1. Building a Large- Scale Knowledge Base for Machine Translation Kevin Knight and Steve K. Luk Presenter: Cristina Nicolae

2. Linguistic resources combined into PANGLOSS PENMAN Upper Model (Bateman 1990) –top-level network of 200 nodes implemented in the LOOM KR language –makes extensive use of syntactic-semantic correspondences (taxonomy  grammar) ONTOS (Carlson & Nirenburg 1990) –top-level ontology designed to support machine translation Longman’s Dictionary (LDOCE) –words with definition, usage, syntactic code ( [B3] for adj+to), semantic code ( [H] for human), pragmatic code ([ECZB] for economics/business) WordNet (Miller 1990) –semantic word database Collins Bilingual Dictionary –Spanish-English dictionary

3. Merging resources

4. Merging resources – contributions LDOCE: syntax and subject area WordNet: synonyms and hierarchical structuring the upper structures: organize the knowledge for NLP in general and the English generation in particular the bilingual dictionary: lets us index the ontology from a second language

5. two word senses should be matched if their two definitions share words looks also at related words and senses (e.g. synonyms) LDOCE (batter_2_0) “mixture of flour, eggs and milk, beaten together and used in cooking” (batter_3_0) “a person who bats, esp. in baseball – compare BATSMAN” WordNet (BATTER-1) “ballplayer who bats” (BATTER-2) “a flour mixture thin enough to pour or drop from a spoon” Match: –(batter_2_0) with (BATTER-2) –(batter_3_0) with (BATTER-1) Definition Match Algorithm

6. Definition Match Algorithm – Results low ambiguity wordshigh ambiguity words Ran algorithm on all nouns from LDOCE and WordNet.

7. Hierarchy Match Algorithm uses sense hierarchies inside LDOCE and WordNet once two senses are matched, it is a good idea to look at their respective ancestors and descendants for further matches Match: –animal_1_2 with ANIMAL-1 –and their respective animal-subhierarchies start with unambiguous words and match them, then look downward and upward in the hierarchies rooted at them and match those too

8. Hierarchy Match Algorithm – Results In the end, the algorithm produced 11,128 noun sense matches at 96% accuracy.

9. Bilingual Match Algorithm goal is to annotate the ontology with a large Spanish lexicon from: –mappings between Spanish and English words (from Collins) –mappings between English words and ontological entities (from WordNet) –conceptual relations between ontological entities we obtain: –direct links between Spanish words and ontological entities

10. Discussion each merge algorithm presented above is verified by humans afterwards (humans are faster at verifying info than generating it from scratch) semi-automatic merging brings together complementary sources of information also allows us to detect errors and omissions where resources are redundant