a′ a m a meets a′ a a′ mi a met by a′ a′ a

a overlaps a′

a a, a′ a′

a overlapped by a′ a′ a′, a a d a during a′ a′ a, a′ a′ di a contains a′ a a′, a a s a starts a′ a, a′ a′ si a started by a′ a′, a a f a finishes a′ a′ a, a′ fi a finished by a′ a a′, a = a equals a′ a, a′

Allen’s interval relations as strings.

SLIDE 28

Improving String Processing for Temporal Relations Introduction Allen’s Interval Relations

Extraction of ARs as Strings from TimeML

https://www.scss.tcd.ie/~dwoods/isa14/

SLIDE 29

Improving String Processing for Temporal Relations Introduction Superposition

Combining Strings

Superposition allows us to condense the information from multiple strings into a more compact form.

SLIDE 30

Improving String Processing for Temporal Relations Introduction Superposition

Combining Strings

Superposition allows us to condense the information from multiple strings into a more compact form. The simplest version of the operation is just componentwise union of two strings of equal length: α1α2 · · · αn & α′

1α′ 2 · · · α′ n := (α1 ∪ α′ 1)(α2 ∪ α′ 2) · · · (αn ∪ α′ n)

SLIDE 31

Improving String Processing for Temporal Relations Introduction Superposition

Combining Strings

Superposition allows us to condense the information from multiple strings into a more compact form. The simplest version of the operation is just componentwise union of two strings of equal length: α1α2 · · · αn & α′

1α′ 2 · · · α′ n := (α1 ∪ α′ 1)(α2 ∪ α′ 2) · · · (αn ∪ α′ n)

Example (Basic Superposition) a b c & a a d = a a, b c, d

SLIDE 32

Improving String Processing for Temporal Relations Introduction Superposition

Asynchronous Superposition

We handle strings of unequal length by introducing stutter using an inverse block compression, such that the operands are padded to the point that basic superposition is possible.

SLIDE 33

Improving String Processing for Temporal Relations Introduction Superposition

Asynchronous Superposition

We handle strings of unequal length by introducing stutter using an inverse block compression, such that the operands are padded to the point that basic superposition is possible. Noting that every string s′ ∈ b c−1b c(s) is equivalent under compression, we define asynchronous superposition: s &∗ s′ := {b c(s′′) | s′′ ∈ b c−1b c(s) & b c−1b c(s′)}

SLIDE 34

Improving String Processing for Temporal Relations Introduction Superposition

Asynchronous Superposition

Example (Asynchronous Superposition) x z &∗ x y z = { x x, y z , x x, y x, z z , x y, z z , x x, y y, z z , x x, z y, z z }

SLIDE 35

Improving String Processing for Temporal Relations Introduction Superposition

Issues

Superposition doesn’t account for coreferent fluents.

SLIDE 36

Improving String Processing for Temporal Relations Introduction Superposition

Issues

Superposition doesn’t account for coreferent fluents. This causes overgeneration of strings as a result.

SLIDE 37

Improving String Processing for Temporal Relations Introduction Superposition

Issues

Superposition doesn’t account for coreferent fluents. This causes overgeneration of strings as a result. Each string produced must be examined and thrown out if invalid.

SLIDE 38

Improving String Processing for Temporal Relations Introduction Superposition

Issues

For example, “x before y” – x y superposed with “y before z” – y z generates 270 strings, only one of which is correct ( x y z ).

SLIDE 39

Improving String Processing for Temporal Relations Introduction Superposition

Issues

For example, “x before y” – x y superposed with “y before z” – y z generates 270 strings, only one of which is correct ( x y z ). The remaining 269 do not retain the original information that both x is before y, and y is before z.

SLIDE 40

Improving String Processing for Temporal Relations Introduction Superposition

Issues

For example, “x before y” – x y superposed with “y before z” – y z generates 270 strings, only one of which is correct ( x y z ). The remaining 269 do not retain the original information that both x is before y, and y is before z. This is costly! We can do better.

SLIDE 41

Improving String Processing for Temporal Relations Improvements Notation

Notation

Some quick extra notation!

SLIDE 42

Improving String Processing for Temporal Relations Improvements Notation

Vocabulary

The vocabulary of a string s will be said to be the union of each of its components: voc(s) :=

n

αi Example (“John slept through the fire alarm last Tuesday”) voc( lt lt, js lt, js, fa lt, js lt ) = {lt, js, fa} lt = Last Tuesday js = John sleeps fa = Fire alarm sounds

SLIDE 43

Improving String Processing for Temporal Relations Improvements Notation

Reduct

For any set A, the A-reduct of a string s is defined as the componentwise intersection of s with A: ρA(α1α2 · · · αn) := (α1 ∩ A)(α2 ∩ A) · · · (αn ∩ A) Example ρ{lt,fa}( lt lt, js lt, js, fa lt, js lt ) = lt lt lt, fa lt lt

SLIDE 44

Improving String Processing for Temporal Relations Improvements Projection

Projection

We’ll say that a string s projects to another string s′ if the voc(s)-reduct of s block compresses to s′: b c(ρvoc(s′)(s)) = s′

SLIDE 45

Improving String Processing for Temporal Relations Improvements Projection

Reduction to Desired Information

We can use this idea of compression along with specified vocabulary to determine the Allen Relation between any two fluents in an arbitrary string.

SLIDE 46

Improving String Processing for Temporal Relations Improvements Projection

Reduction to Desired Information

We can use this idea of compression along with specified vocabulary to determine the Allen Relation between any two fluents in an arbitrary string. Example lt lt, js lt, js, fa lt, js lt projects to the string lt fa, lt lt , which corresponds to “fa during lt”.

SLIDE 47

Improving String Processing for Temporal Relations Improvements Projection

Projection-based Constraints

Additionally, we can say that the resulting string(s) of a superposition should project back to both of their input strings in order to retain the initial information. Example x y superposed with y z produces 270 strings, but of these, only x y z projects back to both of its “parents”.

SLIDE 48

Improving String Processing for Temporal Relations Improvements Projection

Improving Superposition with Projection

By building the notion of projection into superposition, we can prevent invalid strings from being generated at all. No more need to test for inconsistency or loss of information after generation, testing is built-in. This vastly increases execution speed of the operation, and paves the way for handling more complex situations than we have seen so far.

SLIDE 49

Improving String Processing for Temporal Relations Improvements Projection

Improving Superposition with Projection

Algorithm details are in the paper!

As high as 94% in certain cases!

SLIDE 50

Improving String Processing for Temporal Relations Improvements Projection

Improving Superposition with Projection

Algorithm details are in the paper! Suffice to say that on average (for all pairs of strings which each contain two fluents), we saw a 58% speed up when going from the old generate-then-test approach to the new superposition which incorporates projection.

As high as 94% in certain cases!

SLIDE 51

Improving String Processing for Temporal Relations Improvements Projection