Experiments in Linear Template Combination using Genetic Algorithms

Natural Language Generation systems typically have two parts ­ strategic (" what to say ") and tactical (" how to say "). We present our experiments in building an unsupervised corpus­driven template based tactical NLG system. We consider templates as a sequence of words containing gaps. Our idea is based on the observation that templates are grammatical locally (within their textual span). We posit the construction of a sentence as a highly restricted sequence of such templates. This work is an attempt to explore the resulting search space using Genetic Algorithms to arrive at acceptable solutions. We present a baseline implementation of this approach which outputs gapped text. NLG is the task of generating natural language from non­linguistic inputs. Most NLG systems can be classified into two broad camps ­ template based and statistical. Template based systems are generally characterized by structure gapped text (slot­filler structure) which is predominantly manually created (Reiter 1995) which generally result in high quality text but also limited linguistic coverage. Statistical systems such as (Langkilde 1998) on the other hand, use data­driven algorithms for text generation and have little­to­no reliance on repetitive manual resources making them more adaptable and maintainable, albeit with lesser text quality (Reiter 1995). In this work, we consider generation on a sentence level and output gapped text. We define templates as gapped text which can be filled to generate textual output. A (partial) sentence is a linear sequence of such templates. Since there are a large number of choices (templates) at every step of sentence generation (sequence of templates), it naturally gives rise to a search space which contains all such sequences ­ grammatical and ungrammatical. The challenge is to navigate this large search space and arrive at reasonable grammatical sentences. RELEVANT WORK Template based systems have been considered to have a very shallow linguistic representation and mapping of the non­linguistic data to the generated text. General consensus among the NLG community has been that template based systems are not as flexible, maintainable and expressive (linguistic coverage) as full fledged NLG systems (Reiter 1995). (Reiter 1997) mentions that template based systems and NLG systems are " turing 1