Summarizing Narratives

Most research on na r ra t i ve tex t summarisation has been conducted w i t h i n the paradigm of experimental psychology. But recent language processing research i n a r t i f i c i a l i n t e l l i g e n c e suggests that the predominant theory of tex t summarisation requi res fu r the r examinat ion. Seemingly minor s t r u c t u r a l mod i f i ca t ions of a s tory can reau l t in a i g n i f i c a n t a l t e r a t i o n s of summary behavior . In t h i s paper, h i g h l i g h t s of summary data from 72 subjects are presented and analyzed in terms of two competing summarization models: (1) the s to ry grammar model of psychology, and (2) the p lo t u n i t model developed in a r t i f i c i a l i n t e l l i g e n c e . We w i l l show how selected story grammar p red ic t ions compare to p lo t un i t p red ic t ions f o r short term summarization and then i d e n t i f y two compl icat ing fac to rs that have a major impact on summarisation behavior . 1 . I n t roduc t i on While experiments using people may seem less re levant t o a r t i f i c i a l i n t e l l i g e n c e than experiments using computers, the work described in t h i s paper represents more than an i n t e l l e c t u a l commitment to psychologica l v a l i d i t y in AI (a l though tha t commitment is a lso p resen t ) . The task of t ex t summarisation is one of the most complicated chal langes fac ing na tu ra l language processing and poasib ly fac ing AI a l t oge the r . We need a l l the help we can g e t . We have been working in recent years at Yale on a s to ry understanding system (BORIS) (2] whose t h e o r e t i c a l roo ts go back to the MARGIE system [ 1 0 ] . While an i n i t i a l implementation toward summarisation has been completed in BORIS [7] , we have not yet been able to tes t our most important i n t u i t i o n s using BORIS. We probably won't be able to u t i l i s e BORIS f o r t h i s purpose u n t i l i t can process a large number of proto type s t o r i e s . In the meantime, we have been conducting exp lora tory psychologica l experiments, to see how our ideas hold up against systems that can summarize a large number of s t o r i e s . The r e s u l t s have been both su rp r i s ing and encouraging. This work was sponsored in part by the Advanced Research Pro jects Agency under cont ract N0014-75-C-1111 and in par t by the Nat iona l Science Foundation under cont ract IST7918463. 2. Plot U n i t s : A Heterarch ica l Representat ion Plot un i t s ( o r i g i n a l l y ca l l ed " a f f e c t u n i t s " ) 13,4] lead to general graphs that are t y p i c a l l y c y c l i c i n t h e i r a t ruc tu re ra ther than h i e r a r c h i c a l . A p lo t un i t ia a f i xed con f i gu ra t i on of mental s tates and events i nvo l v i ng one or more characters in a atandard behaviora l pa t te rn or i n t e r a c t i o n . For example, there are p lo t un i t s that represent honored requests , problem r e s o l u t i o n , r e t a l i a t i o n , broken promises, and o b l i g a t i o n s . A l l p l o t un i t s are constructed from smaller p r i m i t i v e components ca l l ed " a f f e c t s t a t e s . " There are three kinds of a f f e c t s t a t e s , designed to encode mentsl s t a tes , p o s i t i v e events, and negat ive events. The i n t e r n a l s t ruc tu re of a p lo t u n i t is governed by s t r i c t ru les to con t ro l the var ious ways a f f ec t s tates can be l inked to each other [ 3 ] , but the most important proper ty of a p lo t un i t representa t ion l i e s in the fac t that p lo t un i t s can over lap w i t h each o the r . Two p lo t un i t s over lap when they both conta in a common a f f e c t s t a t e . For example, suppose John asks Mary to buy an a r t i choke , and she does so. The un i t descr ib ing John's honored request w i l l over lap w i t h the un i t descr ib ing Mary's success because they both share the same p o s i t i v e event of Mary buying an a r t i choke . This event is a c r i t i c a l component fo r both Mary's success and John's request :