Automatic Abstracting of Textual Material

Our i n i t i a l e f f o r t s a t automat ic a b s t r a c t i n g began in 1969 as p a r t of a more genera l q u e s t i o n answering system (Tharp , 1969; Tharp and K r u l e e , 1969) which made use of shor t s t o r i e s about f a mous d i s c o v e r i e s taken from a c h i l d r e n ' s encyc lo p e d i a . The o r i g i n a l t e x t was mapped i n t o a set of p r e d i c a t e s r e p r e s e n t i n g the l o g i c a l content of the s t o r y . Although the main emphasis in t h i s system was on q u e s t i o n a n s w e r i n g , t h e r e were two quest ions t h a t could be asked t h a t made use of a summarizing and a b s t r a c t i n g c a p a b i l i t y . These were : "Who is the c e n t r a l c h a r a c t e r in the s t o r y ? " and "What is the main theme of the s tory?" In our c u r r e n t e f f o r t s ( T a y l o r , 1 9 7 5 ) , we have extended Tharp 's methods in order to dea l more e x p l i c i t l y w i t h the problem of automatic abs t r a c t i n g . The system represents the meaning of a t e x t in terms of the semantic networks of Simmons (1973) which are based on the case grammar r e l a t i o n s h i p s o f F i l l m o r e ( 1 9 6 8 ) . In t h i s form o f r e p r e s e n t a t i o n , the nodes are words or concepts w h i l e the arcs represen t a case grammar r e l a t i o n sh ip t h a t e x i s t s between a p a i r of nodes. Using the g r a p h i c a l techniques of Ramamoorthy ( 1 9 6 6 ) , the system i d e n t i f i e s a p o r t i o n of the o r i g i n a l ne twork , namely, the maximal ly connected subgraph (or g r a p h s ) . Then us ing the techniques of s i g n a l f l ow graph a n a l y s i s , the system i d e n t i f i e s nodes t h a t are most i n f l u e n t i a l w i t h i n the maximal ly connected subgraph. By proceeding i t e r a t i v e l y , us ing t h i s p a i r of t e c h n i q u e s , a subgraph is obt a i n e d which serves as an a b s t r a c t of the o r i g i n a l t e x t . As a f i n a l s t e p , aga in us ing a t e c h nique due to Simmons and Slocum ( 1 9 7 2 ) , the subgraph is conver ted back i n t o a se t of n a t u r a l l a n guage sentences as the f i n a l ou tpu t . The r e s u l t s o f these i n i t i a l a t tempts are reasonably encouraging a l though c e r t a i n p r a c t i c a l d i f f i c u l t i e s have been encountered. For example, even a shor t sample of t e x t (one or two pages) leads to the f o r m a t i o n of a complex network t h a t i s d i f f i c u l t t o s t o r e . However, computer t ime f o r the processing of these networks is not excessive ( l e s s than ten seconds on a CDC 6400) and l i m i t a t i o n s of space are more ser ious than of computer t i m e . Dur ing the past y e a r , we have in t roduced a s e r i e s o f m o d i f i c a t i o n s i n t o the o r i g i n a l system (L indner , 1 9 7 6 ) . One of these has to do w i t h t e x t s t h a t have m u l t i p l e themes, such as a main theme and some s u b s i d i a r y themes. In our o r i g i n a l system, the f i r s t p a r t o f the process leads to the i d e n t i f i c a t i o n o f s e v e r a l maximal ly connected subgraphs (MSC's) . However, only one of t h e s e — t h e l a r g e s t i s chosen as the bas is f o r developing an a b s t r a c t . This subgraph u s u a l l y does c o n t a i n the main theme plus some r e l a t e d d e t a i l s . A secondary theme may w e l l be conta ined in a second MCS. Thus, by choosing on ly a s i n g l e MCS, one b iases the system towards the development of an a b s t r a c t t h a t overemphasizes d e t a i l s r e l a t i n g t o the main theme w h i l e i g n o r i n g secondary themes. Accord ing ly , we experimented w i t h longer mul t i p l e t h e m e t e x t s and w i t h a procedure t h a t would s e l e c t the l a r g e s t MCS plus one or more a d d i t i t i o n a l MCS's. The r e s u l t i n g a b s t r a c t appears to be much improved. For example, using as a t e x t a book r e v i e w , t h ree important MCS's were i d e n t i f i e d , the f i r s t concerned p r i m a r i l y w i t h a d i s cussion of the author and the second and t h i r d w i t h the main c h a r a c t e r of the book and the theme of the n o v e l , r e s p e c t i v e l y . Thus, an a b s t r a c t making use of a l l t h r e e leads c l e a r l y to a much more balanced p r e s e n t a t i o n than does an a b s t r a c t o f equal l ength t h a t r e f e r s only to the au thor . Our o r i g i n a l program was unable to handle networks of more than 300 nodes. B a s i c a l l y , the r e v i s e d technique deals w i t h the t e x t , paragraph by paragraph. As a f i r s t s t e p , the network assoc ia ted w i t h each paragraph is reduced, one at a t i m e , thus o b t a i n i n g a sequence of reduced n e t works, one from each paragraph. These networks are recombined i n t o a s i n g l e network and conver ted back i n t o an a b s t r a c t . These a b s t r a c t s obta ined are not u n l i k e those obta ined through the use of the o r i g i n a l method. Most i m p o r t a n t l y , w i t h t h i s m o d i f i c a t i o n , i t i s the l eng th o f each paragraph t h a t i s c r i t i c a l r a t h e r than the length o f the t e x t as a whole , a l t h o u g h , as the number of p a r a graphs i n c r e a s e s , one aga in runs i n t o d i f f i c u l t i e s in s t o r i n g the reduced networks f o r each of a s e r i e s of paragraphs. We now f i n d ourselves faced w i t h the problem of i n t r o d u c i n g m o d i f i c a t i o n s t h a t would lead to a s i g n i f i c a n t improvement i n the c a p a b i l i t i e s o f t h i s a b s t r a c t i n g system. We propose to in t roduce some r a d i c a l changes of which the f o l l o w i n g a re perhaps the most s i g n i f i c a n t . 1 . In many r e s p e c t s , networks r e p r e s e n t a t i v e o f the meaning of a t e x t make use of l o g i c a l p r e d i cates as the bas ic semantic u n i t s and the h i g h e r l e v e l semantic u n i t t h a t i s computed i s the p rod uct of a "bottom-up" form of a n a l y s i s w i t h the p r o p o s i t i o n s be ing r e l a t e d i n p a i r s u n t i l the N a t u r a l L a n * u a s e 6 : T a y l o r 1 1 7 o v e r a l l network emerges as a f i n a l p roduct . As an a l t e r n a t i v e , we propose t h a t the h i g h e r l e v e l s e mant ic a n a l y s i s should be "top-down" in the sense of making p r e d i c t i o n s about the o v e r a l l themat ic s t r u c t u r e o f the m a t e r i a l be ing processed. Moreo v e r , t h i s a n a l y s i s should make use of a h i g h e r l e v e l or semantic grammar much l i k e the themat ic grammar t h a t Rumelhart (1975) has proposed f o r the themat ic a n a l y s i s o f c h i l d r e n ' s s t o r i e s . 2 . Secondly , we are assuming t h a t a b s t r a c t i n g is n o r m a l l y a dynamic process t h a t should be p r i m a r i l y l o g i c a l o r q u a l i t a t i v e in form and making use of what we might r e f e r to as s u b s t i t u t i o n or condensing o p e r a t o r s . F o r m a l l y , these operators might resemble an axiom or theorem in mathemat ics , s t a t i n g t h a t c e r t a i n content can b e s u b s t i t u t e d f o r c e r t a i n o ther c o n t e n t . These operators w i l l "condense" in the sense of t a k i n g a network r e p r e s e n t i n g a p o r t i o n of t e x t and r e p l a c i n g i t w i t h a s i m p l i f i e d network or perhaps a s i n g l e node. Sometimes, the bas is f o r condensing the t e x t is made e x p l i c i t in the t e x t , in which case the opera t o r is not u n l i k e a procedure f o r i d e n t i f y i n g c e r t a i n types of phrases in c o n t e x t . For example, one o f t e n encounters sentences in the form: "There are two main methods f o r the p roduc t ion of ." or "When c o n s t r u c t i n g a , one immediate ly faces t h r e e types of problems." Thus, in the a b s t r a c t , one wants to i d e n t i f y in one case the two main methods and in the o ther case the t h r e e types of problems. Moreover , one probably wants to name the methods or problems w h i l e i g n o r ing a l l o f the a m p l i f y i n g d e t a i l s . U n f o r t u n a t e l y , under many c i rcumstances , such s t rong c lues may not be g iven e x p l i c i t l y and must be i n f e r r e d in much the same way t h a t answers not e x p l i c i t l y cont a i n e d in a d a t a base can be i n f e r r e d by problem s o l v i n g or theorem prov ing methods. Thus, in our r e v i s e d system we want to i n c l u d e a c a p a b i l i t y f o r "p rov ing" t h a t a set of summarizing statements can be i n f e r r e d from the o r i g i n a l da ta base . In s h o r t , we are proposing two major m o d i f i c a t i o n s to our present a b s t r a c t i n g program in order to make the system per form in a moce "humanl i k e " f a s h i o n and in order to develop a system w i t h a s i g n i f i c a n t l y improved l e v e l of competence.