Augmented-word language model
AssigneeMicrosoft Corporation
InventorsEric K. Ringger, Lucian Galescu
Life of this patent
Estimating the term…
Abstract
A language model comprising a plurality of augmented-word n-grams and probabilities corresponding to such n-grams. Each n-gram is comprised of a sequence of augmented words. Each augmented word is comprised of the orthographic representation of the word together with a tag representing lexical information regarding the word, such as syntactic or semantic information. Also disclosed are a method of building such a language model, a method of automatically recognizing speech using the language model and a speech recognition system that employs the language model.
Claims15
A computer-readable medium having stored thereon a data structure comprising:
- a first data field containing data representing a first word and comprising an orthography subfield containing data representing an orthographic representation of the first word and a tag subfield containing data representing a tag, the tag encoding lexical information regarding the first word;
- and
- a probability data field containing data representing the probability of the first word occurring in a word sequence.
A method of building a language model comprising:
- receiving a training corpus comprising a body of text;
- augmenting words in the training corpus with a tag encoding lexical information regarding the corresponding word;
- selecting a plurality of sequences of n augmented words, n being a positive integer, each selected sequence including a sub-sequence made up of a first n−1 augmented words of the selected sequence;
- and
- for each selected sequence of n augmented words, calculating a probability that, given an occurrences of the sub-sequence in a block of text, an immediately following word will be the nth augmented word of the selected sequence.
A language model, comprising:
- a plurality of word fields, each word field including an orthographic subfield that includes an orthographic representation of an associated word, a lexical subfield that includes lexical information corresponding to the associated word, and a probability subfield including probability information indicative of a probability that the associated word occurs with the lexical information.
Dependencies are inferred from the claim text.