📄 Book Review: Design Patterns in Fluid Construction Grammar edited by Luc Steels

Summary

This is a review by Nathan Schneider and Reut Tsarfaty of a volume edited by Luc Steels on Fluid Construction Grammar (FCG). FCG is a formalism that, following the principles of Construction Grammar (CxG), combines parsing, production, and learning within a single computational system. The reviewers place the book within the "integrative" line of research. It does not aim at broad data coverage but instead shows how different mechanisms work together on a small number of examples. The volume has five parts: the foundations of the formalism, phrasal and argument constructions, search management during processing, complex phenomena in German and Hungarian, and tools for robustness to noise and variation. Each chapter identifies reusable grammar "design patterns." The reviewers see the book's main value in its demonstration that the constructional approach can be implemented as a full-fledged computational system. They regard CxG, with its different conception of modularity, as a serious challenge for NLP. Among the shortcomings, they name the terse and overly technical presentation and the weak connection to other lines of research. The book also lacks learning of constructions directly from data and evaluation on sentences the system has not seen before.

Key points

• Publication: John Benjamins, Constructional Approaches to Language series, vol. 11, 2012, xi+332 pp. FCG is a unification-based formalism, like Embodied Construction Grammar and Sign-Based Construction Grammar.
• Processing in FCG consists of incrementally applying constructions to a transient structure. Templates eliminate boilerplate code, and the chapter on phrasal constructions alone introduces more than twenty of them.
• By default, search is greedy and depth-first, with backtracking. Construction scores and networks of conditional dependencies between constructions sharply reduce processing time. However, in the experiments the sentences were assumed to have already been seen by the system, i.e., there was no held-out sample.
• Three case studies:
  • feature matrices for case indeterminacy in German (in coordination, Frauen is acceptable but Kindern is not);
  • ordered construction sets for Hungarian poly- and monopersonal agreement, where zero affixes act as default values;
  • German spatial terms for interaction with robots.
• Robustness is ensured by a diagnosis and repair mechanism that runs in parallel with parsing and production, including through reverse transduction. The goal of processing is communicative success rather than grammaticality.