Background Document
Description:
The goal of GWA is to establish a Word-wide Association for maintaining, standardising and interlinking wordnets for all languages in the world, likewise preparing the ground for the development of a world-wide multilingual database with wordnets. The main objectives of GWA are:
- standardize the specification of lexical semantic relations, the notion of a synset and degrees of polysemy
- standardize the Inter-Lingual-Index for interlinking the wordnets of different languages
- extend the specification to include all Parts-of-Speech and multiword expressions
- develop a common XML representation for wordnet data
- prepare the development of sense-tagged corpora in all the linked languages.
- sharing and transferring of data, software and specifications across wordnets for different languages
- the development of guidelines and methodologies for building wordnets in new languages
- the development of explicit criteria and definitions for verifying the relations in any language
- the development of consistency checking, comparison and evaluation modules
Lexical semantic relations, synset and polysemy
The first objective involves the coordination and integration of the current initiatives: the Princeton WordNet and EuroWordNet. Their specifications will be compared and we will agree on a common specification for future releases of wordnets. This will include, among others, an extension of Princeton WordNet with references to the Inter-Lingual-Index so that it can directly be compared and mapped to the other wordnets in future.
Inter-Lingual-Index
A major issue of the project is the development of a standardized index of sense-distinctions or so-called Inter-Lingual-Index (ILI) that can be used as a universal and cross-lingual standard for sense differentiation. The current ILI, developed in EuroWordNet, is based on WordNet1.5 and is partially adapted to provide a better matching across the languages. On the one hand, it should be the superset of all concepts occurring in the different languages, on the other hand, it should have a certain granularity or generality of sense-differentiation. A similar adaptation of the sense-differentiation has been applied to WordNet1.6, and is aimed at in Corelex developed at Brandeis University. The granularity of WordNet1.5 is often discussed in relation to comparing and evaluating word-sense-disambiguation techniques. A major theme in the SENSEVAL/ROMANSEVAL competition has been the need for such a standardised list of sense-distinctions.
Some fundamental issues still have to be resolved, before the ILI can be agreed upon as index of word meanings. To what extent is regular polysemy valid across languages, and can it be solved centrally in the ILI, as it is currently done in EuroWordNet? In what sense should the ILI be the universal superset of concepts and what is the status of the concepts that are to be included? In GWA we will develop a fundamental approach to standardised ILI. The status of concepts in the index will be defined on the basis of criteria, such as:
- Universality: in how many languages does the concept occur?
- Usage: how frequently is the concept used across languages?
- Productivity: how easily can similar or related concepts be derived as new concepts?
- Morpho-syntactic markedness across languages: have words a morphological complex structure?
- Relations between index-records: can index records be seen as (productive) specializations or regular extensions of other index records?
- Ontological status of index-records: to which degree the concept is capable of distinguishing word senses in a minimally overlapping way.
These criteria will also clarify the different ways in which concepts can be lexicalized in languages. This will shed light on the cross-linguistic distribution of polysemy, derivation, compounding and multi words or phrases as ways of expressing concepts in languages. The resulting sense-structuring of ILI will be tested in Word-Sense-Disambiguation and Automatic-Document-Classification tasks for all the inter-linked languages. We will participate in international competitions such as multilingual TREC and SENSEVAL/ROMANSEVAL.
Another important aspect of the ILI is that it should be made possible to integrate other monolingual and ontological sources into the ILI (senses from other dictionaries like Hector and ontologies such as Mikrokosmos). This will create a platform for verification and comparison. Techniques for semi-automatically linking these different types of resources will have to be created and tested.
All this will amount to an extension of the present ILI architecture. The ILI will have a layered structure in which different levels will be distinguished. These levels may be the following:
1) resource specific concepts.
The architecture should allow sense distinctions from different resources to be incorporated as a contribution towards the standardisation of the notion of
concept/word sense. For language specific wordnets this will mean creating alternative linking systems involving different sets of ILIs.
2) clusters
Composite ILIs need to be typed according to the resource that is responsible for their clustering. Overlaps between the cluster types need to be established.
3) links between ILIs and ontologies (WN, Top Ontology, Mikrokosmos, Cyc…)
In this way the status of concepts can be defined and standardised in a bottom-up fashion, and GWA can be the multilingual platform for this standardisation effort and the integration/merge of ontologies.
All Parts-of-Speech and multiword expressions
The third objective is closely related to the above lexicalization phenomena. The current European wordnets are limited to nouns and verbs and hardly include phrases and collocations. However, it is obvious that meanings are not just expressed by nouns and verbs or single words. Languages use a variety of ways to express content. When we include a wider range of words and expressions we will also get a clear view on the existence of concepts in language-cultures and the ways in which these are expressed. We see for example that many lexicalized compounds in Germanic languages are often expressed as adjective-noun combinations, noun-preposition-noun combinations or derivations in Romanic languages. The possibility to include phrases, derivations and collocations in wordnet, either in a static way or dynamically, is essential for high-quality applications in the areas of Information Retrieval, Machine Translation, Language Generation and Language Learning applications. The wordnets will make it possible to reduce expressions to their conceptual content but also to generate all possible conceptual combinations of words that are compatible with such content (either within a language or across languages). Similar, it is evident that a language-resource needs to be able to generate derived forms when a direct equivalent for a concept is not present. Information on typical collocation restrictions within this set can then be used to select the typical and intuitive phrases out of all possible combinations. In the future, sense-tagged corpora that are related to the wordnets will become an essential component of the wordnets. GWA will prepare the development of these corpora in a uniform way. This involves the selection of comparable corpora and semantic tag sets.
Sense-tagged corpora in all languages
The Brown corpus with WordNet1.5 sense tags plays a crucial role in the development of Word-Sense-Disambiguation techniques. To both test the differentiation of senses across languages, and to evaluate the ILI as a fund of universal sense-distinctions it is therefore crucially important that similar corpora will be created for the other languages. These corpora can be tagged with senses from the local wordnets, but we will also tag them with ILI concepts, WordNet1.5 classes and the EuroWordNet top-ontology classes. The latter taggings can be used to compare the corpora, the sense codings and to test the Word-Sense-Disambiguation techniques across languages.
Software and Data Sharing across all languages
A standardised specification of the relations and a standardised list of sense-distinction in the form of the Inter-Lingual-Index will make it possible to share and compare both data and software modules across languages. Data sharing via the ILI will involve both language-neutral and language-specific data:
- definitions
- examples
- subject domains
- ontologies
- lexical relations
In the latter case, lexical relations are transferred from one wordnet to another wordnet. This can either be done for special relations, such as roles, meronymy or causal relations, but also for more fundamental relations such as hyponymy and synonymy, when wordnets are being built (see below).
Software that can be shared ranges from databases, editors, consistency checking software, conversion software, and hierarchy comparison software to conceptual distance and density measurements, corpus verification, word-sense-disambiguation and automatic term-extraction and classification. These modules, which are now available in different environments, will be wrapped up and standardised in a shared toolkit and evaluation module. This will make it possible to directly evaluate and test new wordnets that are to be integrated in the database.
Methodologies and guidelines for building wordnets
Currently, a number of methods and principles have been developed to build wordnets. A common approach is to start from an agreed set of important concepts, the so-called Base Concepts, which have to be represented in a local language and from which the core wordnets are being built. The shared top-ontology plays an important role as a common framework for guiding interpretation and formulating conceptual constraints on the possible relations. Furthermore, there is now a list of tests and definitions for semantic relations that can be extended and improved, especially in combination with the top-ontology and being supported by inheritance devices.
Extensions of the core wordnets can take place along different strategies, e.g.: an expand or a merge approach. In the case of an expand approach, a wordnet is translated and the lexical semantic relations are taken over. In the case of the merge approach, independent language specific structures are built first, which are translated to the index afterwards. For both approaches, there exist techniques to automatically map or translate wordnets to the ILI or to derive relations from regular definition patterns or morphological structures.
In GWA we will collect these techniques and write guidelines and methodologies that can be used by builders of wordnets for other languages. The methodologies will not only make the construction process more efficient but will also have a strong standardizing effect on the encoding of the relations.
Evaluation Module
Finally, some of the software and the definitions can directly be used to evaluate wordnets or fragments of wordnets. We will agree on a number of evaluation criteria to test the validity and compatibility of wordnets. In so far these criteria are implemented in software packages we will provide automatic evaluation mechanisms. Currently the wordnets are evaluated in terms of:
1. size of synsets, meanings, words (per POS)
2. density of Language Internal relations (per relation type, and per POS)
3. density of Equivalence relations (per relation type, per POS)
4. overlap in Index items covered
5. distribution over Ontology clusters
6. length of the hyperonymy chains
7. overlap of hyponym chains represented by the equivalent ILI-records
8. overlap and coverage with respect to corpora
Further criteria will be developed and combined in a single evaluation module.
GWA directly builds on the success of the Princeton WordNet and EuroWordNet. It will fulfil a direct need for multilingual semantic resources and will make it possible to share domain-ontologies across languages or upgrade ontologies to multilingual dimensions.
On a longer term, GWA aims at the development of wordnets for all languages in the world and to extend the wordnets to full coverage and all parts-of-speech. We will investigate what regulations and organisation are needed to support this. We will start co-operations with institutes and national science foundations that are willing to develop these wordnets or extend existing wordnets with the framework of GWA.