Context-sensitive description of objects
Post on 05-Jan-2016
21 Views
Preview:
DESCRIPTION
Transcript
1
Context-sensitive description of objects
Mariët Theune(joint work with Emiel Krahmer)
2
Introduction
An important question:How to generate distinguishing descriptions of objects?
State of the art: Incremental Algorithm, Dale & Reiter (1995)
Today’s aims:• Show that Dale and Reiter’s algorithm can be refined by taking salience into account• … which opens the way for several interesting
extensions
3
Overview
• Dale & Reiter’s Incremental Algorithm• A modified Incremental Algorithm• Extensions:
– Pronouns– Relational descriptions
• Implementation • Evaluation• Concluding remarks • Related work
4
The Incremental Algorithm
Terminology
Distinguishing description
An accurate description of the intended referent r, but
not of any other object in the current context set
Distractors
The objects from which r has to be distinguished (= all
objects other than r)
5
Terminology (cont.)
Preferred attributes
The properties that human speakers and hearers prefer
for a specific domain
Best value
The value that is closest to the basic level value of a
property, and that still rules out the maximal number of
distractors
6
The Incremental Algorithm
Strategy
Iterate through the list of preferred attributes,
• adding the best value of an attribute if: - it rules out any distractors not previously ruled out - or the attribute is ‘type’
• terminating when a distinguishing description has been constructed (all distractors have been ruled out)
7
The Incremental Algorithm
Example
Domain (Dale & Reiter 1995:258):
d1 <type, chihuahua>, <size, small>, <colour, black>
d2 <type, chihuahua>, <size, large>, <colour, white>
d3 <type, siamese cat>, <size, small>, <colour, black>
Preferred attributes: < type, colour, size >
8
Example (cont.)
animal
DOG CAT d1
chihuahua poodle siamese cat
Describe d2:1. Property ‘type’, best value = ‘dog’ d2
2. Property ‘colour’, best value = ‘white’
d3Result: < white, dog >
9
The Incremental Algorithm
Good points
• Fast and efficient due to lack of backtracking• Psychologically realistic
Still lacking
• Construction of natural language expressions• Context sensitivity
10
A modified algorithm
Adding salience
Intuition
A definite description refers to the most salient object
which has the properties expressed by it
Salience (Lewis 1979):The dog got in a fight with another dog
11
Adding salience (cont.)
Salience weights (sw)
In each state, every object is assigned a natural number
between 0 (not salient) and 10 (maximally salient)
Salience weight assignment• In the initial state, salience weight is 0 for all objects• If an object is mentioned, its salience weight increases• If an object is not mentioned, its salience weight
decreases (to a minimum of 0)
12
A modified algorithm
Strategy
Same as the Incremental Algorithm, except:
• The distractors are those objects in the domain with a salience weight that is equal to or higher than that of the intended referent
• An NP tree is built within the algorithm, to check the expressibility of properties (cf. Horacek 1997)
13
A modified algorithm
Example (1) d1
Input:
• Object r = d2 d2
• State s0 :
sw(d1)= sw(d2) = sw (d3) = 0
• P = < type, colour, size … > d3
• L = <>
Result: the2 white dog
14
A modified algorithm
Example (2) d1
Context = The2 white dog and the3 cat …
Input
• Object r = d2 d2
• State s : sw(d2)= sw(d3) > sw (d1)
• P = < type, colour, size … >
• L = <> d3
The2 dog …
15
A modified algorithm
Example (3) d4
Context = The2 white chihuahua …
Input d2
• Object r = d2
• State s : sw(d2) > sw (d1,3,4) d1
• ... d3
The2 dog …
16
Extensions
PronounsIf• r is currently the single most salient object • and there is an antecedent for r
Then pronominalise reference to r
The2 white chihuahua was fast asleep.
It2 was dreaming of tasty bones.
17
Extensions
Relational descriptions
• Add relations as attributes• Add a hierarchy of relations spatial
NEXT_TO IN
left_of right_of
If a relation with object r’ is included when describing r,then recursively call the algorithm to describe r’
18
Relational descriptions (cont.)
Input: d2 , s0 , P = <type, …, spatial>, L = <>
• First property rules out d1 and d4, best value is ‘dog’
• Next properties (‘colour’, ‘size’) rule nothing out
• The spatial relation rules out d3; best value ‘next to’
d1 d2 d3 d4
19
Relational descriptions (cont.)
Recursive call, input: d1, s0 , P, L = < next_to (d2,d1) >
• The spatial relation in L rules out all distractors• The ‘type’ property is included by default
• Resulting description of d1: the snowman
d1 d2 d3 d4
20
Relational descriptions (cont.)
Recursive call, input: d1, s0 , P, L = < next_to (d2,d1) >
• The spatial relation in L rules out all distractors• The ‘type’ property is included by default
• Resulting description of d1: the snowman
d1 d2 d3 d4
The2 dog next to the1 snowman
21
Implementation
The modified algorithm has been implemented in LGM,
IPO’s data-to-speech system. Applications of LGM are:
• DYD: information about Mozart compositions • GoalGetter: soccer reports • OVIS: train time information• VODIS: in-car route descriptions • Pavlov: toy system, testing the modified algorithm
22
Evaluation
The basic assumptions underlying the modifiedalgorithm are that an anaphoric reference:
1. Contains fewer properties2. Uses more general phrasing3. Is pronominalised whenever possible4. Obeys 1 and 2 also after an intervening sentence
All hypotheses were experimentally confirmed, except 2(which turns out to depend on wording)
23
Concluding remarks
Context Description
-
the white chihuahua
the black and the white chihuahua
the white chihuahua
the dog / it
the white dog
Use of salience allows for the generation of
context-sensitive descriptions:
24
Related work
Krahmer, van Erk & Verleg (2001):• Domain represented as a labeled, directed graph• Property selection is subgraph construction
d1 d2 d3
d3
white
dog
dog
white
d2
d1
snowmannext-to
dog
snowmannext-to
25
Related work
Van der Sluis & Krahmer (in progress): generating
referring expressions in a multi-modal context
• Three kinds of salience: linguistic, inherent, and focus space salience
• Pointing decision and determiner choice are added
that white one left of the blue one
* *this blue one
top related