RANLP, Borovets Sept Evaluating Algorithms for GRE (Going beyond Toy Domains) Ielka van der Sluis Albert Gatt Kees van Deemter University of Aberdeen, Scotland, UK
RANLP, Borovets Sept Outline GRE: Generation of Referring Expressions TUNA project: Corpus and Annotation Evaluation of Algorithms –Furniture Domain –People Domain [ Evaluation in the real world: STEC ]
RANLP, Borovets Sept 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!)
RANLP, Borovets Sept TUNA results Elsewhere: –Reference to sets (e.g., Gatt 2006, 2007) –Gradable/vague properties (Van Deemter 2006) –Pointing (Van der Sluis & Krahmer 2007) –Large domains (Paraboni et al. 2007) This talk: empirical issues –Testing classic algorithms –Method: compute similarity to human-generated NPs
RANLP, Borovets Sept Method (overview) Elicitation 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
RANLP, Borovets Sept 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 shirt the man with the funny haircut the man on the left the chair in the top right
RANLP, Borovets Sept Furniture trial
RANLP, Borovets Sept People trial
RANLP, Borovets Sept 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
RANLP, Borovets Sept 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]
RANLP, Borovets Sept Classic GRE Algorithms Full Brevity (FB; Dale 1989) –Generation of a minimal description Greedy Algorithm (GR; Dale 1989) –Always add property that removes the most distractors Incremental Algorithm (IA; Dale and Reiter 1995) –Add next useful property from an ordered list of properties. (“Preference Order” = PO)
RANLP, Borovets Sept 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
RANLP, Borovets Sept Other evaluation studies General limitations: Limited numbers of subjects/referents Few attempts at balancing the corpus. (E.g., Viethen & Dale 2006 let subjects decide what to refer to.) IA: no teasing apart of preference orders
RANLP, Borovets Sept 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.
RANLP, Borovets Sept 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)
RANLP, Borovets Sept Evaluation metric Dice Coefficient: 2 x |Common properties| |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
RANLP, Borovets Sept (Assumptions behind DICE) Deletion of a property is slightly worse than addition of a property 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” )
RANLP, Borovets Sept 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)
RANLP, Borovets Sept Furniture: -LOCATION Significant FB/GR
RANLP, Borovets Sept 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! = possible POs
RANLP, Borovets Sept IA: Preference Orders for People Domain Little psycholinguistic evidence for choosing between all 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 }
RANLP, Borovets Sept Preference Orders and frequency Mean (std)Sum type hasGlasses hasBeard HairColour hasHair orientation.2173 age.1034 hasTie.0412 hasSuit.014 hasShirt.013 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
RANLP, Borovets Sept Results People Domain IA-BASE Significant Significant by subjects GR
RANLP, Borovets Sept 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 Some of these did much worse: –IA_BHCG had DICE=0.6, making it significantly worse (by subjects) than GR!
RANLP, Borovets Sept Summary People domain gives much lower DICE scores than Furniture domain Difference between “good” and “bad” POs was enormous in People domain
RANLP, Borovets Sept 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!
RANLP, Borovets Sept 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 doesn’t say much either …
RANLP, Borovets Sept 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
RANLP, Borovets Sept Other uses of this corpus: STEC 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)
RANLP, Borovets Sept Other uses of this corpus: STEC Future STEC: –beyond “first mention” –beyond Content Determination –more hearer-oriented experiments
RANLP, Borovets Sept STEC results 1.The more minimal the descriptions generated by these 22 systems were, the worse their DICE scores were
RANLP, Borovets Sept. 2007
RANLP, Borovets Sept 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
RANLP, Borovets Sept Thank you
RANLP, Borovets Sept Annotator agreement Semantic markup was applied manually to all descriptions in the corpus. 2 annotators were given a stratified random sample Comparison used Dice. meanmode Furniture0.89 (A/B) 1 (71.1%) Annotator A0.93 (A/us) 1 (74.4%) Annotator B0.92 (B/us) 1(73%) People0.89 (A/B) 1(70%) Annotator A0.84 (A/us) 1(41.1%) Annotator B.78 (B/us) 1(36.3%)