Structure-Driven Lexicalist Generation

Proceedings of COLING 2012: Technical Papers, pages 2027–2042, COLING 2012, Mumbai, December 2012. Structure-Driven Lexicalist Generation Shashi Narayan1 Claire Gardent2 (1) Université de Lorraine/LORIA, Nancy, France (2) CNRS/LORIA, Nancy, France shashi.narayan@loria.fr, claire.gardent@loria.fr Abstract We present a novel algorithm for surface realisation with lexicalist grammars. In this algorithm, the structure of the input is used both top-down to constrain the selection of applicable rules and bottom-up to filter the initial search space associated with local input trees. In addition, parallelism is used to recursively pursue the realisation of each daughter node in the input tree. We evaluate the algorithm on the input data provided by the Generation Challenge Surface Realisation Task and show that it drastically reduce processing time when compared with a simpler, top-down driven, lexicalist approach. Title and Abstract in Hindi ������ - p���� ш�� - ����� u����� �� ш�� - ����� ���к�� к� ��� �� ��k�� к� ��� ������ к� ���� eк ��� �к��к p��� � к� ��� ���। i� �к��к ��� a��� ������ к� u���� ����� “ ��� �� ���� ” u��� k ������ к� ��� �� ��� “ ���� �� ��� ” a��� ������ ��� ���� � ������ �� ��� �� �� �� p�����к ����� ��� к� ����� к��� ��� �к�� ��� ��। i�к� ��� , ��������� к� u���� a��� �� ��� к� ш�х�� к� ��� ������ ���� к� ш� ���к� �� �к�� ��� ��। �� a��� �к��к к� �� ����к� “ u����� p�������� : ��� ������ ” к� a��к��� �� к��� ��� ��� �� �� ��х���� ��� �к p��� � ���к� �� ��� ������ “ ��� �� ���� ” p���� ш�� - ����� ���к�� к� �� к���� u����� ��� к� a����к ��� ���� ��। Keywords: Generation, Tree Adjoining Grammar, Surface Realization. Keywords in Hindi: u����� , �� � - �� - ���к�� , ��� ������. 2027 1 Introduction Depending on the type of semantic representation encoded by the grammar, two main types of al- gorithms have been proposed for generating sentences with bi-directional, unification-based gram- mars such as CCG (Combinatory Categorial Grammar, (Espinosa et al., 2010)), HPSG (Head- Driven Phrase Structure Grammar, (Carroll et al., 1999)) and TAG (Tree Adjoining Grammar, (Gardent and Kow, 2005)). For recursive semantic representations such as first-order logic formulae, head-driven algorithms (Shieber et al., 1990) have been argued to be best because they restrict the combinatorics inherent to bottom-up search; they avoid non termination by using lexical items to guide the search ; and they allow for semantically nonmonotonic grammars (i.e., grammars where the semantics of a rule at the left hand side need not be subsumed by the semantics of the rule at the right hand side). One main issue with this approach however is the so-called logical form equivalence problem (Shieber, 1993). A logic formula may have several logically equivalent but syntactically distinct formulae. For instance p∧q is logically equivalent to q ∧ p. In general though, a grammar will associate with natural language expressions only one of these logically equivalent formula. Hence a generator will be able …

Content truncated to 500 words