Declarative Programming for Natural Language Generation

Home

The task brings two kinds of obstacles.

• Specifications. NLG systems require large and complex resources that describe how to use language. The research challenge is to construct such resources quickly, correctly and generally.
• Algorithms. Once you develop suitable communicative resources, the research challenge is to use them soundly and efficiently. Formulating the needed procedures can be difficult and involved.

Because SPUD constructs sentences by reasoning from a declarative grammar, SPUD brings distinctive opportunities for exploring specifications in NLG. I explore these questions, which occupy me presently, in a later subsection.

Algorithms for NLG
NLG is typically broken down into stages of discourse planning (to select information and organize it into coherent paragraphs), sentence planning (to choose words and structures to fit information into sentence-sized units), and realization (to determine surface form of output, including word order, morphology and final formatting or intonation). The SPUD system combines the generation steps of sentence planning and surface realization by using a lexicalized grammar to construct the syntax and semantics of a sentence simultaneously.

The basic input to SPUD consists of a set of communicative goals to achieve. SPUD seeks to satisfy three kinds of constraint in mapping these goals onto a sentence that accomplishes them. Any fact to be communicated must be contributed directly or by inference from an element in an abstract grammatical structure, including lexical items. Any reference to a domain entity must be elaborated into a description that distinguishes the entities from its distractors (the salient alternatives to it in context). And all this conceptual material must be organized into a grammatical surface form.

SPUD builds its sentence step-by-step. Each step adds to the ongoing sentence a lexicalized entry encoded as an elementary tree in Lexicalized Tree-Adjoining Grammar (LTAG). By using these LTAG structures, SPUD always works with concrete grammatical surface forms. Each tree is paired with logical formulae that, by referring to a rich discourse model, characterize the semantic and pragmatic contribution that the tree makes to the sentence. By keeping track of these contributions as it builds the sentence, SPUD is able to assess the shared information that the sentence uses to identify discourse entities and the new information that the sentence adds about them.

The advantages of this scheme include these:

• Syntactic constraints are handled early and naturally, ensuring the expressibility of selected concepts.
• The order of adding content is flexible, allowing varied sentences without complicated search.
• Grammatical knowledge is stated once only, inside the generator.
• It enables efficient descriptions that refer implicitly to parameters from the context.
• It enables efficient descriptions that exploit the hearer's ability to recognize inferential links to material elsewhere in the sentence.

Specifications for NLG
In NLG, the ideal way to specify linguistic resources is by example. In most NLG applications, you can readily find or construct gold-standard texts that you would like your system to generate. These texts themselves should determine NLG resources with minimal further effort. The research problem is to articulate methodologies for preparing texts, and procedures for automatic text analysis, that suffice to bootstrap a generator from data.

Now, clearly we need to annotate the syntactic structure behind the text - a tree or some similar formal description. We also need to know something about the speaker's intentions in producing the text - for example, the domain elements that the speaker intended to refer to. But what else?

We explored this problem manually, by building a SPUD grammar for a small set of sentences. We were surprised to discover a methodology that allows specifications to be developed systematically, by applying simple heuristics connected with clear corpus evidence. We have since been working to automate this methodology. For example, a slightly simpler syntactic formalism allows us to go directly from a treebank to syntactic grammar entries. Meanwhile we hypothesize that we will then be able to go automatically from an annotated utterance to a schema for the context the utterance presupposes and the communicative goals that the utterance addresses. These two steps amount to an algorithm that compiles a potential output sentence into the abstract input and the grammatical resources that would be required for SPUD to create the sentence.