ICoS-5 Conference Report
From SIGSEM
A conference report by Johan Bos
Contents |
What's this remarkable building?
This is the impressive dome of Buxton, England (on a nice day, many Buxtonians would add). When it was built by Henry Currey in 1859 it was said to be the world's largest unsupported dome (larger even than the Pantheon in Rome). The building was in use by the Devonshire Royal Hospital for a long time, before becoming part of the University of Derby. Inside the building one can see a model of the Foucault Pendulum, a device demonstrating the rotation of the earth.
And this is the building where two dozens of computational semanticists from all around the globe met on the 20th and 21st of April, 2006, to confer about the relationships between language, computer science, logic, and inference. The event, known to insiders as ICoS-5, the fifth international conference on Inference in Computational Semantics, was in the organisational hands of Ian Pratt-Hartmann, Allan Ramsay and Allan Third. The programme of ICoS-5 featured three invited speakers, twelve regular papers, and a poster and demo session.
But what exactly happened on these two days? And who was the winner of the traditional Richard Montague Memorial Award? Read on to find out!
Thursday, April 20th, 2006
The conference was given an excellent start by the first invited speaker, Stephen Pulman, who explained to the audience why it so difficult to bridge the gap between formal ideas and computational implementations. He showed this by taking a fairly worked out linguistic theory of adjectival comparatives, embedding this into a question-answering system to answer questions such as "Is Bill Clinton taller than Bill Gates?" But as linguistic theories are not destructively tested for implementation purposes, first-order theorem provers lack facilities for procedural attachment and can't cope with higher-order constructs efficiently, this turned out to be both a demanding and interesting exercise.
After the coffee break Allan Ramsay took the stage and talked about how utterances can change the belief states of dialogue participants, but that intentions of the speaker are only achieved when the hearer has the right set of beliefs. He illustrated this with the following example: two birdwatchers are bored seeing pigeons all day so one of them decides to cheer up his fellow companion by means of a little sarcasm and asserting that there is an albatross. However, this speech act would not achieve its goal if they would not have the mutual belief that this assertion is false! Then Farid Nouioua presented a method for modelling natural language semantics using Answer Set Programming (a relatively new paradigm in logic programming). He applied this approach to the domain of descriptions of road accidents, with the aim of finding the cause of the accident. And just before the lunch break Rodrigo Agerri introduced an inference-based approach to interpret metaphorical use of language.
Lunch was offered by the site's own restaurant, and enjoyed in the impressive space under the great dome. Many of us were fascinated by the Foucault Pendulum, and it was given a big push.
While the Pendulum was swinging away, our second invited speaker, Christian Ebert, discussed two desiderata for underspecified representation formalisms. He came to the remarkable conclusion that no formalism can be expressively complete and avoid combinatorial explosion at the same time. Then Sela Mador-Haim described a system that uses controlled language for geographical information system queries, combining a categorial grammar with a compositional semantics based on lambda-calculus and SQL.
After the tea break, Hilke Reckman compared Davidsonian states with Kimian states with the aim of developing an elegant theory of entailment relations between verbs and their nominalisations. Satoshi Tojo refurbished a multi-dimensional modal logic for temporal expressions classifying events and states. The first day was closed by Ariel Cohen who formalised a system for anaphora resolution based on default logic and minimal models.
Friday, April 21st, 2006
The second day of the workshop started with an inspiring talk about harvesting knowledge required for inference in computational semantics, given by the third invited speaker Patrick Pantel. He introduced us to corpus-based methods applied to corpora of various sizes and to the challenges of fusing the acquired knowledge into semantic repositories. Then Lauri Karttunen and Rowan Nairn presented a strategy for detecting author commitment to the truth or falsity of complement clauses introduced by implicational verbs, claiming that this would be useful for automatic recognition of textual entailment. Remarkably, the theoretical fundaments for their approach were already established in an article written by Lauri in 1971, but they were implemented some 35 years later...
The rest of the morning saw the most dynamic part of the workshop: the poster and demo session! This was organised by a quick succession of five-minute flash talks were the presenters could advertise their poster, followed by an informal but intense discussion around the posters. Posters were presented by David Ahn on temporal question answering, by Philipp Cimiano on first-order logic and bridging anaphora, by Albert Goldfain on inference for arithmetic explanation, by Ingo Glöckner on semantic networks and inference, and by Paul Piwek on a theorem prover for dependent type systems.
One incident should not be left unnoted. On the last slide of his
flash presentation, Philipp summarised his theory of bridging with
quite a peculiar axiom. The audience, flabbergasted, couldn't believe
their eyes, but there it was, loud and clear: men are a subset of
women. While the female part of the audience, belonging to the minority,
applauded this proposal, the male fraction was left in disbelief and
demanded clarification. Yet Philipp stood by his theory up until the
last moment, when he could be convinced to introduce a negation
in his axiom...
After a well deserved lunch, we were ready for the last series of regular presentations. Nikhil Dinesh introduced formal verification techniques to provide a way to automatically determine consistency of regulatory documents. Patrick Pantel presented the Espresso system, an algorithm for extracting binary semantic relations aiming for high recall without losing much precision. Livio Robaldo argued that we ought to take branching quantifiers seriously and presented a formalism based on dependency tree semantics to model them. The last talk of the conference was given by Stefan Thater, who presented an algorithm for eliminating redundant readings from scopally underspecified semantic representations.
The workshop was closed with the formalities of the traditional Richard Montague Memorial Award for the best presentation delivered at ICoS, as voted by the participants. This was the third occurrence of the event since it was introduced by Patrick Blackburn in Siena at ICoS-3 (awarded to Alexander Koller). The second winner of this prestiguous award was Carlos Areces (at ICoS-4 in Nancy). And this year's winner is ... (drum rolls, etc.) ... Patrick Pantel, who got an overwhelming majority of the votes for his talk on harvesting knowledge from corpora. Congratulations, Patrick!
And just when Foucault's Pendulum arrived at a complete standstill, the workshop came to an end on a late Friday afternoon. And what do you do on a Friday night in Britain? Right! So off we went to one of the nicest pubs in Buxton for a pint (and another one, for some of us).
Group Picture
Upper row: Stefan Thater (Germany), Debby Field (UK), Paul "Pet" Piwek (UK), Rodrigo Agerri (UK), Allan Third (UK), Matthew Stone (UK), Sela Mador-Haim (Israel), Christian Ebert (Germany), Philipp Cimiano (Germany), Ian Pratt-Hartmann (UK). Middle row: Allan Ramsay (UK), Satoshi Tojo (Japan), Lauri Karttunen (US), Rowan Nairn (US), David Ahn (Netherlands), Nikhil Dinesh (US), Ingo Glöckner (Germany), Matthias Irmer (Germany), Livio Robaldo (Italy). Lower seats: Albert Goldfain (US), Ariel Cohen (Israel), Hilke Reckman (Netherlands), Steve Pulman (UK), Alexander Koller (Germany), Patrick Pantel (US), Johan "Muts" Bos (Italy).
>> back to Conference Reports
