Advanced topics
The following topics describe advanced features and techniques, including:
- Importing grammars with different tag formats: The grammar compiler accepts mixed-format imports and handles them in a consistent way. Since well-designed grammars are modular and encourage re-use, and because some tag formats are highly portable (SISR syntax)—and others are directed to Nuance Recognizer (swi syntax)—you may encounter situations in which the syntax of semantic tags in a child grammar varies from the syntax of the parent. This topic describes the permutations of grammar imports in detail.
- Multiple parses: Ideally, each possible sentence in the grammar has a unique parse. Occasionally, a grammar may allow multiple parses of a single sentence.
- SWI_listClass: In grammars that have one or more rules composed of large lists of phrases referenced in multiple places, you can define the rule as a SWI_listClass to reduce memory usage and improve recognition accuracy. This topic provides an example and some guidelines for using the rule.
- Applying bigram language models: A bigram model is useful for tuning speech recognition accuracy when you have reason to believe that some words or word sequences will occur more often than others. Applying a bigram boosts scores on more likely phrases compared to less likely ones. See this topic for guidelines on applying a bigram language model to boost accuracy for difficult recognition tasks.
- Getting raw recognition results: The raw recognition format, which is returned in wordlattice format, can be used for semantic analysis and other purposes. This topic describes the XML-based format and how to obtain the results.
- Adding natural language capabilities: Natural language technologies make it possible for a grammar to recognize and extract meaning from a very wide range of different user utterances, without explicitly defining every word that the user may say. Instead, natural language grammars use statistical models that avoid complex grammar rules, derived from actual results. This topic provides an introduction to NLU technologies and techniques.
