Development of a weighted average language model by approximating held‐out data*

Address correspondence to: Peter O, Boyle, The Queen's University of Belfast, Department of Computer Science, Q.U.B., Belfast BT7 1NN, N. Ireland. E‐mail: [email protected], [email protected].

Abstract

We present new variants of the class of statistical models of language which exploit function‐based smoothing and examine their performance by comparing them with some held‐out language models. We can construct a weighted average line‐segment function for combining probability sources the details of which are empirically derived and whose performance matches that of an equivalent held‐out language model.

Notes

Address correspondence to: Peter O, Boyle, The Queen's University of Belfast, Department of Computer Science, Q.U.B., Belfast BT7 1NN, N. Ireland. E‐mail: [email protected], [email protected].