Book Reviews Coherence in Natural Language: Data Stuctures and Applications

In his blurb on the back cover, Mark Liberman calls this book " the biggest step forward [in research on discourse structure] since Aristotle. " Given this eminent recommendation , I read the book with great interest and some anticipation. Wolf and Gibson's book contributes to the growing body of work on computational models of text structure. Following a tradition originating with Mann and Thompson (1988) and Hobbs (1985), Wolf and Gibson assume a " relational " account of coherence: the coherence of a text is attributed to the presence of relations holding at various levels of hierarchy between its constituent elements. Traditionally, relations hold between pairs of propositions; canonical examples are CAUSE, TEMPORAL SEQUENCE, and ELABORATION. Work on relational models of text has pursued two related questions. One question concerns how to characterize the set of special relations whose presence can confer coherence on a text. How many such relations are there, and how should each relation be defined? A second question concerns where in a text we expect these special relations to be found. Answering this question involves using relations as building blocks in a theory of text structure. Wolf and Gibson's book addresses this latter question: their main aim is to propose and argue for a particular theory about the structure of relations in coherent text. Broadly speaking, there are two alternative theories about the arrangement of relations in coherent text. Both are expressed using the terminology of graph theory: the sequence of atomic discourse segments in a text is interpreted as a set of nodes in a graph, ordered by adjacency, and relations introduce arcs (and sometimes additional nodes) into this graph. One theory proposes that the relations in a coherent text must form a tree structure. In this model, each relation holds between two adjacent nodes, and creates a new node representing a unit of text spanning both nodes which can itself be linked to other units by other relations; see Figure 1 for an example. The other theory is less restrictive, requiring only that the relations in a coherent text form a connected graph. In this model, arcs directly denote relations; see Figure 2 for an example. Note that in the graph model, relations can hold between discontiguous text segments and participate in crossing dependencies between segments, and that a single text segment can be involved in several relations. Wolf and Gibson's central proposal …