| 发明名称 |
|
| 摘要 |
PURPOSE: To estimate a language model having high and accurate prediction probability by using a small number of parameters. CONSTITUTION: A word string hypothesis record is generated 10 by combining individual records for respective words of a hypothesis. The hypothesis record of one unit and the word string hypothesis record of one unit are generated 20, 22 from the combination of the conditioned generation probability of estimation for 1st class words consisting of recorded words when a context consisting of unrecorded words for the word string hypothesis is generated and conditioned generation probability based on the estimation of recorded word when the 1st class words and a context are generated. Then the best candidate for one or more word string hypotheses having the best word string hypothesis record are recorded 24 and plural word strings are classified to obtain an effective hypothesis record. |
| 申请公布号 |
JPH05289692(A) |
申请公布日期 |
1993.11.05 |
| 申请号 |
JP19920350244 |
申请日期 |
1992.12.04 |
| 申请人 |
INTERNATL BUSINESS MACH CORP <IBM> |
发明人 |
PIITAA FUITSUJIYUU BURAUN;SUTEFUEN ANDORIYUU DERA PIITORA;BINSENTO JIYOSEFU DERA PIITORA;ROBAATO REROI MAASAA;FUIRITSUPU ESU REZUNIKU;SUTANREI ESU CHIEN |
| 分类号 |
G10L15/10;G06F9/44;G06F17/27;G10L15/18;G10L15/28 |
主分类号 |
G10L15/10 |
| 代理机构 |
|
代理人 |
|
| 主权项 |
|
| 地址 |
|
