Merging specialist taxonomies and folk taxonomies in wordnets

Merging specialist taxonomies and folk taxonomies in wordnets - a case study of plants, animals, and foods in the Danish wordnet Bolette Sandford Pedersen, Sanni Nimb*, Anna Braasch University of Copenhagen, * Danish Society for Language and Literature

Outline • Introduction to Dan. Net § Food taxonomies in Dan. Net based on DDO: ideal cases and problems • Interrelating natural and functional taxonomies in Dan. Net • Conclusions LREC 2010 2

Dan. Net – a wordnet for Danish • Joint work: University of Copenhagen & Society for Danish Language and Literature • Monolingually-based approach: Danish lexical sources based on corpora: The Danish Dictionary (DDO); SIMPLE-DK • International frameworks: Princeton Word. Net, Euro. Word. Net • 60, 000 synsets ≈ concepts • 200, 000 semantic relations • Open source – download: wordnet. dk LREC 2010

Introduction DDO • Monolingual approach • Semi-automatic establishment of hyponym hierarchies by extraction of identified genus expressions in DDO LREC 2010 4

Food taxonomies in Dan. Net based on DDO: ideal cases and problems Princeton Word. Net: large number of e. g. fishes and vegetables, represented with both an animal or plant sense as well as a food sense -> same goal for edible animals and plants in Dan. Net: 2 hierarchies with a high number of synsets LREC 2010

Food taxonomies in Dan. Net based on DDO: ideal cases and problems Reuse of DDO, ideal cases : edible plants and animals should have two separate senses in DDO with two different genus expressions: 1= plant/animal. Biological genus expression = ”animal”, ”plant”, ”part of animal”, ”part of plant” … 2= food. Functional genus expression = ”food”, ”meat”, ”vegetable” … LREC 2010

Food taxonomies in Dan. Net based on DDO: ideal cases and problems (turkey) Genus expression sense 1= gallinaceous bird Genus expression sense 2 = meat LREC 2010

Food taxonomies in Dan. Net based on DDO: ideal cases and problems Study: 74 DDO lemmas with polysemy plant/animal and food Conclusion: More than half demand manual treatment. Why? (artichoke) (flower bud) (edible) artichoke: • only one sense • genus expression: flower bud • part of definition: edible, but no food sense established LREC 2010

Food taxonomies in Dan. Net based on DDO : ideal cases and problems 1. DDO = traditional -> zoological and botanical domains strongly influenced by professional descriptions: • Preference given to biological senses (in spite of corpus evidence) • Genus expression: Botanical/zoological terminology (tomato = fruit, spinach = herb, pheasant = gallinacean ) 2. First version of DDO - a printed dictionary -> food sense left out due to space limits 3. DDO not ontology-driven -> descriptions made ´bottom up’. -> No predefined vocabulary of definition -> No systematic treatment of regular polysemy LREC 2010

‘Physical’ versus ‘functional’ genus expressions • Observation: DDO definitions of concrete objects tend to use a ‘physical’ genus expression (object, stick , ball. . ) -> few ’functional’ genus expressions (toy, tool, equipment, food. . ) Examples, Princeton Word. Net: ‘ball’ = round object that is hit or thrown or kicked in games’ BUT hypernym in Word. Net = game equipment (NOT object) ‘doll = a small replica of a person, used as a toy’. BUT hypernym in Word. Net = toy (NOT replica) • Conclusion: Genus expressions are not fully reliable for the establishment of functional hierarchies LREC 2010

Interrelating taxonomies in Dan. Net Task of Dan. Net editors: • Establish regular polysemous synsets where missing in DDO in order to account for layman’s food perspective • Harmonize the different DDO approaches in a consistent way • Ensure that clash of taxonomies does not mess up the general structure of the network LREC 2010

Harmonize the different DDO approaches in a consistent way We strive towards individual, but interrelated taxonomies in Dan. Net: § Food taxonomy of meat, vegetables and fruit i. e. FUNCTIONAL KINDS in Cruse’s terminology (Cruse 2002) § Botanical/zoological taxonomies (in practice incomplete) i. e. NATURAL KINDS in Cruse’s terminology LREC 2010

Interrelating taxonomies Encoding strategy in Dan. Net: Synsets interrelate with the two taxonomies either § By means of regular polysemy Example: lamb_1 (animal), lamb_2 (meat: food) or § By means of multiple inheritance Examples: tomato (fruit/part-of-plant AND vegetable: food) rubarb (stalk/part-of-plant AND fruit: food) LREC 2010

Interrelating taxonomies Zoological Food animal meat lamb_1 lamb_2 Regular polysemy Food Botanical fruit vegetable tomato_1 LREC 2010

Not all concepts represented in both taxonomies Only in food taxonomy: § § § § Only in natural taxonomy: grøntsag (vegetable) rodfrugt (root vegetable) krydderurt (spice herb) suppeurt (potherb) vildt (game) flæsk (pork) indmad (offals) § § § plante (plant) skærmplante (umbelliferous plant) rod (tuber) stilk (stalk) indvolde (entrails) LREC 2010

Tricky cases (≈ false friends) Examples of these are such as § frugt (fruit) § nød (nut) § bær (berry) In principle, these cases should evoke two (unrelated) synsets, one in each taxonomy if you actually encode also the botanical ones LREC 2010

Excerpt from Food taxonomy (offals and meat) LREC 2010

Offals and their relations LREC 2010

Conclusions (1) § Foods taxonomies are typically folk taxonomies emerged spontaneously depending on the goods available and on cooking traditions of a particular region. § Inspired by and related to botanical and zoological taxonomies § The fact that terms are taken over from these natural taxonomies causes problems that require a consistent framework § Monolingual dictionaries do not always have a clear strategy LREC 2010

Conclusions (2) § We have developed a framework that enables to distinguish and interrelate between the natural taxonomies and the functional taxonomies of the network. § Even if you have monolingual lexical resources that are inconsistent in this respect, the monolingual starting point is important in order to capture correct conceptual structures of a given language Work in progress! See wordnet. dk andreord. dk LREC 2010