Acquisition of historical knowledge from encyclopedic texts
Richard Hull · Journal of International Crisis and Risk Communication Research · 1997
Automatic acquisition of historical knowledge from encyclopedic texts involves addressing three issues in the field of artificial intelligence: natural language understanding, knowledge representation, and knowledge acquisition. Semantic interpretation, a component of natural language understanding, is responsible for the construction of logical forms from syntactic relations produced by the parser. Nominalized verbs, or nominalizations, contribute the same knowledge as their verbal counterparts, and behave like their verbal forms in that they license prepositional phrases. The nominalizations found in these texts are often ambiguous. Unlike previous treatments of nominalizations which skirted the ambiguity issue, the problems of choosing between the verbal and non-verbal senses of nominalizations and handling the polysemy of the nominalized verb were addressed head on. Another key principle for semantic interpretation is the the creation of verbal knowledge, in the form of verb meaning (VM) rules and verbal concepts. Both of these constructs require a hierarchy of concepts to specify selectional restrictions. Over 220 new VM rules and 140 new verbal concepts were written, covering a wide range of relations. Testing of the system with this verbal knowledge and its supporting ontology showed that the correct sense of the verb and the correct attachment and meaning for prepositional phrases were determined more than 90% of the time. An investigation into how to recognize semantic connections between sentences and how to use those connections to acquire implicit knowledge was performed. Temporal connections are particularly important because they provide a foundation for acquiring other, more challenging relations such as causal connections. Representations of temporal and causal relations are presented along with methods for automatically acquiring them. These algorithms were tested on articles of the electronic version of the World Book Encyclopedia. An application was created which accepts questions about historical figures and using knowledge acquired from encyclopedic articles, answers them. This application, called SNOWY-BIOS, was tested on ninety-one different English questions. The answers produced by the system were compared against answers produced by two human subjects using the same version of the encyclopedia. Recall and precision scores of 54% and 85% respectively were found when compared against the humans' performance.