A Unified Theory of Inference for Text Understanding

Natural languages, like English, are difficult to understand not only because of the variety of forms that can be expressed, but also because of what is not explicitly expressed. The problem of deciding what was implied by a text, of "reading between the lines" is the problem of inference. For a reader to extract the proper set of inferences from a text (the set that was intrended by the text's author) requires a great deal of general knowledge on the part of reader, as well as a capability to reason with this knowledge. When the reader is a computer$\dots$ Past approaches to the problem of inference have often concentrated on a particular type of knowledge structure (such as a script) and postulated an algorithm tuned to process just that type of structure. The problem with this approach is that it is difficult to modify the algorithm when it comes time to add a new type of knowledge structure. An alternative, unified approach is proposed. This approach is formalized in a computer program named FAUSTUS. The algorithm recognizes six very general classes of inference, classes that are not dependent on individual knowledge structures. Rather, the classes describe very general kinds of connections between concepts. New kinds of knowledge can be added without modifying the algorithm. Thus, the complexity has been shifted from the algorithm to the knowledge base. To accommodate this, a powerful knowledge representation language named KODIAK is employed. The resulting system is capable of drawing proper inferences (and avoiding improper ones) from a variety of texts, in some cases duplicating the efforts of other systems, and in other cases improving on them. In each case, the same unified algorithm is used, without tuning the program specifically for the text at hand.