Word-dependent transition models in HMM based word alignment for statistical machine translation
First Claim
1. A method for estimating an alignment between words in a source phrase and words in a target phrase, comprising:
- providing at least one set of probabilistic word-dependent transition models, said word-dependent transition models having been automatically learned from at least one training data set comprising known parallel texts representing a source language and a target language;
each word-dependent transition model representing a self-jump probability of a particular word in combination with probabilities of jumping from that word to other particular words;
providing a source phrase in the source language and selecting a corresponding word-dependent transition model for each word in the source phrase;
constructing a Hidden Markov Model (HMM) on the source phrase from the probabilistic word-dependent transition models for each word of the source phrase in combination with other HMM components including word emission models;
evaluating the HMM to determine an alignment between the source phrase and a target phrase in the target language; and
storing the alignment between the source phrase and the target phrase for later use.
2 Assignments
0 Petitions
Accused Products
Abstract
A word alignment modeler uses probabilistic learning techniques to train “word-dependent transition models” for use in constructing phrase level Hidden Markov Model (HMM) based word alignment models. As defined herein, “word-dependent transition models” provide a probabilistic model wherein for each source word in training data, a self-transition probability is modeled in combination with a probability of jumping from that particular word to a different word, thereby providing a full transition model for each word in a source phrase. HMM based word alignment models are then used for various word alignment and machine translation tasks. In additional embodiments sparse data problems (i.e., rarely used words) are addressed by using probabilistic learning techniques to estimate word-dependent transition model parameters by maximum a posteriori (MAP) training.
34 Citations
20 Claims
-
1. A method for estimating an alignment between words in a source phrase and words in a target phrase, comprising:
-
providing at least one set of probabilistic word-dependent transition models, said word-dependent transition models having been automatically learned from at least one training data set comprising known parallel texts representing a source language and a target language; each word-dependent transition model representing a self-jump probability of a particular word in combination with probabilities of jumping from that word to other particular words; providing a source phrase in the source language and selecting a corresponding word-dependent transition model for each word in the source phrase; constructing a Hidden Markov Model (HMM) on the source phrase from the probabilistic word-dependent transition models for each word of the source phrase in combination with other HMM components including word emission models; evaluating the HMM to determine an alignment between the source phrase and a target phrase in the target language; and storing the alignment between the source phrase and the target phrase for later use. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8)
-
-
9. A computer-readable medium having computer executable instructions stored thereon for determining a probabilistic mapping between a source phrase and a target phrase, comprising instructions for:
-
providing an automatically learned word-dependent transition model for each source language word in a training set comprising known parallel texts in a source language and a target language; each word-dependent transition model representing a self-jump probability of a particular source language word in combination with probabilities of jumping from that word to other particular source language words; receiving a source phrase in the source language; selecting a corresponding one of the word-dependent transition models for each word in the source phrase; constructing a source phrase based Hidden Markov Model (HMM) including the selected word-dependent transition models; and determining a probabilistic mapping between the source phrase and a target phrase in the target language by evaluating the HMM. - View Dependent Claims (10, 11, 12, 13, 14, 15, 20)
-
-
16. A process for mapping a source phrase in a source language to a target phrase in a target language, comprising steps for:
-
receiving one or more sets of automatically learned word-dependent transition models, each set of word-dependent transition models corresponding to a pair of parallel texts in a unique pair of source and target languages; each word-dependent transition model in each set corresponding a specific one of the words in one of the parallel texts of one of the source languages; each word-dependent transition model representing a self-jump probability of a particular source language word in combination with probabilities of jumping from that word to other particular source language words; selecting one of the pairs of unique source and target languages; receiving a source phrase in the selected source language; selecting a corresponding one of the word-dependent transition models for each word in the source phrase; constructing a source phrase based Hidden Markov Model (HMM) including the selected word-dependent transition models and other HMM components including word emission models; and determining a probabilistic mapping between the source phrase and a target phrase in the corresponding target language by evaluating the HMM. - View Dependent Claims (17, 18, 19)
-
Specification