We have been working on something to read contracts. It reads about 50,000 words, and accesses its knowledge structure about 200,000 times during the process of reading. The high usage, and the accuracy and precision required, mean that an ontology is quite unsuitable. What the system has read in the previous sentence, the previous clause, the previous page, all affect how it should interpret what it reads.
The high usage of the knowledge structure comes from
| Disambiguating anaphora |
| Unravelling prepositional chains (is it ((A of B) of C) or (A of (B of C))?) |
| Determining passive subjects or objects of ditransitive verbs |
| Determining the sense of a prepositional or noun phrase |
To throw some numbers around, if the knowledge structure was 99.9% accurate, it could still make 200 mistakes on average. While not all the mistakes would be disastrous to the meaning, out of the 200 there might be 10 that damaged the meaning, and ten is too many when a mistake can turn into financial loss.
An ontology which is invariant to what has already been read will have a poor reliability, even poorer if the document introduces its own internal definitions, which the ontology can know nothing about. It seems surprising that people put so much store by them.
Evolutionary ontologies are the next hope. In a document, you may be told something once. It is difficult to see how an evolutionary approach will be useful under the circumstances. Worse, evolution implies failure. What cost the failure before a successful evolution of meaning?