Automatic Abstracting of Textual Material

AUTOMATIC ABSTRACTING OF TEXTUAL MATERIAL

Stephen L. Taylor Wichita State University Wichita, Kansas 67208

Gilbert K. Krulee Northwestern University Evanston, Illinois 60201

Lawrence T. Henschen Northwestern University

Evanston, Illinois 60201

In what follows, we want to describe a sys­ tem for the automatic abstracting of textual ma­ teriaL In designing the system, major theoreti­ cal questions have arisen not unlike those that arise in dealing with any natural language sys­ tem. In addition, we want to describe some pro­ posed revisions which have important theoretical implications and which should lead to significant improvements in the capabilities of the present system.

Our initial efforts at automatic abstracting began in 1969 as part of a more general question­ answering system (Tharp, 1969; Tharp and Krulee, 1969) which made use of short stories about fa­ mous discoveries taken from a children's encyclo­ pedia. The original text was mapped into a set of predicates representing the logical content of the story. Although the main emphasis in this system was on question-answering, there were two questions that could be asked that made use of a summarizing and abstracting capability. These were: "Who is the central character in the

story?" and "What is the main theme of the story?"

In our current efforts (Taylor, 1975), we have extended Tharp's methods in order to deal more explicitly with the problem of automatic ab­ stracting. The system represents the meaning of a text in terms of the semantic networks of Simmons (1973) which are based on the case grammar rela­ tionships of Fillmore (1968). In this form of representation, the nodes are words or concepts while the arcs represent a case grammar relation­ ship that exists between a pair of nodes. Using the graphical techniques of Ramamoorthy (1966), the system identifies a portion of the original network, namely, the maximally connected subgraph (or graphs). Then using the techniques of signal flow graph analysis, the system identifies nodes that are most influential within the maximally connected subgraph. By proceeding iteratively, using this pair of techniques, a subgraph is ob­ tained which serves as an abstract of the origi­ nal text. As a final step, again using a tech­ nique due to Simmons and Slocum ·(1972), the sub­ graph is converted back into a set of natural lan­ guage sentences as the final output.

The results of these initial attempts are reasonably encouraging although certain practical difficulties have been encountered. For example, even a short sample of text (one or two pages) leads to the formation of a complex network that is difficult to store. However, computer time for the processing of these networks is not excessive (less than ten seconds on a CDC 6400) and limita­ tions of space are more serious than of computer time.

During the past year, we have introduced ~ series of modifications into the original system (Lindner, 1976). One of these has to do with texts that have multiple themes, such as a main theme and some subsidiary themes. In our original system, the first part of the process leads to the identification of several maximally connected sub­ graphs (MSC's). However, only one of these--the largest--is chosen as the basis for developing an abstract. This subgraph usually does contain the main theme plus some related details. A secondary theme may well be contained in a second MCS. Thus, by choosing only a single MCS, one biases the system towards the development of an abstract that overemphasizes details relating to the main theme while ignoring secondary themes.

Accordingly, we experimented with longer mul­ tiple-theme texts and with a procedure that would select the largest MCS plus one or more additi­ tional MCS's. The resulting abstract appears to be much improved. For example, using as a text a book review, three important MCS's were identi­ fied, the first concerned primarily with a dis­ cussion of the author and the second and third with the main character of the book and the theme of the novel, respectively. Thus, an abstract making use of all three leads clearly to a much more balanced presentation than does an abstract of equal length that refers only to the author.

Our original program was unable to handle networks of more than 300 nodes. Basically, the revised technique deals with the text, paragraph by paragraph. As a first step, the network associated with each paragraph is reduced, one at a time, thus obtaining a sequence of reduced net­ works, one from each paragraph. These networks are recombined into a single network and converted back into an abstract. These abstracts obtained are not unlike those obtained through the use of the original method. Most importantly, with this modification, it is the length of each paragraph that is critical rather than the length of the text as a whole, although, as the number of para­ graphs increases, one again runs into difficulties in storing the reduced networks for each of a series of paragraphs.

We now find ourselves faced with the problem of introducing modifications that would lead to a significant improvement in the capabilities of this abstracting system. We propose to introduce some radical changes of which the following are perhaps the most significant.

1. In many respects, networks representative of the meaning of a text make use of logical predi­ cates as the basic semantic units and the higher­ level semantic unit that is computed is the prod­ uct of a "bottom-up" form of analysis with the propositions being related in pairs until the overall network emerges as a final product. As an alternative, we propose that the higher-level se­ mantic analysis should be "top-down" in the sense of making predictions about the overall thematic structure of the material being processed. More­ over, this analysis should make use of a higher level or semantic grammar much like the thematic grammar that Rumelhart (1975) has proposed for the thematic analysis of children's stories.

2. Secondly, we are assuming that abstracting is normally a dynamic process that should be primar­ ily logical or qualitative in form and making use of what we might refer to as substitution or con­ densing operators. Formally, these operators might resemble an axiom or theorem in mathematics, stating that certain content can be substituted for certain other content. These operators will "condense" in the sense of taking a network repre­ senting a portion of text and replacing it with a simplified network or perhaps a single node. Sometimes, the basis for condensing the text is made explicit in the text, in which case the oper­ ator is not unlike a procedure for identifying certain types of phrases in co~text. For example, one often encounters sentences in the form: "There are two main methods for the production of . " or "When constructing a one immediately faces three types of problems." Thus, in the abstract, one wants to identify in one case the two main methods and in the other case the three types of problems. Moreover, one probably wants to name the methods or problems while ignor­ ing all of the amplifying details. Unfortunately, under many circumstances, such strong clues may not be given explicitly and must be inferred in much the same way that answers not explicitly con­ tained in a data base can be inferred by problem solving or theorem proving methods. Thus, in our revised system we want to include a capability for "proving" that a set of summarizing statements can

be inferred from the original data base.

In short, we are proposing two major modifi­ cations to our present abstracting program in order to make the system ?erform in a m01:e "human­ like" fashion and in order to develop a system with a significantly improved level of competence.