SELLC Winter School 2010 Evaluating Algorithms for GRE Kees van Deemter (work with Albert Gatt, Ielka van der Sluis, and Richard Power) University of Aberdeen,

SELLC Winter School 2010

Evaluating Algorithms for GRE

Kees van Deemter(work with Albert Gatt, Ielka van der

Sluis, and Richard Power)

University of Aberdeen, Scotland, UK


Outline

• GRE: Generation of Referring Expressions

• TUNA project: Corpus and Annotation

• Evaluation of Algorithms – Furniture Domain– People Domain

• [ Evaluation in the real world: STEC ]


TUNA project (ended Feb. 2007)

• TUNA: Towards a UNified Algorithm for Generating Referring Expressions.

1. Extend coverage of GRE algorithms (plurals, negation, gradable properties,…)

2. Improve empirical foundations of GRE

• Focus on – Content Determination– “First mention” NPs (no anaphora!)

Background

• Dale and Reiter hypothesised that the Incremental Algorithm (IA) led to “better” output than other algorithms

– “better”: more human-like– other algorithms: see below



Other GRE Algorithms

• Full Brevity (FB; Dale 1989)– Generation of minimal descriptions– For example, by first trying all descriptions of length 1,

then length 2, and so on.

• Greedy Algorithm (GR; Dale 1989) – Always add property that removes the most

distractors


Elicitation experiment

• Participants were told that we wanted to test an AI program that interprets referring expressions

• Participants were shown a series of domains• Each domain included 1 or 2 target objects• Participants entered their descriptions,

then the referents were removed• To make the interaction seem real, we sometimes

removed the wrong object! (25% of trials)– The experiment was later repeated without this feature – Essentially the same outcomes were found

• For generality: two types of domains (furniture, people)


Furniture trial


People trial


Method (overview)• Experiment leads to transparent corpus of referring

expressions:– referent and distractors are known– Domain attributes are known

• Transparent corpora can be used for many purposes

This talk: Compare some classic algorithms– giving each algorithm the same input as subjects– computing how similar algorithm’s output is to subjects’ output– We count semantic content only


Elicitation Experiment

• Furniture (simple domain)– TYPE, COLOUR, SIZE, ORIENTATION

• People (complex domain)– Nine annotated properties in total

Location:– Vertical location (Y-DIMENSION)– Horizontal location (X-DIMENSION)

the green desk facing backwards

the sofa and the desk which are red

the young man with a white shirtthe man with the funny haircut

the man on the left

the chair in the top right


Corpus setup• Each corpus was carefully balanced, e.g. between

singulars and plurals.

• Between-subjects design:

-Location: Subjects discouraged from using locative expressions.+Location: Subjects not discouraged.

-FaultCritical: Subjects could correct their utterances+FaultCritical: Subjects could not correct their utterances

• After discounting outliers and (self-reported) non-fluent speakers, 45 subjects were left


• Experiment design: Furniture (-Location)

• 18 trials (C=Colour, O=orientation, S=size)– 1 referent: minimal identification uses

{c}, {o}, {s}, {c,o}, {c,s}, or {o,s} [6 trials]– 2 “similar” referents

{c}, {o}, {s}, {c,o}, {c,s}, or {o,s} [6 trials]– 2 “dissimilar” referents

{c}, {o}, {s}, {c,o}, {c,s}, or {o,s} [6 trials]


Other evaluation studies

Limitations:

• Limited numbers of subjects/referents

• Few attempts at balancing the corpus

• IA: no teasing apart of preference orders

NB Some of these studies were more ambitious in some respects, looking at context, and going beyond identification


Other evaluation studies

• Jordan 2000, Jordan & Walker 2005– More than just identification (Jordan 2000)

• Siddharthan & Copestake 2004– References in linguistic context

• Gupta & Stent 2005– Realisation mixed with Content Determination

• Viethen & Dale 2006– Only Colour and Location


Extensions to the classics• Plurality: (van Deemter 2002)

– Extend each algorithm to search through disjunctions of increasing length

• Location: (van Deemter 2006)– Locatives treated as gradable: “the leftmost table/person”– E.g., suppose the referent x is located in column 3

=> “x is left of column 4”, “x is left of column 5” …=> “x is right of column 2”, “x is right of column 1”…

• Type:– People tend to use TYPE (Dale & Reiter 1995)– Here: All algorithms added TYPE.


Evaluation aims

• Hypothesis in Dale & Reiter 1995: – IA resembles human output most

• Our main questions: – Is this true?– How important are parameters (PO) for the IA?

• More generally: – assess ‘quality’ of classic GRE algorithms :– calculate average match between the description

generated by an algorithm and the descriptions produced by people (for the same referent)


Evaluation metric

• Dice Coefficient:

2 x |Common properties|

|total properties|

corpus: {A,B,C}

algorithm: {B,C} Dice = …

corpus: {A,B,C}

algorithm: {A,B,C,D} Dice = …


Evaluation metric



|total properties|

corpus: {A,B,C}

algorithm: {B,C} Dice = (2*2)/5 = 4/5

corpus: {A,B,C}

algorithm: {A,B,C,D} Dice = (2*3)/7 = 6/7


Evaluation metric



|total properties|

• A coefficient result of 1 indicates identical sets. 0 means no common terms

• We also used this to measure agreement between annotators of the corpus


Assumptions behind DICE

• The discriminatory power of a description does not matter

• All properties are equidistant

See Gatt & Van Deemter 2007, “Content Determination in GRE: evaluating the evaluator”


Evaluation (I): Furniture• Which preference orders for the IA?

– Psycholinguistic evidence:

• COLOUR >> {ORIENTATION, SIZE}(Pechmann 89; Eikmeyer & Ahlsen 96; Belke & Meyer 02)

• Y-DIMENSION >> X-DIMENSION(Bryant et al, 1992; Arts 2004)

• Split data: +LOCATION vs –LOCATION This talk: focus on –LOCATION –LOCATION = approx. 800 descriptions

• Compare algorithms to a randomized IA (RAND)


Furniture: -LOCATION

SignificantSignificant

FB/GR


Beyond Toy Domains• More on Furniture corpus:

Gatt et al. (ENLG-2007)

• With complex real-world objects:– Many different attributes can be used– Number of PO’s explodes– Few psycholinguistic precedents

• People domain attributes:– { hasBeard, hasGlasses, age, hasTie,

hasSuit, hasSuit, hasHair, hairColour, orientation }– 9 Attributes, so 9! = 362880 possible POs


IA: Preference Orders for People Domain

• Little psycholinguistic evidence for choosing between all 362880 possible PO’s

• Focus on the most frequent Attributes: G=hasGlasses, B=hasBeard, H=hasHair, C=haircolour– Assumption: H and B must precede C– This leaves us with eight POs:

{ GBHC, GHBC,HBGC,HBCG, HGBC,BHGC, BHCG, BGHC }


Preference Orders and frequency

Mean Sum

type 1.39 475

hasGlasses .68 231

hasBeard .66 226

HairColour .61 210

hasHair .46 158

orientation .21 73

age .10 34

hasTie .04 12

hasSuit .01 4

hasShirt .01 3

• For attributes other than {G,C,H,B}, we let corpus frequency determine the order

• E.g, IA-GBHC uses

type, G,B,H,C, age,

hasTie, hasSuit,hasShirt

as its PO


Results People Domain

IA-BASE

Significant Significant by subjects

GR


Results People domain

• IA_base performs very badly now

• So much about the best IA’s that start with {B,H,G,C} and end with <age,hasTie,hasSuit,hasShirt>

• Some of these did much worse:– IA_BHCG had DICE=0.6, making it

significantly worse (by subjects) than GR!


Summary

• People domain gives much lower DICE scores than Furniture domain

• Difference between “good” and “bad” POs was – small (but significant) in the Furniture domain, – big (and significant) in the People domain


Summary• The “Incremental Algorithm” (IA):

– not an algorithm but a class of algorithms

• The best IA beats all other algorithms, but the worst is very bad ...

• GR performs remarkably well.

• How to choose a suitable PO?– Furniture: few attributes; psycholinguistic precedent

• Still, there is variation.

– People: more attributes; no precedents• Variation even greater!


Discussion• Suppose you want to build a GRE

algorithm for a new and complex domain, for which no transparent corpus is available.

• Psycholinguistic principles are unlikely to help you much

• If corpus is also not balanced, then frequency may not say much either …


Other uses of this method: STEC

• Summer 2007: First NLG Shared task Evaluation Challenge (STEC)

• STEC involved GRE only, focussing on Content Determination

• 22 GRE Algorithms were submitted and evaluated (6 teams)

• Reported in UCNLG+MT workshop, Copenhagen, Sept 2007


Other uses of this corpus: STEC

• An even bigger STEC one year later

• Each algorithm was compared with the TUNA corpus (minus 40% training set) – Both Furniture and People domain – DICE measured “humanlikeness”– Singulars only

• Each algorithm was also tested in terms of identification time (by human reader)


Some STEC results

1. The more minimal the descriptions generated by these 22 systems were, the worse their DICE scores were



2. No relation between humanlikeness and identification time

– Best system in terms of DICE was worst-but-one in terms of identification time

• More research needed on the different criteria for judging NLG output


Thank you


Annotator agreement

• Semantic markup was applied manually to all descriptions in the corpus.

• 2 annotators were given a stratified random sample

• Comparison used Dice.

mean mode

Furniture 0.89 (A/B)

1 (71.1%)

Annotator A 0.93 (A/us)

1 (74.4%)

Annotator B 0.92 (B/us)

1(73%)

People 0.89 (A/B)

1(70%)

Annotator A 0.84 (A/us)

1(41.1%)

Annotator B .78 (B/us)

1(36.3%)

SELLC Winter School 2010 Evaluating Algorithms for GRE Kees van Deemter (work with Albert Gatt, Ielka van der Sluis, and Richard Power) University of Aberdeen,

Documents

sellc winter school

location slide

referent slide

right slide

trials c

stec slide

people trial slide

furniture trial slide