Optimizing parameters for machine translation

A computer-implemented method comprising:

• accessing, at a computing device including a processor, a translation lattice that represents a plurality of candidate translations, the translation lattice including a plurality of paths including nodes connected by edges;
• calculating, at the computing device, posterior probabilities for each n-gram represented in the translation lattice including:
• determining, for each node, a score for each n-gram that is represented by a path from a source node of the translation lattice to the node, where the score is a highest posterior probability for edges on paths that include the n-gram;
• and
• performing, at the computing device, decoding on the translation lattice using the scores to convert a sample text from a first language to a second language,
• where calculating the posterior probabilities includes calculating:
• P
• ⁡
• (
• w
• |
• Ψ
• )
• =
• ∑
• E
• ∈
• Ψ
• ⁢
• ∑
• e
• ∈
• E
• ⁢
• f
• ⁡
• (
• e
• ,
• w
• ,
• E
• )
• ⁢
• P
• ⁡
• (
• E
• |
• F
• )
• ,
• where P(w|Ψ) is the posterior probability of the n-gram w in the translation lattice;
• Ψ is a space defined by the translation lattice;
• E is a candidate translation;
• F is the sample text in the first language, e is an edge;
• and f(e,w,E)=1 when wεe, P(e|Ψ)>P(e′|Ψ), and e′ is an edge that precedes e on E;
• otherwise, f(e,w,E)=0.

A computer program product, encoded on a non-transitory program carrier, operable to cause a processor of a computing device to perform operations comprising:

• accessing a translation hypergraph that represents a plurality of candidate translations, the translation hypergraph including a plurality of paths including nodes connected by edges;
• calculating first posterior probabilities for each edge in the translation hypergraph;
• calculating second posterior probabilities for each n-gram represented in the translation hypergraph based on the first posterior probabilities;
• and
• performing decoding on the translation hypergraph using the second posterior probabilities to convert a sample text from a first language to a second language,
• where calculating the second posterior probabilities includes calculating:
• P
• ⁡
• (
• w
• |
• Ψ
• )
• =
• ∑
• E
• ∈
• Ψ
• ⁢
• ∑
• e
• ∈
• E
• ⁢
• f
• ⁡
• (
• e
• ,
• w
• ,
• E
• )
• ⁢
• P
• ⁡
• (
• E
• |
• F
• )
• ,
• where P(w|Ψ) is the posterior probability of the n-gram w in the translation hypergraph;
• Ψ is a space defined by the translation hypergraph;
• E is a candidate translation;
• F is the sample text in the first language, e is an edge;
• and f(e,w,E)=1 when Wεe, P(e|Ψ)>P(e′|Ψ), and e′ is an edge that precedes e on E;
• otherwise, f(e,w,E)=0.

A computer program product, encoded on a non-transitory program carrier, operable to cause a data processing apparatus to perform operations comprising:

• accessing a translation lattice that represents a plurality of candidate translations, the translation lattice including a plurality of paths including nodes connected by edges;
• calculating posterior probabilities for each n-gram represented in the translation lattice including:
• determining, for each node, a score for each n-gram that is represented by a path from a source node of the translation lattice to the node, where the score is a highest posterior probability for edges on paths that include the n-gram;
• and
• performing decoding on the translation lattice using the scores to convert a sample text from a first language to a second language,
• where calculating the posterior probabilities includes calculating:
• P
• ⁡
• (
• w
• |
• Ψ
• )
• =
• ∑
• E
• ∈
• Ψ
• ⁢
• ∑
• e
• ∈
• E
• ⁢
• f
• ⁡
• (
• e
• ,
• w
• ,
• E
• )
• ⁢
• P
• ⁡
• (
• E
• |
• F
• )
• ,
• where P(w|Ψ) is the posterior probability of the n-gram w in the translation lattice;
• Ψ is a space defined by the translation lattice;
• E is a candidate translation;
• F is the sample text in the first language, e is an edge;
• and f(e,w,E)=1 when wεe, P(e|Ψ)>P(e′|Ψ), and e′ is an edge that precedes e on E;
• otherwise, f(e,w,E)=0.