Generation of Referring Expressions: the State of the Art LOT Winter School, Tilburg 2008 Kees van Deemter Computing Science University of Aberdeen.

Generation of Referring Expressions: the State of the Art LOT Winter School, Tilburg 2008

Kees van Deemter

Computing Science

University of Aberdeen

Open Questions in GRE

Open Questions in GRE

Based on experiences in TUNA project Your input is welcome!

suggestions about other open questions? ideas about answering them

OQ1: Reference in context

Project proposal submitted: How can existing GRE algorithms be adapted to produce appropriate references in a (discourse or) dialogue context? Much work exists on the choice between broad

categories, e.g., pronoun vs. full NP vs demonstrative (Poesio et al; Piwek). This does not help to decide what NP to choose. Integration with GRE is needed.

Pioneering accounts are available (Krahmer & Theune 2002, Siddharthan & Copestake 2004, Stoia et al 2007), but these are tentative and largely untested.

Dialogue requires modelling of interaction between speaker and hearer (e.g., alignment and collaboration)

OQ2 Issues regarding knowledge and belief

How should mismatches in knowledge between Speaker and Hearer be modelled?

GRE so far has kept epistemic operators implicit: all the information in the crucial part of the KB was “shared”.

What if S and H differ?

OQ2 Issues regarding knowledge and belief

S: [[P]]={a,b}, [[Q]]={b,c}H: [[P]]={b,c}, [[Q]]={c,d}

If S says “P&Q” to refer to b, the H will misunderstand this as referring to c

Shared Knowledge is not the intersection of [[P]]S and [[P]]H (or else “P” would refer to b)

When S and H differ, S can initiate a clarification dialogue or try to step into H’s shoes:

BelSBelH(P)?

BelSBelHBelS(P)? -- Etc.

Relevant to the simplifications made by current GRE algorithms. E.g.: “The present king of France”

(Frege/Russell/Strawson) “Whoever it will be, the winner of this year’s Tour de

France will be less proud than last year’s winner” It’s possible that the winner of the lottery will win 20

million. X believes that a witch ...; Y believes that she ...”. “The man with the martini is the murderer”, when it’s

actually a soft drink (Donnellan)

More radical new departures needed?

Consider texts about genuinely complex domains: “We examine the problem of generating

definite noun phrases that are appropriate referring expressions” (Opening sentence of the abstract of D&R 1995.)

“Bush’s Middle-East policies are a disaster. Even his closest aids have started to withdraw their support”

What do these NPs refer to? Is it realistic to want to generate them from a shared KB?

OQ3: Reference to sets

We know how to refer to sets making use of properties that hold of their elements.

We have gained some understanding of the role of intra-NP lexical coherence (Gatt’s thesis, Gatt & Van Deemter 2007). But, It’s not clear how lengthy descriptions are best avoided computational efficiency remains a difficult issue Little is known about the use of properties that hold

collectively of the referent set as a whole Interaction between salience and sets is still a mystery.

OQ3: Incrementality

Studies of the TUNA corpus suggest that incremental GRE can work very well ... ... but only if you have a good preference order

How can good preference orders be found? Does every new domain necessitate new empirical

studies? Or are there general principles that underlie preference

orders? (E.g., frequency or complexity of a property) Psycholinguistic issues:

Do human speakers check properties in a fixed order? What’s the link with left-to-right order? (J.Sedivy et al.

1999)

OQ4: Hearer-oriented GRE

Most empirical work on reference (in the GRE community and by psycholinguists) has focussed on production. Exceptions: Paraboni et al. (2007). Preliminary study in

first STEC (Belz & Gatt 2007). How might one build generators that optimise for the

hearer? (High processing speed, low likelihood of errors)

And what if it turns out that speakers are bad at this? If it’s practical GRE you’re interested in then this

allows GRE programs to do better than human speakers. Implications for theoretical GRE are less clear.

OQ5 Multimodality

How does textual GRE interact with non-linguistic issues, such as speech (e.g. pitch accent on “given” information; other

prosodic issues; cf. Theune’s thesis) pointing (e.g. Van der Sluis & Krahmer [to appear]) salience as determined by physical proximity (as well

as textual recency, intrinsic importance of objects, etc.) facial expressions such as gaze, eyebrow movements.

These and other issues to be explored in Krahmer’s new VICI project on GRE (Tilburg, 2008-2012).

OQ6 Realisation & Lexical Choice

Much of what we discussed focusses on Content Determination

But referring expressions require words and syntactic constructions as well!

But surface phenomena can be difficult and interesting too One example: Gatt’s exploration of lexical

coherence Another example: PhD project by Imtiaz Khan

(Aberdeen)

OQ6 Realisation & Lexical Choice

Siddharthan & Copestake (2004) observed: words can introduce ambiguities. E.g.

“The old president” = the previous present, or the president who is old (i.e., aged)

Khan: Syntax can be ambiguous as well: “the man on the hill with the telescope” “the old men and women”

OQ6 Realisation & Lexical Choice

One possible position: “avoid ambiguities at all cost”.

Khan: ambiguous strings are not only often generated, but sometimes also preferred by hearers “the old men and women” preferred over “the old men and the old women”

Hypothesis: surface ambiguities are balanced against other issues (e.g. brevity)

Challenge: test this hypothesis, and find an algorithm that mirrors human behaviour

OQ7 Reference in spacial domains

There is preliminary work (e.g. by Gatt), based on simple domains

What happens when you want to refer to an area of a country? Ross Turner’s PhD project (Aberdeen) Input: a set of points in Scotland where ice is

predicted to hamper road traffic Example output: “icy patches are expected in

the North East and on high grounds”

OQ7 Reference in spacial domains

Ross Turner’s PhD project (Aberdeen) Input: a set of points in Scotland where ice is

predicted to hamper road traffic. (Each point is on a road.)

Example output: “Icy patches are expected in the North East and on high grounds”

This is GRE ... but with a twist : it may not be necessary to include all target points it may not be necessary to exclude all other points

Referential success becomes a graded affair!

OQ8 Integration with the rest of NLG

GRE is arguably the most mature area of NLG: Linguistic Realisation is the main other contender most GRE practitioners use the same assumptions the fact that the first NLG STEC focused on GRE

confirms this Ultimately, the GRE problem is “linguistically

complete”: if we had a flawless GRE algorithm then this algorithm

could easily be transformed into an equally flawless algorithm for all of NLG ...

OQ8 Integration with the rest of NLG

For example, John walks [S] (The person who) walks [ref NP]

Or A man saw a girl with earrings [S] (The man who) saw a girl with earrings [ref NP]

Or Someone saw a beautiful girl with incredibly elaborate

jade earrings bought in Paris (...) [S] (The person who) saw a beautiful girl with incredibly

elaborate jade earrings bought in Paris (...) [ref NP]

OQ8 Integration with the rest of NLG

Even in the absence of a flawless GRE algorithm, there is no reason why GRE should “stand alone”:

If GRE is a microcosm for all of NLG, then the rest of NLG should be able to learn from GRE – and the other way round.

Getting a better integration between GRE and other areas of NLG is a major challenge.

OQ9: Non-atomic knowledge

Most GRE algorithms assume that the (shared) knowledge in the input KB consists of atoms and their negations.

For example ...

Atomic/non-atomic knowledge

Type: furniture (abcde), desk (ab), chair (cde) Origin: Sweden (ac), Italy (bde) Colours: dark (ade), light (bc), brown (a) Price: 100 (ac), 150 (bd) , 250 ({}) Contains: wood ({}), metal ({abcde}), cotton(d)

(Closed World Assumption: if an item is not listed then it’s not in the extension of the property)

But suppose ...

Every desk comes with a lamp. It would be possible to add lamps {x,y} to the KB. Additionally, we could add relations between desks

and lamps to the KB, e.g. has_lampa (y)

has_lampb (x)

This would enable us to refer to “the lamp of b” To achieve the same effect, we could just add one

general rule: m(desk(m) (!n(lamp(n) & has_desk(m,n)))

Atomic/non-atomic knowledge

More complex examples are easy to think of. For example, perhaps we know

m(desk(m) (n(lamp(n) &has_desk(m,n))) lamp(x) price(x)>50 It follows that desk a cannot have 2 lamps

(since 2 lamps cost more than a’s 100 pounds). Therefore

Desk a has exactly one lamp. Therefore We can refer (uniquely) to a’s lamp.

“a’s lamp”, “the lamp” (bridging)

Atomic/non-atomic knowledge

To anyone familiar with KR, this will be obvious: KR with positive & negative literals only is limited. Even PROLOG, for example, has rules, e.g.

grandfather(X Y) :- father(X Z), parent(Z Y)

Someone needs to add rules to GRE! (First attempt using Conceptual Graphs : Croitoru & Van Deemter, IJCAI 2007. There is much room for elaboration and improvement!)

OQ10: Integration between GRE and other areas of linguistics Integration with psycholinguistics:

(NLG more generally: G.Kempen et al, A.Roelofs et al. Recent book by M.Guhe.)

GRE: modest beginnings in Dale & Reiter 1995 (inspiration from Levelt’s book)

investigations into optimal preference orders (Gatt et al 2007, Van der Sluis et al 2007)

Coherent reference to sets (Gatt’s thesis, 2007)

Krahmer’s VICI project aims to take this issue very seriously

OQ10: Integration between GRE and other areas of linguistics Integration with syntax has so far been

meagre interleaving of Linguistic Realisation and

Content Determination (Stone & Webber 1998; Krahmer & Theune 2002)

c-command constraints on the use of reflexives and pronouns (J.Odijk, DYD/ GOALGETTER systems)

OQ10: Integration between GRE and other areas of linguistics Integration with formal semantics and

pragmatics has been limited Various attempts by Stone and colleagues.

E.g. DeVault & Stone 2004 on vagueness (based on Kyburg & Morreau 2000)

Use of salience (mainly for category choice; see also Krahmer & Theune 2002)

GRE uses simple extensional semantics, whereas formal semantics focusses on intensionality and quantification

OQ10: Integration between GRE and other areas of linguistics It would be interesting to let GRE explore core areas

of formal semantics, e.g. Use a flat KB as input (just like in GRE), to generate

quantified NPs like “Five rats died”, “A few rats died”, “Not all rats died”.

Find principles for choosing the quantifier pattern that’s most appropriate in the utterance situation

Early attempts by N.Creaney (2002), but limited progress so far.

Q11: Problematic referents

“the water in this pond” “water” “5”, “2+3” “virtue”, “power”

(cf. G.Chierchia’s invited address)

Plenty of challenges for enthusiastic young researchers!