JPH07200604A - Apparatus and method for language translation using contenxt-based translation model - Google Patents

Abstract

PURPOSE: To provide a device for translating a series of source words of a first language to a series of target words of a second word. CONSTITUTION: Concerning a series of inputted source words at least two target hypotheses including respectively series of target words are generated. Each target word is provided with a context including at least one other word of a target hypothesis. Concerning each target hypothesis, a language coincident score includes the estimated value of the generation probability of the series of works. At least one alignment combining each source word with at least one of the target words of the target hypotheses is identified. Concerning each source unit and each target hypothesis, a translation coinciding score includes the estimated value of the conditional generation probability of the source word when the target word of the target hypothesis connected with the source word and the context of the target hypothesis of the target word combined with the source language are given. Concerning each target hypothesis, the translation coinciding score includes the combination of the word coinciding score of the target hypothesis and the source word of a series of inputted source words. A target hypothesis coinciding score includes the combination of the language model coinciding score of the target hypothesis and the translation coinciding score of the target hypothesis. A target hypothesis with a best target hypothesis coinciding score is outputted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to computer language translation, such as computer translation of French to English.

[0002]

[Prior Art] Peter F. Bro
wn) Other "Method and system For Natural Language Tr
In US patent application Ser. No. 07 / 736,278, filed Jul. 25, 1991, entitled "anslation", the source language text F is translated into the target language text E, the entire contents of which are incorporated herein by reference. A computer translation system is described for which, for each of several hypothetical target language texts E, the conditional probability P of the target language text E given a source language text F. Evaluate (E | F).
The hypothetical target language text E hat with the highest conditional probability P (E | F) is selected as the translation of the source language text F.

[Equation 1] Is represented by E hat.

Using the Bayesian theorem, the conditional probability P (E | F) of the target language text E, given the source language text F, can be expressed as:

## EQU00002 ## P (E | F) = P (F | E) P (E) / P (F) (1)

Since the probability P (F) of the source language text F, which is the denominator of Equation 1, is independent of the target language text E, the hypothetical target language text E hat having the highest conditional probability P (E | F). Is the highest product P (F | E) P
(E) will also be included. Therefore, the following equation is obtained.

[Equation 3]

In equation 2, the probability P of the target language text E is
(E) is a language model matching score, which is estimated from the target language model. Although any known linguistic model may be used to estimate the probability P (E) of the target linguistic text E, the above-referenced related US patent application Ser. No. 07/73627.
The issue describes an n-gram language model including a 1-gram model, a 2-gram model, and a 3-gram model combined by parameters whose values are obtained by interpolation estimation.

Conditional probability P (F | E) in equation 2
Is the translation match score. As described in the above related patent application, a source text F containing a series of source words given a target hypothesis E containing a series of target words.
The translation matching score P (F | E) of is estimated by finding all possible alignments that combine the source word in source text F with the target word in target text E, and one of these alignments is Or, an alignment in which a plurality of source words are not connected to any target word is included, but an alignment in which a source word is connected to a plurality of target words is not included. For each alignment and each target word e in the target text E combined with the φ source word in the source text F, the probability of probability n (φ | e) that the target word e is connected to the φ source word in the alignment. Is estimated. Also,
For each source word f in source text F, and for each target word e in target text E linked to source word f by alignment, source word f occurs when the occurrence of the linked target word e is given. , Vocabulary probability t (f |
e) is estimated.

For each alignment and each source word f, the related patent application cited above further assumes that the target word e connected to the source word f is at position a j in the target text E, and the source text F contains Assuming that there are m words, the distortion probability a (j | aj, m) in which the source word f exists at the position j of the source text F is estimated.

The alignment probability and the probability probabilities of all target words e in the target text E are combined, and the result is the number of target words φ 0 that are not combined with any of the source words in the alignment.
Probability of

[Equation 4] By multiplying by, the fertility score of the target text E and the alignment is obtained assuming that the sum φ of the fertility of all target words in the target text E being aligned is given.

Combining the alignment vocabulary probability and the vocabulary probabilities of all source words in the source text F,
The alignment lexical score is obtained.

The alignment distortion probability is combined with the distortion probabilities of all source words in the source text F associated with the target word in the alignment, and the result is 1 / φ0 !. By multiplying (where φ0 is the number of target words in the target text E that are not combined with any source word), the distortion score of the alignment is obtained.

Finally, combining the alignment fertility score, vocabulary score, and distortion score,
The result is a combination coefficient

[Equation 5] The translation match score for the alignment is obtained by multiplying by. (Section 8. of the above related patent application.
2. (See)

The translation match score P (F | E) of the source text F and the target hypothesis E may be the sum of the translation match scores of all the allowed alignments between the source text F and the target hypothesis E. The translation match score P (F / E) of the source text F and the target hypothesis E is preferably the translation match score of the alignment estimated to have the highest probability.

Using Equation 2, the hypothetical target language text E
And the target hypothesis agreement score P (F |
E) P (E) can also be estimated directly. However, in order to simplify the language model P (E) and the translation model P (F | E), and to estimate the parameters of these models from a manageable amount of training data, The target hypothesis matching score P (F | E) P (E) of the simplified intermediate forms E ′ and F ′ of the target language text E and the source language text F, respectively, is estimated. Each intermediate target language word e'represents a class of related target language words.
Each intermediate source language word f'represents a class of related source language words. The source language conversion mechanism converts the source language text F into an intermediate form F '. Best hypothesis match score P (F '|
The target language text E-hat 'in the hypothetical intermediate form with E') P (E ') is estimated from Eq. A target language translation mechanism translates the best matching intermediate target language text E-hat 'into a target language text E-hat.

In the language translation system of the above related patent application, the vocabulary probability of each source word f is a conditional probability of each source word f when only the target word e combined with the source word in the alignment is given. It is estimated as t (f | e). The vocabulary probability thus only provides a rough estimate of the probability of the source word f.

[0015]

It is an object of the present invention to provide a set of source words of a first language with a set of target words of a second language different from the first language, which gives an improved estimation of the source word vocabulary probability. An apparatus and method for translating into.

Another object of the invention is the condition when the vocabulary probability of a source word is given a target word connected to the source word in the alignment and the context of the target word connected to the source word. First estimated as the attached probability
An apparatus and method for translating a series of source words in a language into a series of target words in a second language different from the first language.

[0017]

According to the present invention, an apparatus for translating a series of source words in a first language into a series of target words in a second language different from the first language comprises a series of source words. And means for inputting. Means are also provided for generating at least two target hypotheses. Each target hypothesis comprises a series of target words selected from a second language vocabulary. Each target word has a context that includes at least one other word in the target hypothesis.

The language model match score generation mechanism generates, for each target hypothesis, a language model match score including an estimated value of the occurrence probability of a series of words of the target hypothesis. An alignment identification mechanism identifies at least one alignment between the input series of source words and each target hypothesis. The alignment combines each source word with at least one target word of the target hypothesis.

The word match score generation mechanism is given, for each source word and each target hypothesis, a target word of the target hypothesis coupled to the source word and a context of the target word of the target hypothesis coupled to the source word. It is provided to generate a word match score, which includes an estimate of the conditional occurrence probability of the source word if The translation match score generation mechanism generates, for each target hypothesis, a translation match score including a combination of the word match score of the target hypothesis and a source word in the series of source words input.

The hypothesis matching score generation mechanism is provided to generate a target hypothesis matching score for each target hypothesis. Each target hypothesis match score includes a combination of the target hypothesis language model match score and the target hypothesis translation match score. The goal hypothesis with the best goal hypothesis match score is provided at the output.

Each target hypothesis preferably includes a set of target words selected from a vocabulary containing words in a second language and blank words indicating no words.

The alignment identification mechanism may also include means for identifying two or more alignments between the input series of source words and each target word.
Each alignment combines each source word with at least one target word of the target hypothesis. The word match score generation mechanism estimates a conditional occurrence probability of a source word for each source word, each alignment, and each target hypothesis when a target word connected to the source word is given and a context of the target word is given. Generate a word match score, including values.
The translation match score generation mechanism generates, for each target hypothesis, a translation match score that includes a combination of the target hypothesis word match score and a source word in the input series of source words.

The source text input device includes means for converting a series of input source words into a series of converted source words. The alignment means identifies at least one alignment between the series of transformed source words and each target hypothesis. Each alignment binds each transformed source word to at least one target word of the target hypothesis. The word match score generation mechanism, for each converted source word and each target hypothesis, is given the target word of the target hypothesis connected to the converted word and the context of the target word of the target hypothesis connected to the converted source word. Generate a word match score that includes an estimate of the conditional occurrence probability of the transformed source word in the case.

The translation match score generation mechanism generates, for each target hypothesis, a translation match score including a combination of the word match score of the target hypothesis and the converted source word. Output means are provided for synthesizing a series of output words from the target hypothesis having the best target hypothesis match score and for outputting the output words.

The target hypothesis translation match score may include the product of the target hypothesis word match score and the source word in the input series of source words. The target hypothesis match score for the target hypothesis may include the product of the target model hypothesis language model match score times the translation match score for the target hypothesis.

The context of the target word in the target hypothesis bound to the source word can be included in at least one of two or more context classes. An estimated conditional occurrence probability of a source word given a target word coupled to the source word and a context of the target word has at least a value that depends on the class containing the context of the target word.
Including one function.

Alternatively, the estimated conditional occurrence probability of the source word is a value that depends on the part of speech in the target hypothesis of at least one word in the context of the target word bound to the source word, or of the target word bound to the source word. A function may be included that has a value that depends on the identification of at least one word in the context.

The means for outputting the target hypothesis with the best match score can include a display. The input means is a keyboard, a computer disk drive mechanism,
Or it may include a computer tape drive.

Estimating the vocabulary probability of the source word as the conditional probability of the source word given the target word connected to the source word of the alignment and the context of the target word connected to the source word. Thus, the present invention provides a translation match score with improved accuracy.

[0030]

1 is a block diagram of an example of a device for translating words from a first language to a second language according to the present invention. The device includes a source text input device 10 for entering a series of source words. Primitive text input device 10 may include, for example, a keyboard, computer disk drive, or computer tape drive.

The source text input device 10 may further include means for converting the input series of source words into a series of converted source words. Each converted source word may represent a set of related input source words. For example, each input source word {etre, etant, ete, suis, es, est.
sommes, etes, sont, fus, fumes, serai, serons, soi
s, soit, soyons, soyez, soient} is a variation of the French indefinite verb "etre" and can be converted into a tagged "etre" that represents the tense of the original input source word.

The means for converting an input sequence of source words into a sequence of conversion source words may include the source conversion mechanism described in Sections 3, 4, 11 of the above related patent application. Basically, these conversion mechanisms check and correct the spelling of the input source word, check and correct the case of the input source word, detect the document title in the input source word sequence,
And detecting the name in the input series of source words. The translator also tags each input source word with the most probable part of speech of the source word and flags unknown source words (those not included in the stored source word vocabulary). The means for converting a series of source words entered is also
A plurality of word units of the input source word are decomposed into a single converted source word, and the composite input source word is divided into two or more converted source words. The means for converting an input sequence of source words into a sequence of conversion source words further performs a linguistic or morphological conversion of a word into various basic single forms. Finally, the means for converting the input sequence of source words also estimates the meaning of each input source word and assigns this meaning to the converted source word.

Table 1 is an example hypothesis of a series of input source words according to the present invention. In this example, the source word is a French word.

[Table 1] Series of input source words, F f1 f2 f3 f4 f5 f6 La clef est dans la porte

The translation device according to the present invention further includes a target hypothesis generating mechanism 12. The target hypothesis generating mechanism 12 generates at least two target hypotheses. Each target hypothesis includes a set of target words selected from a second language word vocabulary. The second language word vocabulary may be stored in the target language vocabulary storage area 14. Each target word of the target hypothesis has a context that includes at least one other word of the target hypothesis.

An example of a target hypothesis generation mechanism is described in section 14 of the above patent application.

Table 2 shows an example of one of the target hypotheses E1, E2 and E3. In this example, the target word is a word.

[Table 2]

Each target hypothesis includes a set of target words selected from a vocabulary containing words in a second language and a blank word indicating that the word does not exist. In Table 2, all of the target hypotheses Eh are considered to contain "blank" words.

Referring again to FIG. 1, the translation apparatus generates, for each target hypothesis, a language model match score generation mechanism 16 for generating a language model match score including an estimated value of the occurrence probability of a series of words of the target hypothesis. including. An example of a language model match score generation mechanism is described in Sections 6 and 7 of the above related patent applications. Although any known model can be used to estimate the probability of occurrence of a series of words in the target hypothesis, in the above-mentioned related patent application, a 1-gram model, 2-gram model combined by parameters whose values are obtained by interpolation estimation. -Models and n-gram language models including 3-gram models are described.

The translation device further includes an alignment identification mechanism 18 for identifying at least one alignment between the input series of source words and each target hypothesis. The alignment combines each source word with at least one target word of the target hypothesis.

2, 3 and 4 schematically show examples of possible alignments between the set of hypothesis input source words in Table 1 and the set of hypothesis input target words of the target hypothesis E1 in Table 2. In each alignment, each source word in the input series of source words F is connected to at least one target word of the target hypothesis E1 by a solid line. In the alignment of FIG. 4, there is no solid line from the second occurrence of the source word "La" and is therefore considered to be joined to the "blank" word.

Tables 3, 4, and 5 show different alignments of FIGS. 2, 3, and 4, respectively.

[Table 3]

[Table 4]

[Table 5]

In each table, for each parameter j in the range 1 to m (where m is the number of words in the input source word), another parameter aj having a single value in the range 0 to 1 (Where l is the number of words in the target hypothesis).

For a given alignment, each word fj in the series of source words input is the target hypothesis word ea.
is bound to j.

In general, a series of m source words and a series of l
There are 2 lm possible alignments with the non-blank target words, where each source word is combined with either a blank source word or one or more non-blank target words. M if each source word is constrained to be joined to exactly one blank or non-blank target word
There are (l + 1) possible alignments.

Preferably, only one alignment between the input sequence of source words and each target hypothesis is identified to obtain a word match score for each source word, as described below. The series of source words and one identified alignment for each target hypothesis is preferably generated by the target hypothesis generator, as described in section 14 of the above-referenced patent application.

If the source text input device 10 includes means for converting an input sequence of source words into a sequence of converted source words, the alignment identification mechanism 18 will:
At least one between the series of transformed source words and each target hypothesis
Identify the two alignments. The alignment combines each transformed source word with at least one target word of the target hypothesis.

Referring again to FIG. 1, the translation device further includes a word match score generation mechanism 20. For each source word and each target hypothesis, the word match score generation mechanism 20 is provided with the target word e of the target hypothesis coupled to the source word f, and the context X of the target word e of the target hypothesis coupled to the source word f.
Conditional occurrence probability P of the source word f when is given
Generate a word match score that includes an estimate of (f | e, X).

Table 6 shows a hypothetical example of the context X of each target word eaj of the target hypothesis E1 of Table 2 for the alignment A1,1 of Table 3 with the input series of source words F of Table 1.

[Table 6]

As shown in Table 6, in this hypothesis example, the context X of the selected target word is three target words preceding the selected target word in the target hypothesis and three target words following the selected target word. Composed of target words. This context also includes punctuation and the absence of words.

In general, the context of the target word eaj of the target hypothesis E combined with the source word fj is included in at least one of two or more context classes. The estimated conditional occurrence probability of a source word given the target word of the target hypothesis connected to the source word and the context of the target word connected to the source word is It can include at least one function whose value depends on the class that contains the context.

Alternatively, the context may include at least one word having a target hypothesis part of speech. The estimated conditional occurrence probability of a source word given the target word of the target hypothesis connected to the source word and the context of the target word connected to the source word is At least one function can be included that has a value that depends on the part-of-speech of the target hypothesis of at least one word of the context.

In another example, the target word context of the target hypothesis bound to the source word includes at least one word having an identification. The estimated conditional occurrence probability of a source word given a target word of the target hypothesis connected to the source word and a context of the target word connected to the source word is
Includes at least one function having a value that depends on the identification of at least one word in the context of the target word coupled to the source word.

Equations 3, 4, 5 and 6 are hypothetical examples of functions whose values depend on the context of the target word bound to the source word.

G1 (f, e = key, X) = 1, f = “clef” and the word x3 in X immediately before “key” is “the” (3) = 0, If not above

## EQU00007 ## g2 (f, e = key, X) = 1, f = "clef", and the word x3 in X immediately before "key" is "car" (4) = 0, If not above

G3 (f, e = key, X) = 1, f = “ton” and word X of X immediately before “key” is “the” (5) = 0, not above If

## EQU00009 ## g4 (f, e = key, X) = 1, f = "ton", and a word x4 in X immediately after "key" or in X next to "key" in X When the word x5 is one element of the set {A, B, C, D, E, FG} (6) = 0, otherwise

In Equation 3, the function g1 is the source word f
Is “clef”, the target word e is “key”, and the word in the context X on the straight line of “key” is “the”, it has a value of 1. When these conditions are not met, the context function g1 has a value of 0.

The hypothetical context function g2 of the mathematical expression 4 is the source word f
Is “clef”, the target word e is “key”, and the word in the context X immediately before “key” is “car”, it has a value of 1. If these conditions are not met, the function g2
Has a value of 0.

[Mathematical formula-see original document] In the mathematical expression 5, in the context function g3, the source word f is "ton", the target word e is "key", and "key".
Has a value of 1 if the word in context X immediately preceding the is "the". When these conditions are not met, the context function g3 has a value of 0.

Finally, the hypothetical context function g4 in Equation 6
Is a word in context X that comes immediately after "key" and whose source word f is "ton" and target word e is "key", or a word in context X that comes next to "key" Is a set {A, B, C,
When it is one element of D, E, F, G}, it has a value of 1.
When these conditions are not met, the context function g4 has a value of 0.

Table 7 shows the source word f = “clef” and the target word e.
= ”Key”, the target word “key” of the target hypothesis E1 in Table 2
3 shows an evaluation of a context function g (f, e, X) for a context X of.

[Table 7]

As shown in Table 7, the context function g1 has a value of 1, and the context functions g2, g3, g4 have a value of 0.

As described above, the word match score of each source word and each target hypothesis is given by the target word e of the target hypothesis connected to the source word f and the context of the target word e connected to the source word f. It contains an estimate of the conditional occurrence probability P (f | e, X) of the source word f when X is given. The word match score can be obtained using, for example, a model defined by the following mathematical formula.

[Equation 10]

In equation 7, the function gi (f, e, X) is
A function having a value that depends on the context X of the combined target word e of the source words f in the alignment between the input series of source words and the target hypothesis. Parameter λ
(e, i) is a parameter indicating the relative strength of each context function gi when predicting the source word f from the target word e in the context X. The quantity N (e, X) is a normalization coefficient that depends on the target word e and the context X of the target word e, as shown in Expression 8.

[Equation 11]

When the target word e = “key”, the above equation 3
For hypothetical context functions g1 to g4 of ~ 6, Equation 9 is a hypothesis model that produces a word match score for the source word f.

[Equation 12]

With respect to Equation 9, the normalized N (e, X) is given by Equation 10.

[Equation 13] (10)

In order to explain how to use this model, the hypothetical values of the model parameters are λ (e = key, 1) = 0.12 and λ (e = e
key, 2) = 0.34, λ (e = key, 3) = 0.09, λ (e = key, 4) = 0.40
And

Table 8 shows the source word "clef" and the target word "key".
In the case of, the calculation of the word matching score from Expressions 7 and 9 for the context X of "key" in the target hypothesis E1 of Table 2 is shown.

[Table 8]

[Equation 14]

In this hypothesis example, the target word "key" of the target hypothesis combined with the source word "clef" is given and the source word "key" is given.
Conditional occurrence probability P ("clef" of the source word when the context of the target word "key" of the target hypothesis linked to clef "is given
| "Key", X) is 0.507 (from equations 9 and 10).

When the alignment identifying mechanism 18 identifies two or more alignments between the input sequence of source words and each target hypothesis, the word match score generator will use each source word and each alignment and each target. Generate a word match score for the hypothesis. Each word match score can be estimated using the model of Equation 7.

When the source text input device 10 includes means for converting a series of input source words into a series of converted source words, the word matching score generating mechanism causes the word matching score generation unit to match the words for each converted source word and each target hypothesis. Generate a score.
The word match score is the conditional occurrence probability of the converted source word given the target word of the target hypothesis connected to the converted source word and the context of the target word of the target hypothesis connected to the converted source word. Including estimated value. The word match score of the converted word can also be estimated using the model of Equation 7.

The translation apparatus according to the present invention further includes a translation match score generation mechanism 22. Translation match score generation mechanism 2
2 generates, for each target hypothesis, a translation match score that includes a combination of the target hypothesis and the word match score of the source word in the series of source words input. The translation match score for the target hypothesis includes, for example, the product of the word match scores of the source words in the series of source words input with the target hypothesis.

Table 9 shows the calculation of the translation match score for the input series of source words fj of Table 1 and the target word eaj of the target hypothesis E1 of Table 2.

[Table 9]

[Equation 15] (In order to explain the present invention, a fertility score and a distortion score are assumed in this example, and the combination coefficients are all 1).

The word matching score P (fj | eaj, X) is obtained from the equation 7. The numbers in Table 9 are hypothetical numbers. To illustrate the present invention, the Fertility Score and Strain Score of the above related patent applications are assumed to be one.

Referring again to FIG. 1, the translation device includes a hypothesis matching score generator 24 which generates a target hypothesis matching score for each target hypothesis. Each target hypothesis match score includes a combination (eg, product) of the target model hypothesis language model match score and the target hypothesis translation match score. As mentioned above, the linguistic model match score can also be obtained from known linguistic models, such as the n-gram linguistic model described by the above-referenced patent applications.

When the alignment identifying mechanism 18 identifies two or more alignments between the input series of source words and each target hypothesis, the translation match score of each target hypothesis is the word match of the target hypothesis. It includes a combination of a score, an alignment and a source word in the set of source words entered.

When the source text input device 10 includes means for converting a series of input source words into a series of converted source words, the translation match score is a combination of the word match score of the target hypothesis and the converted source word. Including combinations.

The translation device according to the present invention further includes an output unit 26.
including. The output unit 26 outputs the target hypothesis having the best target hypothesis matching score. The output unit 26 can include, for example, a display device or a printer.

If the source text input device 10 includes means for converting a series of input source words into a series of converted source words, the output unit 26 outputs a series of target hypotheses having the best target hypothesis matching score. It includes means for synthesizing output words. The means for synthesizing a series of output words from a target hypothesis with a target hypothesis match score may include a target transformation mechanism as described in Section 5 of the above-referenced patent application. For example, the target word "be" with a tag indicating the tense of the original source input word is a synthetic output word {be, was, were, been, am, are, is, being which is a variation of the infinitive "be". } Is converted into one of.

As mentioned above, Equation 7 is an example of a model that can be used to obtain a word match score in accordance with the present invention. Context function g i (f,
e, X) and the parameter λ of the word match score model
(e, i) can be obtained as follows.

The candidate context function gi (f, e, X) is, for example, as shown in the example of Table 6 above, the context X of the target word in the target hypothesis is defined by the three words to the left of the target word e and the target word e.
It can be obtained by limiting to the right 3 words of.

Next, the training texts of the source language sentence and the target language sentence, which are mutual translations, are obtained. Corresponding source language sentences and target language sentences, which are mutual translations, can be fixed, for example by a trained translator. Corresponding source and target language sentences, which are mutual translations, are preferably automatically identifiable, for example by the methods described in sections 12 and 13 of the above-referenced patent applications.

For each corresponding source and target sentence pair in the training text, the estimated most probable alignment between the source and target words is calculated using the method described above used by the alignment identification mechanism 18. Find out. Next, each occurrence of the target word eaj in the training text is identified, and the target word eaj in each alignment is identified.
Tag with the source word fj combined with aj. The target word eaj is also tagged with the context X in each target sentence.

Table 10 shows a hypothetical example of a training event for the target word eaj = "key" in the training text of the aligned source language sentence and target language sentence.

[Table 10]

The candidate context function gi (f, e, X) is calculated using the training event of the target word eaj from the training text,
It is obtained by first identifying the form of the context function.
For example, one form of context function tests for the presence of a particular word at one or more positions in the context. Another form of context function tests for the presence of a particular word class (eg, part of speech) at one or more positions in the context. The specific word or word class to be tested in the context function of the target word e is the target word e in the training text.
Can be obtained from a word or word class in the context of.

Alternatively, the candidate context function gi (f, e,
X) can be obtained by classifying the training events according to context X using the method described in section 7 of the related patent application above.

First, all the parameters λ (e, i) of the word match score model of Equation 7 are set to zero.

For each candidate context function gi (f, e, X), the “measure of merit” G (i) is calculated according to equation 11.

[Equation 16]

[Equation 17]

[Equation 18]

[Formula 19]

In equations 11 to 14, the combined target word e
The conditional probability P (f | X) of the source word f given the context X of is obtained from Eq. 7 using the most recent values of the model parameters. Probability P (X) of context X
Is obtained by counting the occurrences of the target word e and context X in a training text event of the form shown in Table 10, and dividing that count by the total number of events in the training text with the target word e. The probability P (f, X) between the source word f and the context X counts the occurrences of the target word e, the context X, and the source word f in a training text event of the form shown in Table 10, and each count is the target word Obtained by dividing by the total number of events in the training text, where e is e.

“Measures of Merit” from Equation 11 G
The context function gi (f, e, X) with the highest (i) is
Selected as the context function to be used in. Parameter λ
For (e, i), first set λ (e, i) = 0 and then the quantity Δλ (e,
It is obtained by solving the following Expression 15 for i).

[Equation 20]

The new value of λ (e, i) is equal to the previous λ (e, i) by Δλ.
Obtained by adding (e, i). Then, using the new value of λ (e, i), Equation 1 for the new value of Δλ (e, i)
Solve 5. This process is repeated until the value of Δλ (e, i) is less than or equal to the selected limit value. This method is known as iterative scaling.

New model of word match score (Equation 7)
To recalculate the "measure of merit" G (i) of Equation 11 for the remaining candidate context functions g i (f, e, X), and the remaining with the highest "measure of merit" G (i) Identifies the context function of. The best remaining context function is added to the word match score model of Equation 7 and all new values of the parameter λ (e, i) are calculated using the iterative scaling method and Equation 15. When the word match score in Expression 7 includes two or more parameters λ (e, i), each parameter λ (e, i)
Is updated exactly once per iteration, so all parameters λ (e, i) converge at the same iteration. This process continues until the best context function "measure of merit" is less than or equal to the selected limit.
Iterate over (f, e, X).

As a summary, the following matters will be disclosed regarding the configuration of the present invention.

(1) An apparatus for translating a series of source words in a first language into a series of target words in a second language different from the first language, and means for inputting the series of source words, and The target hypothesis comprises a series of target words selected from a vocabulary of words in a second language, each target word being at least one of the target hypotheses.
Means for generating at least two target hypotheses having a context that includes two other words; and for each target hypothesis, generating a language model match score that includes an estimate of the probability of occurrence of a series of words of the target hypothesis. Means for identifying at least one alignment between the input series of source words and each target hypothesis, which combines each source word with at least one target word of the target hypothesis; and for each source word and each target hypothesis, A word match that contains an estimate of the conditional occurrence probability of the source word given the target word of the target hypothesis connected to the source word and the context of the target word of the target hypothesis connected to the source word A translation match score, including means for generating a score and for each target hypothesis, a combination of the target hypothesis word match score and the source word in the series of source words entered. A means for generating, and a means for generating a target hypothesis matching score for each target hypothesis, each target hypothesis matching score including a combination of a language model match score for the target hypothesis and a translation match score for the target hypothesis; Means for outputting a target hypothesis having a score. (2) The apparatus according to (1) above, wherein each target hypothesis includes a series of target words selected from a vocabulary containing words in a second language and a blank word indicating no words. (3) A means for identifying at least one alignment comprises two or more between an input sequence of source words and each target word, each combining each source word with at least one target word of the target hypothesis. Identify the alignment,
For each source word, each alignment, and each target hypothesis, the word matching score generation mechanism estimates the conditional occurrence probability of the source word given the target word connected to the source word and the context of the target word. A translation match score generating a word match score, including values, wherein the translation match score generation mechanism includes, for each target hypothesis, a combination of the target hypothesis word match score and a source word in the set of source words. The device according to (2) above, which is characterized in that: (4) A series of input means includes means for converting the input series of source words into a series of converted source words, and the alignment means combines each converted source word with at least one target word of the target hypothesis. Identifying at least one alignment between the transformed source word and each target hypothesis, and for each transformed source word and each target hypothesis, a target word of the target hypothesis coupled to the transformed word is provided. And, when the context of the target word of the target hypothesis coupled to the converted source word is given, including the estimated value of the conditional occurrence probability of the converted source word, to generate a word match score, the translation match score generation mechanism, For the target hypothesis, including the combination of the target hypothesis word match score and the converted source word,
Generating a translation match score, the output means including means for synthesizing a series of output words from the target hypothesis having the best target hypothesis match score; and output means for outputting the synthesized output word. The device according to (2). (5) The translation match score of the target hypothesis includes the product of the word match score of the target hypothesis and the source word in the series of input source words, and the target hypothesis match score of the target hypothesis matches the language model of the target hypothesis. The apparatus according to (2) above, which includes a product of the score and a translation match score of the target hypothesis. (6) The target word context in the target hypothesis coupled to the source word is included in at least one of two or more context classes, the target word coupled to the source word is given, and the context of the target word is given. The apparatus according to (2) above, wherein the estimated conditional occurrence probability of the source word when given includes at least one function having a value that depends on a class that includes the context of the target word. (7) The context of the target word of the target hypothesis coupled to the source word includes at least one word having the part of speech of the target hypothesis, the target word coupled to the source word is given, and the context of the target word is given. And the estimated probability of occurrence of the source word comprises at least one function having a value that depends on the part of speech of the target hypothesis of at least one word in the context of the target word coupled to the source word, The device according to (2). (8) In the case where the context of the target word of the target hypothesis coupled to the source word includes at least one word having an identification, the target word coupled to the source word is given, and the context of the target word is given. (2) above, wherein the estimated conditional occurrence probability of the source word comprises at least one function having a value dependent on the identification of at least one word in the context of the target word bound to the source word The described device. (9) The above-mentioned (2), wherein the means for outputting the target hypothesis having the best match score includes a display.
The device according to. (10) The apparatus according to (2) above, wherein the input means includes a keyboard. (11) The apparatus according to (2) above, wherein the input means includes a computer disk drive mechanism. (12) The apparatus according to (2) above, wherein the input means includes a computer tape drive mechanism. (13) A method of translating a series of source words in a first language into a series of target words in a second language different from the first language,
Inputting a sequence of source words, each target hypothesis comprising a sequence of target words selected from a vocabulary of words in a second language, each target word having a context comprising at least one other word of the target hypothesis , Generating at least two target hypotheses; generating, for each target hypothesis, a language model match score that includes an estimate of the probability of occurrence of a series of words in the target hypothesis; Identifying at least one alignment between the input series of source words and each target hypothesis that is associated with one target word, and for each source word and each target hypothesis, the target hypothesis associated with the source word. Including an estimate of the conditional occurrence probability of the source word given the target word and the context of the target word of the target hypothesis coupled to the source word, Generating a word match score, generating, for each target hypothesis, a translation match score that includes a combination of the target hypothesis word match score and a series of source words of the source word, and a language model of the target hypothesis. Generating a target hypothesis match score for each target hypothesis, including a combination of the match score and the translation match score of the target hypothesis; and outputting the target hypothesis with the best target hypothesis match score. (14) The method according to (13) above, wherein each target hypothesis includes a series of target words selected from a vocabulary containing words in a second language and a blank word indicating no words. (15) The step of identifying at least one alignment comprises two or more alignments between the input series of source words and each target word, each source word being combined with at least one target word of the target hypothesis. The step of generating a word match score includes the step of identifying, for each source word and each alignment and each target hypothesis, the source word of the source word given the target word associated with the source word and given the context of the target word. A translation match score generation step that includes generating a word match score, including an estimate of the conditional probability of occurrence, and for each target hypothesis, a combination of the word match score of the target hypothesis and a source word of the input series of source words. Generating a translation match score that includes the method according to (14) above. (16) The input step includes a step of converting the input series of source words into a series of converted source words, and the alignment step combines each converted source word with at least one target word of a target hypothesis. Of at least one alignment between each transformed source word and each target hypothesis, wherein a word match score generation step includes, for each transformed source word and each target hypothesis, a goal of the target hypothesis associated with the transformed word. Generating a word match score, including an estimate of the conditional occurrence probability of the transformed source word given the word and the context of the target word of the target hypothesis coupled to the transformed source word,
The translation match score generation step, for each target hypothesis,
Generating a translation match score that includes a combination of a target hypothesis word match score and a transformed source word, the output step synthesizing a series of output words from the target hypothesis having the best target hypothesis match score; The method according to (14) above, including an output step of outputting a combined output word. (17) The translation match score of the target hypothesis includes the product of the word match score of the target hypothesis and the source word in the series of input source words, and the target hypothesis match score of the target hypothesis matches the language model of the target hypothesis. The method according to (14) above, comprising a product of the score multiplied by the translation match score of the target hypothesis. (18) A target word context in a target hypothesis coupled to a source word is included in at least one of two or more context classes, a target word coupled to the source word is given, and a target word context is given. Method according to (14) above, characterized in that the estimated conditional occurrence probability of the source word when given comprises at least one function having a value that depends on the class containing the context of the target word. (19) The context of the target word of the target hypothesis coupled to the source word includes at least one word having the part of speech of the target hypothesis, the target word coupled to the source word is given, and the context of the target word is given. The estimated conditional occurrence probability of the source word in the case of including at least one function having a value that depends on the part-of-speech of the target hypothesis of at least one word in the context of the target word bound to the source word. The method according to (14) above. (20) if the context of the target word of the target hypothesis coupled to the source word includes at least one word having an identification, the target word coupled to the source word is given and the context of the target word is given. (14) above, wherein the estimated conditional occurrence probability of the source word comprises at least one function having a value that depends on the identification of at least one word in the context of the target word bound to the source word. The method described. (21) The method according to (14) above, wherein the step of outputting the target hypothesis having the best match score includes the step of displaying the target hypothesis having the best match score.

[0092]

In the translation device according to the present invention, the target hypothesis generating mechanism 12, the language model matching score generating mechanism 16, the alignment identifying mechanism 18, the word matching score generating mechanism 2 are used.
0, the translation match score generator 22, and the hypothesis match score generator 24 can be appropriately programmed general purpose or special purpose digital signal processors. The target language vocabulary store 14 may be a computer storage device such as a random access memory. Means for converting a series of inputted source words into a series of converted source words in the source text input device 10 and means for synthesizing a series of output words from a target hypothesis having a best target hypothesis matching score in the output unit 26 are also suitable. It may be a general-purpose or special-purpose digital signal processor programmed in.

[Brief description of drawings]

FIG. 1 is a block diagram of an example of an apparatus for translating words from a first language to a second language according to the present invention.

FIG. 2 is a schematic diagram showing an example of an alignment between a series of hypothesis source words and a series of hypothesis target words.

3 is a schematic diagram showing an example of a second alignment between the series of hypothetical source words and the series of hypothetical target words in FIG.

4 is a schematic diagram showing an example of a third alignment between the series of hypothetical source words and the series of hypothetical target words in FIG.

[Explanation of symbols]

 10 Primitive Text Input Device 12 Target Hypothesis Generating Mechanism 14 Target Language Vocabulary Storage 16 Language Model Matching Score Generating Mechanism 18 Alignment Identifying Mechanism 20 Word Matching Score Generating Mechanism 22 Translation Matching Score Generating Mechanism 24 Hypothesis Matching Score Generating Mechanism 26 Output Unit

 ─────────────────────────────────────────────────── ——————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————-- # # # # # # # # # # # # # # # # # # # OF # # # = # # # # # # # # # # # # # # # # # # # # # # # # # # # # 2 Pall River Mayer Oval 113 (72) Inventor Vincent Joseph de la Pietra United States 10913 Braubert Sunset Road 129 (72) Inventor Andrew Scott Keller United States 02143 Somerville Massachusetts Beacon Street 326 Number 2 (72) Inventor Robert Leroy Mercer USA Yo Shu countries 10598 New York Kutaun Heights view Land drive 669

Claims (21)

[Claims] 1. An apparatus for translating a series of source words in a first language into a series of target words in a second language different from the first language, comprising means for inputting the series of source words and each target. Means for generating at least two target hypotheses, wherein the hypothesis comprises a series of target words selected from a vocabulary of words of a second language, each target word having a context comprising at least one other word of the target hypothesis; For each target hypothesis, a means for generating a linguistic model match score containing an estimate of the probability of occurrence of the set of words in the target hypothesis, and a set of input sequences connecting each source word to at least one target word in the target hypothesis. Means for identifying at least one alignment between a source word and each target hypothesis, and for each source word and each target hypothesis, a target word of the target hypothesis linked to the source word is provided, and Means for generating a word match score, including an estimate of the conditional probability of occurrence of a source word given the context of the target word of the target hypothesis coupled to the word, and for each target hypothesis, the word match of the target hypothesis Including a combination of a score and a source word in the set of source words entered,
A means for generating a translation match score, a means for generating a target hypothesis match score for each target hypothesis, each target hypothesis match score including a combination of the target model hypothesis language model match score and the target hypothesis translation match score; Outputting a target hypothesis having a goal hypothesis match score of. 2. The apparatus according to claim 1, wherein each target hypothesis comprises a set of target words selected from a vocabulary containing words in a second language and a blank word indicating no words. . 3. A means for identifying at least one alignment, two between an input sequence of source words and each target word, each combining each source word with at least one target word of a target hypothesis. The above-mentioned alignment is identified, and the word match score generation mechanism, for each source word, each alignment, and each target hypothesis, the source word when the target word connected to the source word is given and the context of the target word is given. Generates a word match score, which includes an estimate of the conditional occurrence probability of, and for each target hypothesis, the translation match score generation mechanism compares the word match score of the target hypothesis with the source word in the set of source words. 3. The device according to claim 2, characterized in that it produces a translation match score containing the combinations. 4. The input means includes means for converting the input series of source words into a series of converted source words, and the alignment means combines each converted source word with at least one target word of a target hypothesis. Identifying at least one alignment between the series of transformed source words and each target hypothesis, and the word match score generator determines, for each transformed source word and each target hypothesis, the target word of the target hypothesis associated with the transformed word. Generates a word match score, which includes an estimate of the conditional occurrence probability of the converted source word given the context of the target word of the target hypothesis that is given and combined with the converted source word, For each target hypothesis, a translation match score is generated that includes a combination of the target hypothesis word match score and the converted source word, and the output means outputs the best target hypothesis match score. 3. The apparatus according to claim 2, comprising means for synthesizing a series of output words from a target hypothesis having a, and output means for outputting the synthetic output word. 5. The translation match score of the target hypothesis includes the product of the word match score of the target hypothesis and the source word in the series of input source words, and the target hypothesis match score of the target hypothesis is the language of the target hypothesis. 3. The apparatus of claim 2 including a product of the model match score times the translation match score of the target hypothesis. 6. A target word context in a target hypothesis bound to a source word is included in at least one of two or more context classes, a target word bound to the source word is given and a context of the target word. 3. An apparatus according to claim 2, characterized in that the estimated conditional occurrence probability of a source word given is given at least one function whose value depends on the class containing the context of the target word. 7. The context of a target word of a target hypothesis bound to a source word comprises at least one word having a part of speech of the target hypothesis, a target word bound to the source word is given and the context of the target word is Given probability of occurrence of a given source word,
3. Include at least one function having a value that depends on the part-of-speech of the target hypothesis of at least one word in the context of the target word coupled to the source word.
The device according to. 8. A target word context of a target hypothesis bound to a source word comprises at least one word having an identification, a target word bound to the source word is given and a target word context is given. 3. The case where the estimated conditional occurrence probability of a source word comprises at least one function having a value that depends on the identification of at least one word in the context of the target word bound to the source word. The device according to. 9. The apparatus of claim 2, wherein the means for outputting the target hypothesis with the best match score comprises a display. 10. The apparatus of claim 2, wherein the input means comprises a keyboard. 11. The apparatus of claim 2 wherein the input means comprises a computer disk drive. 12. The apparatus of claim 2 wherein the input means comprises a computer tape drive. 13. A method of translating a series of source words in a first language into a series of target words in a second language different from the first language, the method comprising inputting a series of source words, each target hypothesis comprising: Generating at least two target hypotheses comprising a set of target words selected from a vocabulary of words in a second language, each target word having a context containing at least one other word of the target hypotheses; For a hypothesis, generating a language model match score that includes an estimate of the probability of occurrence of a set of words in the target hypothesis, and an input set of source words that joins each source word to at least one target word in the target hypothesis. Identifying at least one alignment between the source hypothesis and each target hypothesis, and for each source word and each target hypothesis, a target word of the target hypothesis coupled to the source word is given. And generating a word match score that includes an estimate of the conditional occurrence probability of the source word given the context of the target word of the target hypothesis coupled to the source word, and for each target hypothesis, Generating a translation match score that includes a combination of the word match score and the source word of the input sequence of source words, and each goal that includes a combination of the target model hypothesis model match score and the target hypothesis translation match score. Generating a target hypothesis match score for the hypothesis, and outputting the target hypothesis with the best target hypothesis match score. 14. The method of claim 13, wherein each target hypothesis comprises a set of target words selected from a vocabulary containing words in a second language and a blank word indicating no words. . 15. The step of identifying at least one alignment includes assigning each source word to at least one of the target hypotheses.
A step of identifying two or more alignments between each target word and an input sequence of source words associated with one target word, the word matching score generating step comprising: each source word and each alignment and each target; For a hypothesis, including a target word coupled to the source word and an estimate of the conditional occurrence probability of the source word given the context of the target word,
Including generating a word match score, the translation match score generation step for each target hypothesis
15. The method of claim 14, comprising: generating a translation match score that includes a combination of the target hypothesis word match score and the source word of the input series of source words. 16. The input step includes the step of converting the input series of source words into a series of converted source words, and the alignment step combines each converted source word with at least one target word of a target hypothesis. , A step of identifying at least one alignment between a series of transformed source words and each target hypothesis, wherein a word match score generation step comprises, for each transformed source word and each target hypothesis, a target hypothesis associated with the transformed word. Generating a word match score, including an estimate of the conditional occurrence probability of the transformed source word given the target words of and the context of the target word of the target hypothesis combined with the transformed source word. , Translation match score generation step, for each target hypothesis,
Generating a translation match score that includes a combination of a target hypothesis word match score and a transformed source word, the output step synthesizing a series of output words from the target hypothesis having the best target hypothesis match score; 15. The method according to claim 14, characterized in that it comprises an output step for outputting a synthetic output word. 17. The translation match score of the target hypothesis includes the product of the word match score of the target hypothesis and the source word in the series of input source words, and the target hypothesis match score of the target hypothesis is the language of the target hypothesis. 15. Method according to claim 14, characterized in that it comprises the product of the model match score times the translation match score of the target hypothesis. 18. A target word context in a target hypothesis associated with a source word is at least one of two or more context classes.
, The estimated conditional occurrence probability of the source word given the target word combined with the source word and the context of the target word is a value that depends on the class including the context of the target word. 15. Method according to claim 14, characterized in that it comprises at least one function having. 19. The context of a target word of a target hypothesis bound to a source word comprises at least one word having a part of speech of the target hypothesis, a target word bound to the source word is given, and the context of the target word is Estimated conditional occurrence probability of the source word, if given, comprises at least one function having a value that depends on the part of speech of the target hypothesis of at least one word in the context of the target word bound to the source word. And
The method according to claim 14. 20. A target word context of a target hypothesis bound to a source word comprises at least one word having an identification, a target word bound to a source word is given and a target word context is given. 15. The case where the estimated conditional occurrence probability of a source word comprises at least one function having a value that depends on the identification of at least one word in the context of the target word bound to the source word.
The method described in. 21. The step of outputting the target hypothesis with the best match score comprises displaying the target hypothesis with the best match score.
The method according to 4.