News

The LOGON on-line web interface is now open for public access. The site allows one to translate individual sentences (within the LOGON domain and vocabulary of backcountry tourism), including full access to intermediate representations. For comparison to other approaches to Norwegian–English MT, the demonstrator attempts to scrape further translations off the Internet. Computational resources are limited; be gentle on the demonstrator!

On February 8 and 9, the national KUNSTI funding scheme presents project results to the scientific and general public. The two-day conference is attended by approximately 50 participants, representing academia, industry, media, and the public sector. LOGON is represented in the program with an ‘open-microphone’ demonstration and overview poster, as well as with an in-depth summary presentation of major project results. On the second day of the conference, the nationwide newspaper Dagbladet prints a comprehensive summary of contemporary developments in machine translation, including a detailed discussion of LOGON (in Norwegian).

The 17th Computational Linguistics in the Netherlands conference invites Stephan Oepen to provide an overview of LOGON results to date. His presentation at CLIN highlights the utility of Efficient HPSG Realization for Precision Machine Translation.

The developers team delivers the final LOGON integration, release 1.0 (codenamed Fjell). Completing an intense, multi-site blood-toil-tears-and-sweat effort, the final system version arrives at 64.8 per cent end-to-end coverage on the Jotunheimen development corpus. Only hours following the final release, LOGON evaluators make available the (previously carefully protected) held-out test data from the same corpus, and it emerges that system performance on unseen data appears comparable to that on the development corpus: At 71.4 per cent coverage on the Jotunheimen test corpus with vocabulary known to the system and (an unexpectedly high) 56.1 per cent on test data with unknown vocabulary, there is little reason to fear that the final system is overly tuned to its 4000-sentence development corpus. Obviously, end-to-end coverage is but only one parameter in judging the final system version, where human judgements on translation quality are to be compiled by evaluators within the months to come.

Furthermore, looking at coverage evolution throughout the development cycle (compare to March 2005 below), it would seem that the earlier projection of a linear increase in development time was surprisingly accurate, if not underestimating system evolution. Looking at LOGON releases 0.5 through 1.0, a third-grade polynomial actually yields a better correlation than the original linear regression. Developers celebrate on the streets of Norway (and San Franciso). Sadly, with the completion of LOGON, there will not be further releases to determine how far the project can push this favourable pattern of incremental system improvement.

LOGON members Victoria Rosén and Paul Meurer continue the proud tradition of presenting at the annual Fifth International Workshop on Treebanks and Linguistic Theory (in Prague, Czech Republic). In joint work with the project TrePil at UiB, and accompanied by TrePil project manager Koenraad de Smedt, they present their paper Towards a Toolkit Linking Treebanking and Grammar Development.

LOGON stipendiary Lars Nygaard completes his MSc thesis Fra Ordbok til Ordnett (‘Dictionary to Word Net’) at the UiO Department of Linguistics and Nordic Studies. Supervised by Ruth Vatvedt Fjeld, the thesis develops a methodology for the acquisition of lexical semantic relations like synonymy or hyperonymy from a published dictionary of Norwegian.

During a one-week visit of Zhang Yi of Saarland University, he, Stephan, and Dan discuss the adaptation of the selective unpacking technique used in the LOGON generator (see October 2005) to the PET parsing engine. Only days after the completion of his visit to Norway, Zhang Yi releases an extended version of PET that, when applying the ERG to the LOGON corpus, yields more than a two-fold reduction in average parsing time (and up to a six-fold reduction in memory usage on complex inputs). Dan (and others) will now be so much more efficient in extending the ERG (or other DELPH-IN grammars).

Right after completion of his teaching duties at ESLLI 2006, Carl Jesse Pollard of the Ohio State University arrives at UiO for a five-week research visit. Carl works closely with Dan (and to a lesser degree Jan Tore and Stephan) on the syntax and semantics of degree specification (almost but not quite every summit), adverbs (nearly vs. often), derived adjectives (e.g. red-tailed), pronoun binding, and other topics of common interest. During his stay, Carl enjoys the Norwegian late summer and provides an advanced tutorial on his emerging framework of Higher Order Grammar.

LOGON is represented with no less than four project members at the Joint conference of the International Committee on Computational Linguistics and the Association for Computational Linguistics (ACL|COLING), held in Sydney, Australia; with its many hundreds of participants, ACL|COLING is the by far largest scientific gathering in our field (and yet remains among the most competitive conferences). LOGON results are represented in the conference program by virtue of an Interactive Presentation Re-Usable Tools for Precision Machine Translation (by Jan Tore Lønning and Stephan Oepen) and through the paper Statistical Ranking in Tactical Generation (by Erik Velldal and Stephan Oepen) that was accepted for EMNLP 2006.

During a one-week research visit of Micha Jellinghaus from Saarland University, Micha and LOGON developers at UiO produce a bi-directional German–English prototype instantiation of the LOGON MT architecture. Similar to on-going Japanese–English efforts, this strain of work is based on a pre-existing, MRS-enabled comprehensive HPSG implementation of German (GG). As part of his MSc thesis at Saarland University, supervised by Berthold Crysmann, Micha investigates the Automatic Acquisition of Semantic Transfer Rules for MT.

Just in time for a beautiful mid-summer night, Ivan Sag of Stanford University stops over at UiO as part of a visit to Europe. Ivan presents on his recent psycholinguistic investigations, Explaining Islands without Island Constraints, joint work with others at Stanford.

The LOGON consortium organizes the 11th Annual Conference of the European Association for Machine Translation. The two-day conference is held on the campus of the University of Oslo and attracts about one hundred partcipants. LOGON is represented in the scientific programme of the conference with two presentations, one on Identifying Complex Phenomena in a Corpus via a Treebank Lens (by Dan Flickinger), another on Using a Bi-Lingual Dictionary in Lexical Transfer (by Lars Nygård and others).

Immediately preceding the EAMT conference, LOGON members Dan Flickinger and Stephan Oepen host the Second Annual DELPH-IN Summit in the Norwegian mountains. A little over thirty DELPH-IN members from Europe, the US, Asia, and Australia gather for five days at Fefor Høyfjellshotell to exchange research results and make joint plans for on-going DELPH-IN activities.

LOGON stipendiary Arnt Richard Johansen completes his MSc thesis Human Benchmarks in Automatic Evaluation of Machine Translation at the NTNU Department of Language and Communnication Sciences. Supervised by Torbjørn Nordgård, the thesis reports on empirical results contrasting automatic MT evaluation metrics with the quality of translations produced by middle school language learners.

LOGON stipendiary Hanne Moa completes her MSc thesis Search Engines and Linguistics: With a Case Study of an Automated Compound Translator Using Search Engines. at the NTNU Department of Language and Communnication Sciences. Under the supervision of Torbjørn Nordgård, the thesis investigates the translation of Norwegian compounds into English.

Late in April, Francis Bond from the Communication Sciences Research Laboratory of Nippon Telegraph and Telephone Corporation (NTT, Japan) returns to UiO for a five-week stay as a LOGON guest researcher. Based on the LOGON technology, Francis is engaged in a joint effort with colleagues at the Nara Institute of Science and Technology to develop a bi-directional Japanese–English adaptation of the translation system.

At the semi-annual project gathering (held at Fefor), developers present the results of the latest LOGON integration, release 0.7 (codenamed Vår). Since the LOGON Jotunheimen development corpus became available in May 2005, this is the third release targeting this particular text collection and genre. Looking at the evolution of end-to-end coverage in releases 0.5 (codenamed Knut) through 0.7, it looks as if the consortium has accomplished a contiguous linear increase in end-to-end coverage. Developers are encouraged by this observation, though sceptical about their ability to hold up that ambitious trend throughout remaining integrations. Watch this space to see whether they will!

LOGON stipendiary Gisle Ytresøl completes his MSc thesis Universelle Representasjoner av Norske Klokkeslett (‘General Representations for Norwegian Times-of-Day’) at the UiB Department of Linguistics and Literature. Under the supervision of Helge Dyvik, this thesis proposes an MRS account of time-of-day expressions that is implemented within the LOGON analysis component.

LOGON stipendiary Gyri Smørdal Losnegaard completes her MSc thesis The Syntax and Semantics of Norwegian Numerals: An Analysis for MT at the UiB Department of Linguistics and Literature. Supervised by Helge Dyvik, the thesis develops a compositional account of the internal structure of complex number expressions that is integrated with the LOGON analysis grammar NorGram.

At the Second International Joint Conference on Natural Language Processing, John Carroll and Stephan Oepen receive the IJCNLP Best Paper award for their joint work on the LOGON generation component. Their paper High-Efficiency Realization for a Wide-Coverage Unification Grammar reviews a series of novel algorithmic contributions to chart generation, defines a procedure for selective enumeration of n-best results, and presents an in-depth empirical evaluation. IJCNLP 2005 was a highly competitive conference this year, accepting just above thirty per cent of the nearly 300 submissions that were received. Congratulations, John and Stephan!

At the Tenth MT Summit, held in Phuket (Thailand) this year, LOGON research was presented through a total of three presentations, viz. one by Dan Flickinger et al. on SEM-I Rational MT—Enriching Deep Grammars with a Semantic Interface for Scalable Machine Translation, one by Erik Velldal and Stephan Oepen on Maximum Entropy Models for Realization Ranking, and a joint contribution with colleagues from DFKI Saarbrücken (Germany), NTT Japan, and the University of Cambridge (UK) on Open Source Machine Translation with DELPH-IN.

LOGON researchers Dan Flickinger and Stephan Oepen co-organize the First Annual Meeting of the Deep Linguistic Processing with HPSG Initiative (DELPH-IN). The meeting, co-located with the International HPSG Conference in Lisboa (Portugal), had close to forty participants from all around the world, including (besides Dan and Stephan) another two LOGON members, viz. Lars Hellan and Petter Haugereid from NTNU.

Reflecting on joint work with colleagues at the Communication Sciences Research Laboratory of Nippon Telegraph and Telephone Corporation (NTT, Japan), LOGON members succeed in publishing results in two of the biggest conferences in our field, viz. at the Annual Meeting of the Association for Computational Linguistics in Ann Arbor (MI)—on High Precision Treebanking. Blazing Useful Trees Using POS Information—and at International Joint Conference on Artificial Intelligence in Edinburgh (UK)—on Robust Ontology Acquisition from Machine-Readable Dictionaries.

The International Lexical Functional Grammar Conference 2005 is hosted by, among others, LOGON researchers Helge Dyvik and Victoria Rosén between July 18 and July 20, 2005, at Bergen University. LOGON member Dan Flickinger gives an invited talk on Weighing the Utility of Types in Grammatical Description. Furthermore, LOGON is representend with a paper by Helge Dyvik, Victoria Rosén, and Paul Meurer on LFG, Minimal Recursion Semantics and Translation, and no less than three poster presentations.

Stephan Oepen, the Technical Manager of LOGON, is named an outstanding research talent and awarded a prize by the Faculty of Arts, the University of Oslo, at the event of celebrating the new name of the faculty. You can read more at the faculty pages or at the national research portal forskning.no (both in Norwegian).

Einar Stubhaug, LOGON stipendiary at UiO, completes his MSc thesis on Extracting Lexical Information on the Countability of Norwegian Nouns by Means of Corpus Methods. Einar was supervised by Jan Tore Lønning. Our congratulations go to both of them!

Project members have been very successful at publishing research results in international conferences this spring, including two presentations at the 2nd ACL SIGSEM Workshop on The Linguistic Dimensions of Prepositions (at Essex University, UK), a total of ten presentations (of which six were contributed by junior members of LOGON) at the 15th Nordic Conference of Computational Linguistics (in Joensuu, Finland), and one paper in the 10th Annual Conference of the European Association for Machine Translation (in Budapest, Hungary).

Late in April, Francis Bond from the Communication Sciences Research Laboratory of Nippon Telegraph and Telephone Corporation (NTT, Japan) arrives for a three-month stay as a LOGON guest researcher at UiO. Francis will work with other LOGON members on refining English generation (specifically regarding determiner choice and aspects of noun countability) and, in parallel, aims to combine LOGON resources with the DELPH-IN open-source grammar of Japanese to assemble a bi-directional Japanese–English MT system. Francis introduced himself at UiO by means of a colloquium presentation on The Hinoki Treebank.

While touring Europe, Timothy Baldwin from the Department of Computer Science and Software Engineering at Melbourne University stops over for a research visit with LOGON. While at UiO, Tim presents his work on Countability Learning: The Good, the Bad, and the Ugly.

As of April 1 (no jokes), Dan Flickinger joins the LOGON team at UiO as an associate research professor. Dan is the developer of the LinGO English Resource Grammar (ERG), the generation grammar for LOGON, and has taken a leave of absence from Stanford University to work with LOGON. Welcome, Dan!

Another LOGON stipendiary, Eszter Horvati, completes her MSc thesis at UiO. Supervised by Jan Tore Lønning, Eszter has investigated and evaluated methods for the Automated Acquisition of Norwegian Collocations. Congratulations!

LOGON stipendiary Fredrik Jørgensen is the first to complete his MSc thesis The Semantic Representation of Locatives in Machine Translation under the auspices of the project (supervised by Jan Tore Lønning and Stephan Oepen). Congratulations to Fredrik!

LOGON members present their research at the Third International Workshop on Treebanks and Linguistic Theory, viz. in the form of a paper on Paraphrasing Treebanks for Stochastic Realization Ranking (by Erik Velldal, Stephan Oepen, and Dan Flickinger) and as a poster and demonstration of TreeSearch. A User-Friendly Treebank Search Interface (by Lars Nygaard and Janne Bondi Johannessen).

LOGON is represented at two major MT conferences this fall, viz. The 6th Conference of the Association for Machine Translation in the Americas (AMTA) as part of the Research and Deployment Showcase exhibition and with an accepted paper at The 10th International Conference on Theoretical and Methodological Issues in Machine Translation, in Washington (DC) and Baltimore (MD), respectively.

The MT Research Group at Oslo University joins the Deep Linguistic Processing with HPSG Initiative (DELPH-IN) and contributes the LOGON transfer engine—a general-purpose MRS rewrite system—to the DELPH-IN open-source repository.

Oslo University offers two six-month stipends for MSc thesis projects conducted under LOGON auspices at the UiO Linguistics Institute. Applications including a resume and brief sketch of the proposed MSc project should be submitted by August 26, 2004.

Co-located at the 11th International Conference on Head-Driven Phrase Structure Grammar, Dorothee Beermann and Lars Hellan co-organize a Workshop on Semantics in Grammar Engineering. Dorothee and Lars present a paper on semantic decomposition in HPSG as part of the workshop, and LOGON doctoral researchers Liv Ellingsen and Petter Haugereid give presentations on their work on Norwegian in the main conference and workshop, respectively.

Norway swarms with international visitors this summer: Dan Flickinger from the Center for the Study of Language and Information at Stanford University (US) spends July and August at UiO and October at NTNU as a LOGON guest researcher.

Francis Bond, long-time member of the MT Research Group at Nippon Telegraph and Telephone Corporation (NTT, Japan), visits UiO for two weeks in July to assess the current LOGON prototype and explore adaptations of the technology to Japanese; Francis gives two presentations, one an overview of MT activities in Japan, another on Two Approaches to Generating Articles.

Towards the end of August, Rob Malouf (San Diego State University, CA), developer of the estimate Maximum Entropy toolkit used by LOGON stops over at UiO for one week; Rob gives a presentation on Mining Corpora for Linguistic Insights and chairs a session of the UiO Machine Learning reading group on Support Vector Machines.

Jointly with the EU-funded Deep Thought project, LOGON researchers organize the 3rd International Mini-Workshop on Efficient MRS-Based Generation at Fefor Høifjellshotell. Following the workshop, John Carroll (University of Sussex, UK) and Ann Copestake (Cambridge University, UK) visit UiO for a three-day ‘work camp’, which results in a speed-up to the LKB generator that appears to be exponential in input complexity.

LOGON researchers participate in the preparation of the Workshop on Recent Advances in Scandinavian Machine Translation. Helge Dyvik gives an invited keynote presentation on Semantics for Translation: The Possible Contribution of the Mirrors Method. Jan Tore Lønning and Lars Hellan provide an overview of LOGON results to date, and Torbjørn Nordgård presents a paper On Commercialization of Scandinavian MT (jointly with Tronn Skjerstad of LingIT AS).

With lots of help from Rolf, a first version of the re-worked and scalable LOGON web pages goes public. Many thanks, Rolf!

Liv Ellingsen and Erik Velldal are appointed as LOGON doctoral researchers for three years. Liv investigates the interaction of grammatical and soft constraints and Erik explores the acquisition and use of stochastic processes to manage transfer-level and generator ambiguity in MT.

Gunn Inger Lyse is appointed as a LOGON doctoral researcher for three years. Gunn pursues research on the utilization of semantic information acquired through the Semantic Mirrors method for automated word sense disambiguation.

John Carroll from the University of Sussex (UK) spends two weeks at UiO as a LOGON guest researcher. In joint work with Stephan Oepen, John improves LKB generation performance by two exponential factors.

Dan Flickinger from the Center for the Study of Language and Information at Stanford University (US) spends three months at NTNU as a LOGON guest researcher. Dan is the developer of the LinGO English Resource Grammar (ERG) used for English generation in LOGON; he adapts and tunes the grammar for the tourism domain and builds a treebank of annotated LOGON data.

Stephan Oepen is invited as a keynote speaker at the Second International Workshop on Treebanks and Linguistic Theory. Stephan presents on Treebanks versus Linguistic Theory. — Anyone Need Grammars?.

Petter Haugereid is appointed as a LOGON doctoral researcher for three years. Petter investigates the syntax – semantics interface in scalable constraint-based grammars, specifically in relation to the NorSource initiative on building a broad-coverage Norwegian HPSG implementation.