6.5 Vector Space Retrieval of Knowledge Representations 147
you can extract from them. Concept combination could be a powerful in-
formation retrieval mechanism, provided it is properly interpreted. With a
relatively small number of basic concepts along with a small number of con-
ventional relationships, one can construct a very large number of concept
combinations.
Concept combination, also calledconceptual blendingandconceptual inte-
gration, is an active area of research in linguistics. The meaning of a concept
combination requires a deeper understanding of the relationship between
words and the phenomena in the world that they signify. Based on the earlier
work of Peirce, de Saussure, and others in the field of semiotics, Fauconnier
and Turner (1998, 2002) have developed a theory of conceptual blending that
explains how concepts can be blended. However, this theory is informal.
Goguen has now developed a formal basis for conceptual blending (Goguen
1999; Goguen and Harrell 2004). Furthermore, Goguen and his students have
developed software that automates the blending of concepts, and their sys-
tem has been used to understand and even to create poetry and other narra-
tives. Concept combination is closely connected with human categorization
and metaphor. For an entertaining account of these topics, see Lakoff’s book
with the intriguing titleWomen, Fire and Dangerous Things: What Categories
Reveal about the Mind(Lakoff 1987).
The tool developed by Goguen and his students, mentioned above, is ca-
pable of finding a wealth of concept combinations even when the concepts
are relatively simple. For the words “house” and “boat,” their tool finds
48 complete blends and 736 partial blends. Two of these have become so
common that they are considered single words; namely, “houseboat” and
“boathouse.” Others are less obvious, but still make sense, such as a boat
used for transporting houses, an amphibious recreational vehicle, or a boat
used permanently on land as a house.
As one might imagine, the combinatorial possibilities for combinations be-
come enormous when there are more than two words being combined. A
typical title for a biomedical research article can have a dozen words. Un-
derstanding the meaning of such a title can be a formidable undertaking if
one is not familiar with the subject matter of the article, as we pointed out in
section 1.6. Goguen and Harrell (2004) pointed out that conceptual blending
alone is not sufficient for understanding entire narratives that involve many
such blends (Goguen and Harrell 2004). They introduced the notion ofstruc-
tural blending, also calledstructural integration, to account for the meaning of
whole documents.
Having introduced concept combinations, one still has the problem of how