Wordnets are increasingly widely used to model word meaning in natural language processing tasks. However, there are still many challenges in accurately describing word meanings and making these descriptions useful for both human and machine users. This workshop aims to identify, discuss and start to solve existing challenges. The idea grew from the CLARIN Workshop - Towards Interoperability of Lexico-Semantic Resources1 held at Tartu (2017-01). Wordnets are one of the most common semantic lexical resources and exist now for a large number of languages. Many are constructed manually or under careful manual control. Thus their quality is much higher than the quality of resources extracted automatically. Wordnets can provide a core for Linked Open Data that anchors them in language resources. Diferent wordnets are generally similar but still need some efort to combine in a common interoperable multilingual framework. It can be a first step and valuable case study in developing interoperable solutions for semantic lexical resources. In addition to the Princeton WordNet of English, there are several large wordnets built for diferent natural languages, e.g. BulNet (Bulgarian), Czech Wordnet, DanNet, Dutch Open Source WordNet, enWordNet (English, a significant extension to Princeton WordNet), Estonian WordNet, Finish WordNet, GermaNet, MultiWordNet, plWordNet (Polish), sloWNet, as well as several semantic resources similar to or linked to wordnets. Many of them are very large and provide extensive coverage of vocabulary. All of them are somehow mapped to diferent versions of Princeton WordNet. There are also initiatives on building multilingual platforms linking several wordnets, e.g. Open Multilingual Wordnet. Some wordnets are connected to Linked Open Data that creates a possibility for anchoring named entities, their relational descriptions and ontologies in the natural language lexical semantics. Thus, 1https://www.clarin.eu/event/2017/ clarin-workshop-towards-interoperability-lexico-semantic-resources
wordnets already form a rich set of lexical semantics resources that can be a case study in building interoperable solutions in semantic multilingual lexical resources for the needs of semantic interoperability on several levels: the infrastructure, meta-data, language tools and research applications. However this potential has not been fully explored yet. The main problems that we have identified before the workshop are: • variety of formats for wordnets, • small but significant diferences between wordnet models that results in problems of mapping wordnet relations across language and makes their full use in multilingual applications more dificult, • lack of a common set of applications for wordnet across diferent languages, e.g. for storing and accessing full content of wordnets (including their use as a reference set), text-to-sense mapping, word and text similarity calculation, etc.
The main challenges for wordnet developers seem to be: • common formats for wordnets with their whole richness of semantic knowledge represented and models applied, • inter-operability on the level of the interpretation of wordnet models (correspondence of basic elements and diferent types of relations), • requirements for a basic set of wordnet-related tools, web services and applications, • development of wordnet browsing services or inter-operable services, • common framework for linking wordnets with Linked Open Data or domain knowledge resources for the needs of applications, • gathering and sharing good practices in utilising wordnets and similar semantic resources in applications.
The program for this workshop was a little unusual. First, we asked participants to read the papers ahead of the workshop. Then the workshop itself was split into two sessions. In the first session, each participant had 15 minutes to highlight one or two of the challenges they considered most important, possibly with an outline of a solution. In the second session, we summarized the first session and then followed up with a long moderated discussion with all participating. Here we try to come up with a deeper understanding of the problems and sketch some possible solutions. Brief notes from the summary are included in this preface.
All the papers, the summary and notes from the discussions are online at: http:// clarin-pl.eu/workshop/wn-challenges.html. We would like to thank CLARINPL2 research infrastructure and the Global WordNet Association Board for their support.
Maciej Piasecki, Wrocław University of Science and Technology Francis Bond, Nanyang Technological University Jan Wieczorek, Wrocław University of Science and Technology
Eneko Agirre, University of the Basque Country Francis Bond, Nanyang Technological University Sonja Bosch, University of South Africa, Pretoria Christiane D. Fellbaum, Princeton University Darja Fišer, University of Ljubljana Antoni Oliver Gonzalez, Open University of Catalonia Shu-Kai Hsieh, National Taiwan University John P. McCrae, Insight Centre for Data Analytics, National University of Ireland, Galway Verginica Mititelu, Romanian Academy Monica Monachini, National Research Council of Italy Adam Pease, Articulate Software Bolette Sandford Pedersen, University of Copenhagen Maciej Piasecki (chair), Wrocław University of Technology Alexandre Rademaker, IBM Research German Rigau, Polytechnic University of Catalonia Ewa Rudnicka, Wrocław University of Technology Shikhar Kr. Sarma, Gauhati University Stanisław Szpakowicz, Emeritus Professor, University of Ottawa Veronika Vincze, University of Szeged Piek Th. J. M. Vossen, VU University Amsterdam
Time 13:00‒13:15 13:15‒13:30 13:30‒13:45 14:30‒14:45 14:45‒15:00 15:30‒16:00 16:00‒18:40 18:40‒19:00
Title Welcome How Stable are WordNet Synsets? Inside Baseball: Coverage, quality, and culture in the Global WordNet Two corpus based experiments with the Portuguese and English Wordnets Overview and Future of Czech Wordnet Challenges behind the data-driven Bulgarian WordNet (BulTreeBank Bulgarian Wordnet) The Revision History of Estonian Wordnet Towards Revised System of Verb Wordnet Relations for Polish Classification of Adjectives in BulNet: Notes on an Efort The Concept of Lexical Platform
Maciej Piasecki Eric Kafe Martin Benjamin Fabricio Chalub, Alexandre Rademaker and Cláudia Freitas Adam Rambousek, Karel Pala and Sandra Tukacova Petya Osenova and Kiril Simov Neeme Kahusk and Kadri Vider Agnieszka Dziob, Maciej Piasecki, Marek Maziarz, Justyna Wieczorek and Marta Dobrowolska-Pigoń Tsvetana Dimitrova and Valentina Stefanova Maciej Piasecki, Tomasz Walkowiak, Ewa Rudnicka, Tomasz Naskręt and Francis Bond Francis Bond (Moderator) Francis Bond and Maciej Piasecki
by Francis Bond (based on input from all present) Outline These very rough notes are an attempt to list some of my (and Ed Hovy’s) experiences; discuss a little what we are doing with the OMW; and summarize the main issues raised by the various papers. Views given here are not necessarily shared by all the participants (but we all seemed in general agreement). They were shown as slides to start the discussion and have been reproduced here with minimal changes.
• Some problems identified in Ontonotes • Some problems identified by NTU-MC • Some solutions with the OMW • Summary – Technical – Social – Financial Ontonotes Problems • Technical • Social • Missing senses
– hard to add new terms NTU Multilingual Corpus Problems – Missing senses
named entities – Indistinguishable senses hard to distinguish reliably merged things until IA was 80% hard to tell if important information was lost – especially MWEs and cultural words – non-english senses: gohan “cooked rice” kome “rice grains” – want to add multiple languages at once – small variations are common (sg/pl ’-/space’: night bird vs nightbird) • Indistinguishable senses – distinguishing information may not be seen easily – verb frames, examples, network, … – some things (40 or so) are no distinguishable — merge • Missing POS: pronouns, classifiers, greetings, exclamations, … OMW Solutions • Link projects with a common shared structure – document online and link to database (*) – link through the CILI
also to other resources (SUMO, …) – allow multiple wordnets for the same language
remove bottlenecks, encourage domain resources – more information about senses and variants
allow orthographic variants, label, e.g. scientific names – link to corpora – rank using sense frequencies/automatically created ontologies – retain and publish statistics on usage and downloads (*) – allow online feedback and comments (*) – more tools for checking definition checking, suggest synonyms, panlex, … graph checks, corpus annotation, … Permanent Identifiers • Should we have sense-keys as well as synset identifiers? • Will we use the lexicographer files any more?
– Will PWN keep using these? (yes) • How about version control for the whole wordnet? can we standardize on major.minor.patch? Quality Control • How can we get feedback for input regarding quality and selection? – user feedback – expert feedback • Can we link to live examples? • Can we copy structure more easily from other wordnets, … • What is the best way to build a wordnet? – merge, transfer, some combination?
automatic sources (from wiktionary, PanLex, …) • Meanings change over time: how can we represent this? Named Entities • Should we have them or take them out? • What about fictional names?
– We need them in the Sherlock Holmes Corpus • Can we get them from external resources – Geonames – Species – DBpedia – UMLS Domains and Productive Meaning Extensions • Domains – Should we move them into sub wordnets? – Or just add better domain markers? – Should we keep supertypes (lexicographer files)? • Productive Meaning Extensions how should ee handle these – quick to quickly – actor to actress – interest to interested – clasp to unclasp Definitions and Examples • It is hard to write a good definition? • Should it be unique? Ontology Compatibility • Should it have redundant info from the links (e.g. domain)?
possibly show in diferent interfaces • Should it be disambiguated? • What do we do if it changes? • Can we use wordnet as an ontology? • What are the diferences between WN and Ontologies? • The same thing is called something diferent by diferent groups • Several solutions to similar problems – project to a shared vocabulary (OMW) – document this (calling for volunteers) – lexical gaps, phrasesets, artificial synsets, • Mapping – should we allow a variety of links to CILI? – not just equal-synonymy Social • Restrictive licenses are restrictive Yes Yes Yes Yes No – Unless you can redistribute the data, it is hard to make a large compatible collection, … – Unless the licenses are compatible it is illegal to merge, … • How do we get feedback from end users/applications? e.g. If Kamusi adds something, how does it get back to the WN • How do we get acknowledgments back to the original wordnets
e.g. If someone looks it up in OMW/NLTK/Babelnet how is credit fed back? • We don’t have standard tools that we all use so duplication of efort can we have a fixed data-specification so we can use any tool Financial • It is hard to get money to maintain wordnets
– How can we make things easier to maintain? • How can we persuade funders? • How can we use funding/efort most efectively?