A Summariser based on Human Memory Limitations and Lexical Competition

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 732–741, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Summariser based on Human Memory Limitations and Lexical Competition Yimai Fang Computer Laboratory University of Cambridge 15 JJ Thomson Avenue, CB3 0FD, UK Yimai.Fang@cl.cam.ac.uk Simone Teufel Computer Laboratory University of Cambridge 15 JJ Thomson Avenue, CB3 0FD, UK Simone.Teufel@cl.cam.ac.uk Abstract Kintsch and van Dijk proposed a model of human comprehension and summarisa- tion which is based on the idea of pro- cessing propositions on a sentence-by- sentence basis, detecting argument over- lap, and creating a summary on the basis of the best connected propositions. We present an implementation of that model, which gets around the problem of identi- fying concepts in text by applying coref- erence resolution, named entity detection, and semantic similarity detection, imple- mented as a two-step competition. We evaluate the resulting summariser against two commonly used extractive summaris- ers using ROUGE, with encouraging re- sults. 1 Introduction Kintsch and van Dijk (1978) (henceforth KvD) present a model of human comprehension and memory retention which is based on research in ar- tificial intelligence, experimental psychology and discourse linguistics. It models the processing of incoming text or speech by human memory lim- itations, and makes verifiable predictions about which propositions in a text will be recalled by subjects later. It has been very influential, particu- larly in the 1980 and 1990s in educational (Palin- scar and Brown, 1984; King, 1992) and cognitive (Paivio, 1990) psychology, and is still today used as a theoretical model of reading and comprehen- sion (Baddeley, 2007; Zwaan, 2003; DeLong et al., 2005; Smith, 2004). It has also been used for improving education, particularly for the produc- tion of better instructional text (Britton and Gul- goz, 1991; Pressley, 2006), and for teaching hu- mans how to read for deep comprehension (Coiro and Dobler, 2007; Duke and Pearson, 2002; Koda, 2005; Driscoll, 2005) and to summarise (Hidi, 1986; Brown et al., 1983). In the summarisation community, the model has been commended for its elegant and explana- tory “deep” treatment of the summarisation pro- cess (Lehnert, 1981; Sp¨arck Jones, 1993; Endres- Niggemeyer, 1998), but has not lead to any prac- tical prototypes, mainly due the impossibility of implementing the knowledge- and inference-based aspects the model relies on. We present here an implementation of the model, which attempts to circumvent some of these prob- lems by the application of distributional seman- tics, and by modelling the construction of the co- herence tree as a double competition (firstly of concept partners for word forms, secondly of at- tachment sites for propositions). In the KvD model, a text (e.g. Figure 1) is con- verted into propositions (see Table 1) which have one functor and one or more arguments. The func- tor can be taken either from a fixed list of gram- matical relations (e.g. IS A; AT; BETWEEN; OR) or an open class-set of so-called concepts, (e.g. BLOODY; TEACH). Arguments can be …

Content truncated to 500 words