JP2014216029A - Device and method for evaluating word phrase of intermediate language, and device and method for machine translation - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device and a method for evaluating a word phrase of an intermediate language, a machine translation method, and a machine translation device.SOLUTION: A method for evaluating a word phrase of an intermediate language includes the steps of: determining a first specific attribute exhibited by a word phrase of an intermediate language in regard to a source language; determining a second specific attribute exhibited by the word phrase of the intermediate language in regard to a target language; calculating a reliability score of the word phrase of the intermediate language, on the basis of the first specific attribute and the second specific attribute; and evaluating the word phrase of the intermediate language, on the basis of the reliability score. The word phrase of the intermediate language serves as a bridge for translating a specific word phrase of the source language to a word phrase of the target language.

Description

  The present invention relates to the field of language processing, and in particular, to a method and apparatus for evaluating phrases in an intermediate language, and a machine translation method and apparatus.

  Parallel vocabulary information occupies an important position in cross language application (machine translation, acquisition of cross language information, etc.). However, it cannot be said that parallel vocabulary information of any two types of words can be acquired, and it is quite difficult to obtain parallel vocabulary information related to words that have a very limited range of use. Also, since new words and words come out more and more, it is quite difficult to adapt to the development of new words and words that come out with parallel vocabulary information already known. Therefore, the concept of intermediate language (Pivot Language) has been introduced. Supplement more parallel vocabulary information with an intermediate language. However, for the use of intermediate languages, there are mainly two main obstacles: the first is ambiguity and the second is inconsistent.

  In order to solve the ambiguous problems of Source Language, Intermediate Language, and Target Language, the traditional method uses structured bilingual dictionaries, semantic classes, multiple intermediate language words, and related frequencies. Information such as edit distance is used. In the conventional method, the ambiguity problem can be consistently reflected by the translation probability between the source language and the intermediate language and the translation probability between the intermediate language and the target language, and the highest probability is selected. We believe that by doing so we can solve ambiguous problems.

  The conventional method of obtaining bilingual information by an intermediate language basically follows the following process. First, information on the source language and the intermediate language, and information on the intermediate language and the target language, for example, translation probability, lexicalized translation probability, editing distance, semantic information, and the like are obtained. Next, based on this information, the most reliable pair of source language phrases and target language phrases is selected. However, when the intermediate language includes a plurality of meanings, the conventional method does not perform special processing for this type of case, and still selects the phrase having the highest translation probability as the final result. This type of method leaves behind the feature that the source and target languages from non-parallel language material do not have the same meaning, so if the intermediate language phrase is ambiguous, the translation probability is the source language The semantic relationship between the phrase and the target language phrase cannot be reflected.

  Therefore, a technique that can solve the above-described problem is desirable.

  An object of the present invention is to provide a method for evaluating an intermediate language phrase, an apparatus for evaluating an intermediate language phrase, a machine translation method, and a machine translation apparatus.

  According to one aspect of the invention, a method for evaluating an intermediate language phrase is provided. The method includes intermediate language phrases determining first specific attributes for the source language; intermediate language phrases determining second specific attributes for the target language; based on the first specific attributes and the second specific attributes, intermediate Calculating a confidence score for the language phrase; and evaluating an intermediate language phrase based on the confidence score, wherein the intermediate language phrase converts the source language specific phrase into the target language phrase It is a bridge for translation.

According to another aspect of the present invention, a machine translation method is provided. The method
The intermediate language phrase reliability score is obtained by the intermediate language phrase evaluation method described above; the intermediate language phrase reliability score and the source language specific phrase via the intermediate language phrase as the target language Calculating the translation score of the target language candidate word based on the machine translation score translated into the candidate language; and selecting the target language word as the translation result from the target language candidate word based on the translation score including.

4 is a flowchart of a method for evaluating an intermediate language phrase in one embodiment of the present invention. It is a figure which shows calculation of the reliability score based on the artificial neural network in one Example of this invention. 6 is a flowchart of a method for evaluating an intermediate language phrase according to another embodiment of the present invention. It is a flowchart of the machine translation method in one Example of this invention. 1 is a configuration block diagram of an apparatus for evaluating phrases in an intermediate language according to an embodiment of the present invention. It is a structure block diagram of the apparatus which evaluates the phrase of an intermediate language in the other Example of this invention. It is a structure block diagram of the apparatus which evaluates the phrase of an intermediate language in the other Example of this invention. 1 is a configuration block diagram of a machine translation apparatus in an embodiment of the present invention. It is a block diagram of the example of the calculation apparatus for implementing or implement | achieving the method and apparatus, the machine translation method, and the machine translation apparatus of evaluating the phrase of an intermediate language in this invention.

  Embodiments of the present invention will be described below with reference to the drawings.

  In some cases, there is not enough information to directly associate source language A and target language B, information that directly associates intermediate language E and source language A, and directly associates intermediate language E and target language B. There may be information for In this type of case, consider that one or more phrases P from the intermediate language E are selected as bridges, and that specific phrases S in the source language A are translated into corresponding phrases in the target language B. May be. In other words, the phrase P in the intermediate language E may be a bridge for translating the specific phrase S in the source language A into the corresponding phrase T in the target language B.

  FIG. 1A is a flowchart of a method 100 for evaluating a phrase P of an intermediate language E in one embodiment of the present invention. FIG. 1B is a diagram showing calculation of a reliability score based on an artificial neural network in one embodiment of the present invention.

  As shown in FIG. 1A, in step S102, each phrase P of the intermediate language E determines a specific attribute (that is, a first specific attribute) for the source language A. In other words, the specific attribute that the phrase P exhibits in the source language A is determined.

  In the first specific attribute, each phrase P of the intermediate language E may include a meaning range in the source language A (that is, a first meaning range). The greater the meaning of the word P in the source language A, the greater the first meaning range. For example, the first semantic range may be the number of words corresponding to each word P of the intermediate language E in the source language A, or a function of the number. In some cases, the number of phrases corresponding to the phrase P of the intermediate language E in the source language A exhibits exponentiality, so one function of the number is taken so that the obtained result has linearity. Also good. For example, the function may be logarithmic.

  In step S104, each phrase P of the intermediate language E determines a specific attribute (that is, a second specific attribute) for the target language B. In other words, the specific attribute that the phrase P exhibits in the source language A is determined.

  In the second specific attribute, each phrase P of the intermediate language E may include a meaning range in the target language B (that is, a second meaning range). The greater the meaning of the word P in the target language B, the greater the range. For example, the second semantic range may be the number of phrases corresponding to each phrase P of the intermediate language E in the target language B, or a function of the number. Similarly, the function may be logarithmic.

  In step S106, the reliability score of each word P of the intermediate language E is calculated based on the first specific attribute and the second specific attribute.

  In one embodiment, the first specific attribute and the second specific attribute are characterized, and the reliability score of each phrase P of the intermediate language E may be calculated by, for example, a regression algorithm.

  Preferably, in one embodiment, the reliability score of each word P of the intermediate language E is calculated by an artificial neural network (ANN) algorithm characterized by the first specific attribute and the second specific attribute. Good.

An artificial neural network is essentially a kind of simple mathematical model that can be defined by the equation f: x | y. The term network in an artificial neural network refers to the interconnection between the nerve elements in the nerve elements of each layer in each system. One typical artificial neural network has three layers of nerve elements. The first layer is the input source and is used to receive data and send it to the second layer. Thereafter, the data is sent to the output nerve of the third layer by the second layer nerve whose number is greater than the number of the first layer. More complex artificial neural networks may have more layers. The parameter of storage at each nerve element is called weight, and the weight is calculated with data in the data transmission process. In mathematics, an equation f (x) for one neural network is a combination of a series of equations g _m (x). g _m (x) may be defined as a combination of other series of equations. This equation may be represented as a network configuration diagram as shown in FIG. 1B, where the arrows in this configuration diagram indicate the dependency between each variable, and Pr (S | P) is a phrase in the intermediate language E P represents the translation probability of a specific phrase in source language A, Pw (S | P) represents the lexicalization probability of a phrase in intermediate language E into a specific phrase in source language A, and S (P) Represents that the first meaning range and the second meaning range are included. In FIG. 1B, Pr (S | P), Pw (S | P), and S (P) are shown simultaneously, but only S (P) is used, and Pr (S | P) and Pw (S | P) may not be used.

In an artificial neural network, “non-linear weight sum” is a kind of widely used combination of functions, and is expressed as follows.

F (x) represents the reliability score, K represents the activation function, g _m (x) represents the value of the mth feature, and w _m represents the mth feature. The weight represents bias weight, biasW represents the bias weight, and biasV represents the bias value. In one embodiment, tanh (x) is used as the activation function.

  Thus, the reliability score can be calculated by the artificial neural network. The reliability score may be calculated by other regression methods or other appropriate methods.

  In one embodiment, the reliability of the intermediate language may be defined as the balance between the likelihood that the phrase P becomes a bridge for translating the phrase S into the corresponding phrase T and the semantic accuracy. For example, in the embodiment described above, a regression calculation is performed using an artificial neural network to determine the balance between possibility and semantic accuracy, and finally each word P in the intermediate language E The corresponding reliability score is calculated for.

In one embodiment, the reliability score may be calculated using the first semantic range and the second semantic range as features. In other embodiments, the first semantic range, the second semantic range, the translation probability of the phrase P of the intermediate language E to the specific phrase S of the source language A, and / or the source language A of the phrase of the intermediate language E The reliability score may be calculated using the probability of lexical translation into a specific phrase S as a feature. In other words, in the embodiment, in addition to the first semantic range and the second semantic range, the first specific attribute further includes a translation probability of the phrase P of the intermediate language E to the specific phrase of the source language A, and / or The lexical translation probability of the phrase of the intermediate language E to the specific phrase of the source language A may be included. The translation probability Pr (S | P) of the phrase P of the intermediate language E to the specific phrase of the source language A and the phrase of the intermediate language E to the specific phrase of the source language A by any appropriate method The lexicalized translation probability Pw (S | P) may be obtained. For example, the translation probability Pr (S | P) of the phrase P in the intermediate language E to the specific phrase in the source language A can be calculated by the following equation (2).

Among them, the function N (S, P) represents the number of co-occurrences of the specific phrase S in the source language A and the phrase P in the intermediate language E, and S _i is all in the source language. Represents the phrase.

For example, the lexicalized translation probability Pw (S | P) of the phrase of the intermediate language E to the specific phrase of the source language A can be calculated by the following equation (3).

Among them, the phrase P of the intermediate language E consists of m words, each word is expressed as P _j (j = 1, 2,..., M), and the specific phrase S of the source language A consists of n words Each word is represented as S _i (i = 1, 2,..., N), and the function W (S _i , P _j ) represents the probability of translating the word P _j into the word S _i . In the calculation, it is assumed that the word level alignment relationship of the phrase P to the specific phrase S is known (for example, the first word in the phrase P corresponds to the first to third words of the specific phrase S Al represents alignment. The meaning of (i, j) ∈Al is that when the phrase P and a specific S are aligned, the word P _i and the word S _i correspond to each other. | ... | represents obtaining the number, so the fraction after the product code (Π) represents the number of words in P corresponding to Si, that is, for the result of the subsequent sum (Σ) To find the arithmetic mean. The product code (Π) specifies that the value of i is from 1 to n (ie, from the first word to the last word in the specific phrase S), and to the word Si in the phrase P For all the corresponding words, the probability of translating the word P _j in the phrase P into S _i is obtained, and the sum is obtained. By dividing the number of words P _j in the phrase P corresponding to S _i by this sum, one score is obtained. Then, the product of the obtained scores is set as the lexicalized translation probability of the phrase P to the phrase S.

  Next, in step S108, each phrase P of the intermediate language E is evaluated based on the reliability score.

  Hereinafter, the advantages of the present invention will be described by taking as an example that the source language is Chinese, the target language is Japanese, and the intermediate language is English. In one embodiment of the present invention, the phrase having the narrowest semantic range is selected from the intermediate languages. For example, suppose that a Chinese “water stream” is translated into a corresponding phrase in Japanese. In English, there are one or more selectable phrases, and in the conventional method, one phrase that is most likely to contain the meaning “water flow” is selected, so “flow” is the best choice. However, “flow” has a very low probability of being translated into Japanese “water flow”. Finally, if translated via “flow”, an incorrect Japanese translation is generated. In one embodiment of the present invention, “water flow” in the intermediate language is considered to be the most accurate because it includes the meaning of “water flow” in Chinese and at the same time “water flow” in Japanese. This is because the meaning of is accurately reflected. Therefore, based on the specific attribute in the source language A of the phrase P in the intermediate language E and the specific attribute in the target language B of the phrase P, the reliability with which the phrase P becomes a bridge can be evaluated. Therefore, it is possible to perform translation by selecting an intermediate language phrase having higher reliability as a bridge.

  FIG. 2 is a flowchart of a method 200 for evaluating the phrase P of the intermediate language E in another embodiment of the present invention.

  As shown in FIG. 2, in step S202, the specific phrase S of the source language A and the phrases in the phrase database of the intermediate language E are aligned to obtain at least one first phrase of the intermediate language E. In other words, the specific phrase S is aligned with a phrase having the meaning of the specific phrase S in the phrase database of the intermediate language E, and the obtained phrase is referred to as a first phrase for convenience.

  In step S204, at least one first alignment word / phrase is obtained by removing a portion not corresponding to the specific word / phrase S of the source language A from at least one first word / phrase. In other words, for each first word / phrase, only a part of the first word / phrase may correspond to the specific word / phrase S, and therefore, it is necessary to exclude the non-corresponding part. This word is referred to as a first alignment word. The part removed here refers to a part in one first phrase, and does not refer to some first phrases in a plurality of first phrases.

  In step S206, the phrases in the target language B phrase database and the phrases in the intermediate language E phrase database are aligned to obtain at least one second phrase in the intermediate language E. In some cases, there are several language materials, such language materials include the relationship between the target language B and the intermediate language E, but often the target language B and the intermediate language E are Since there is no one-to-one correspondence, it is necessary to align the target language B and the intermediate language E. For convenience, a phrase that can be aligned with a phrase in target language B in intermediate language E is referred to as a second phrase.

  In step S208, a portion not corresponding to the word / phrase in the word / phrase database of the target language B is removed from at least one second word / phrase to obtain at least one second alignment word / phrase. Similarly, for each second word, only a part may have correspondence with the words in the target language B, so it is necessary to remove the non-corresponding part. The phrase obtained by removing is referred to as a second alignment phrase. The portion removed here refers to a part of one second word, and does not refer to some second words in a plurality of second words.

  In step S210, the words in the common set of the above-described at least one first alignment word and the above-described at least one second alignment word are set as the evaluation waiting word of the intermediate language E. The phrase in the common set is a phrase corresponding to the phrase S in the source language A and has a correspondence with the phrase in the target language B. Therefore, the phrase in the common set uses the phrase S as the target language. There is a possibility that it will be the corresponding phrase when translating.

  For ease of understanding, the above process is described in detail below with an example.

  In this example, the Chinese language is the source language A, the specific phrase S is the Chinese word “josulfuric acid (meaning methionine)”, the intermediate language E is English, and the target language B is Japanese. To do.

  The alignment of the Chinese word “sulfuric acid” with the words in the English word database, the first words “methionine Promix, NEN, Boston, MA”, “-14C] -L-methionine” And so on. The first alignment phrase “methionine” is removed from the first phrase by removing parts that do not correspond to “sulfuric acid”, ie, “Promix, NEN, Boston, MA”, “-14C] -L-”, etc. Get. Here, only one word “methionine” corresponding to “sulfuric acid” is listed, but it is easy to understand that for some words, for example “gas”, the corresponding English There can be more than one number of phrases, for example, “air”, “gas”, etc., so it is possible to get one or more first phrases, and one or more first alignment phrases There is also a possibility to get.

  Next, the word / phrase in the Japanese word / phrase database is aligned with the word / phrase in the English word / phrase database, and the English word / phrase corresponding to each Japanese word / phrase is obtained as the second word / phrase, that is, This corresponds to obtaining a correspondence between a phrase and an English phrase. Similarly, a portion that does not correspond to a Japanese word is removed from an English word to obtain a second alignment word. Of course, the number of second alignment phrases may be greater than the number of first alignment phrases. Finally, the phrase in the common set of the first alignment word and the second alignment word is the English word that is waiting for evaluation, that is, the word whose reliability score needs to be calculated for the word in the common set And

  In one embodiment, prior to the step of setting a phrase in the intersection of the first alignment phrase and the second alignment phrase as an evaluation-pending intermediate language E phrase (ie, step S210), the start of the first alignment phrase Each of the portion and the start portion of the second alignment phrase may be determined to be stable. For example, it is determined whether the starting part of the first alignment word and the starting part of the second alignment word each have a meaning range in the source language A that exceeds the first threshold value. Furthermore, the start part of the first alignment word and the start part of the second alignment word may respectively determine whether the meaning range in the target language B exceeds the second threshold value. If the starting part of the first alignment word and the starting part of the second alignment word are not stable, the starting part is removed. For example, if the source language is Chinese, the intermediate language is English, and the target language is Japanese, the starting definite articles, indefinite articles, interns, adverbs, etc. Ranges and meaning ranges in Japanese are often too large and may be removed as unstable phrases.

  Next, in step S212, the first specific attribute for the source language A is determined by the pending evaluation phrase of the intermediate language E.

  In step S214, the evaluation waiting word of the intermediate language E determines the second specific attribute for the target language B.

  In step S216, the reliability score of the evaluation waiting phrase of the intermediate language E is calculated based on the first specific attribute and the second specific attribute.

  In step S218, the evaluation waiting phrase of the intermediate language E is evaluated based on the reliability score.

  FIG. 3 is a flowchart of a machine translation method 300 in one embodiment of the present invention.

  As shown in FIG. 3, in step S302, the reliability score of the phrase P of the intermediate language E can be obtained by the evaluation method 100 or 200 of the phrase P of the intermediate language E described above.

  In step S304, based on the reliability score of the phrase P of the intermediate language E and the machine translation score for translating the specific phrase S of the source language A into the candidate phrase of the target language B via the phrases of the intermediate language E, Calculate the translation score of the target language B candidate phrase.

  For example, based on the reliability score and the machine translation score, the translation score of the candidate word / phrase of the target language B can be calculated by a CKY (Cocke-Kasami-Younger) algorithm.

For example, from among a plurality of phrases in the intermediate language E, a phrase having a reliability score higher than a predetermined reliability score threshold can be selected as the trust phrase. Assume that the number of trust words is N. Preferably, the translation score P _i of the i-th candidate word / phrase of the target language B is calculated by the following equation (4).

Of these, r _j is the reliability score of the jth trust word in the intermediate language E, and T _ij targets the specific phrase S in the source language A via the jth trust word in the intermediate language E. The number of machine translation points translated into the i-th candidate word / phrase of language B.

Among them, the machine translation score T _i ^j (also referred to as T) may be calculated by the following equation (5).

Of these, _Wi is a translation weight, and _Fi is a feature. Preferably, four features, normal translation probability, reverse translation probability, normal lexicalized translation probability, and reverse lexicalized translation probability are used. ln represents the natural logarithm for each of the four features.

  In step S306, based on the translation score, the target language B phrase is selected as a translation result from the target language B candidate phrases. For example, the word / phrase having the highest translation score in the target language may be selected as the translation result.

  Since it is difficult to estimate the reliability score of the intermediate language, in one embodiment, the parameters used in the above method are adjusted using the swarm intelligence algorithm, for example, parameters in the artificial neural network algorithm, parameters in the CKY algorithm , And a predetermined meaning range threshold value are adjusted.

  The swarm intelligence algorithm is similar to the genetic algorithm. In the swarm intelligence algorithm, first, many system configurations are acquired by performing initialization, and each system configuration exists as one individual. The information included in each individual can constitute the entire system. Each individual is evaluated by one evaluation function. The evaluation score of each individual has a direct influence on the sudden change probability and fertility probability of this individual. The fertility rate of individuals with high evaluation values is high, but the rate of sudden change is low. On the other hand, the probability of sudden change is high although the fertility rate of individuals having low evaluation values is low. The entire herd selects and uses the single individual with the highest evaluation value after several generations of breeding and constitutes the entire system.

Particle Swarm Optimization (PSO) is based on swarms and moves individuals in the swarm to a good area based on the fitness to the environment. However, it does not use an evolution operator for the individual and considers each individual as a single volumeless particle (ie, a point) in the D-dimensional search space, which has a constant velocity in the search space. This speed can be adjusted dynamically based on its own flight experience and accompanying flight experience. The i-th particle is expressed as Xi = (xi1, xi2,..., XiD), of which the dimension number D (Size) is the number of all parameters that need to be adjusted. Corresponds to one specific number of dimensions. The best location experienced is noted as Pi = (pi1, pi2,..., PiD), also referred to as pbest. Of these, the best position refers to the position having the best adaptive value, and the calculation of the adaptive value is to obtain a solution by substituting Xi into the target equation, and the obtained solution value is the adaptive value. is there. In the present invention, the target equation is the above-described equation (4), that is, a formula for calculating the translation score. The index number of the best position experienced by all particles in the swarm is denoted by the symbol g, ie, Pg, also referred to as gbest. The velocity of the particle i is represented by Vi = (vi1, vi2,..., ViD). For each generation, its d-th dimension (1 ≦ d ≦ D) changes according to the following equation.

  Where w is inertia weight, c1 and c2 are acceleration constants, and rand () and Rand () vary within the range of two [0, 1] It is a random value.

  The particle speed Vi is limited by one maximum speed Vmax. If the current acceleration to a particle causes the velocity vid in one dimension of the particle to exceed the maximum velocity vmax, d in the dimension, the velocity of the particle in the dimension is reduced to the maximum velocity vmax, d in the dimension. Limited, ie equal to the maximum velocity vmax, d of the dimension.

  In equation (4), the first part is the inertia of the previous action of the particle, the second part is the “cognition” part, representing the particle's own thoughts, and the third part is “social” "Part, representing information sharing and mutual cooperation between particles."

  The “recognition” part: one enhanced random act is likely to reappear in the future. The act here is “recognition” and it is assumed that having gained accurate knowledge is enhanced, and one such model reduces the error as the particles are excited. Assume that

  “Society” part: When an observer observes that a model enhances the act of jealousy, it increases the probability of performing that act. That is, the recognition of the particle itself is imitated by other particles.

  The PSO algorithm uses the following psychological assumptions: that is, in the recognition process for matching, individuals often remember their beliefs and at the same time consider the beliefs of other individuals . Adjustments are made to adaptation when it is determined that the beliefs of other individuals are better.

  The flow chart of the standard PSO algorithm is as follows: 1) Initialize a swarm of particles (the swarm size is m) and include a random position sum velocity; 2) Adaptation of each particle 3) Compare the adaptation value for each particle with the best position pbest experienced, and if it is determined to be good, it is the current best position pbest; 4) For each particle, Comparing the adaptation value with the best overall position gbest experienced, and if good, reposition gbest index number; 5) change particle velocity and position based on equation (1); 6 ) If the termination condition (usually a sufficiently good adaptation value or one predetermined maximum value Gmax is not met), 2) return.

  In addition to swarm intelligence algorithms, other algorithms such as genetic algorithms, artificial immune systems, random propagation searches, and EM algorithms may be used.

  FIG. 4 is a configuration block diagram of an apparatus 400 for evaluating the phrase P of the intermediate language E in one embodiment of the present invention.

  As shown in FIG. 4, the apparatus 400 for evaluating the phrase P of the intermediate language E may include a first specific attribute determination unit 412, a second specific attribute determination unit 414, a reliability score calculation unit 416, and an evaluation unit 418. . The phrase P of the intermediate language E is a bridge for translating the specific phrase S of the source language A into the phrase of the target language B.

  The first specific attribute determination unit 412 determines the first specific attribute for the language of the intermediate language E for the source language A.

  For example, the first specific attribute may include the first semantic range in the source language A for the phrase of the intermediate language E. The first meaning range is the number of phrases corresponding to the phrase of the intermediate language E in the source language A or a function of the number. The function may be a logarithmic function, for example. In some embodiments, the first specific attribute may further include a translation probability of the phrase P of the intermediate language E to the specific phrase S of the source language A and / or the source language A of the phrase P of the intermediate language E. May include the lexicalized translation probabilities for a specific phrase S.

  The second specific attribute determination unit 414 determines a second specific attribute in which the phrase P of the intermediate language E corresponds to the target language B.

  For example, the second specific attribute may include a second semantic range in the target language B of the intermediate language E phrase. For example, the second meaning range is the number of phrases corresponding to the phrase of the intermediate language E in the target language B, or a function of the number. The function may be a logarithmic function, for example.

  The reliability score calculation unit 416 calculates the reliability score of the phrase P of the intermediate language E based on the first specific attribute and the second specific attribute.

  The evaluation unit 418 evaluates the phrase P of the intermediate language E based on the reliability score.

  FIG. 5 is a block diagram showing the configuration of an apparatus 400 'for evaluating the phrase P of the intermediate language E in another embodiment of the present invention.

  As shown in FIG. 5, the apparatus 400 ′ includes a first alignment unit 402, a first removal unit 404, a second alignment unit 406, a second removal unit 408, a common set determination unit 410, and the components shown in FIG. May include a first specific attribute determination unit 412, a second specific attribute determination unit 414, a reliability score calculation unit 416, and an evaluation unit 418.

  The first alignment unit 402 performs alignment between the specific phrase S of the source language A and the phrases in the phrase database of the intermediate language E, and obtains the first phrase of the intermediate language E.

  The first removal unit 404 removes a portion that does not correspond to the specific word / phrase S of the source language A from the first word / phrase to obtain a first alignment word / phrase.

  The second alignment unit 406 aligns the phrases in the target language B phrase database and the phrases in the intermediate language E phrase database to obtain a second intermediate language E phrase.

  The second removal unit 408 obtains a second alignment word / phrase by excluding a part not corresponding to the word / phrase in the word / phrase database of the target language B from the second word / phrase.

  The common set determination unit 410 sets a word in the common set of the first alignment word and the second alignment word as an evaluation waiting word for the intermediate language E.

  The first specific attribute determination unit 412 determines the first specific attribute for the source language A as the evaluation waiting phrase of the intermediate language E.

  The second specific attribute determination unit 414 determines the second specific attribute for the target language B as the evaluation waiting phrase of the intermediate language E.

  The reliability score calculation unit 416 calculates the reliability score of the pending evaluation phrase of the intermediate language E based on the first specific attribute and the second specific attribute.

  The evaluation unit 418 evaluates the evaluation waiting phrase of the intermediate language E based on the reliability score.

  FIG. 6 is a block diagram showing a configuration of an apparatus 400 ″ for evaluating the phrase P of the intermediate language E in another embodiment of the present invention.

  The difference between the apparatus 400 ″ as shown in FIG. 6 and the apparatus 400 ′ as shown in FIG. 5 is that the apparatus 400 ″ further includes a third removal unit 409.

  The third removal unit 409 determines whether the start part of the first alignment word and the start part of the second alignment word are stable, respectively, and the start part of the first alignment word and the start part of the second alignment word If is not stable, remove these starting parts.

  For example, the third removal unit 409 determines whether the start part of the first alignment word and the start part of the second alignment word each have a meaning range in the source language A that exceeds the first threshold value. Remove the part. Further, it is determined whether or not the start part of the first alignment word and the start part of the second alignment word each have a meaning range in the target language B exceeding the second threshold value, and if it exceeds, the start part is removed. Here, if the source language is still Chinese, the intermediate language is English, and the target language is Japanese, the definite article, indefinite article, interposition, adverb, etc. Since the meaning range in Chinese and the meaning range in Japanese are often too large, they may be removed as unstable phrases.

  FIG. 7 is a block diagram showing the configuration of the machine translation apparatus 700 in one embodiment of the present invention.

  As shown in FIG. 7, the machine translation device 700 includes a device 712 that evaluates an intermediate language E phrase, a translation score calculation unit 714, and a translation result selection unit 716.

  The device 712 for evaluating intermediate language phrases may be, for example, a device 400 as shown in FIG. 4, a device 400 'as shown in FIG. 5, or a device 400 ″ as shown in FIG.

  The device 712 for evaluating the phrase of the intermediate language E can obtain the reliability score of the phrase P of the intermediate language E.

  The translation score calculation unit 714 translates the reliability score of the phrase P of the intermediate language E and the machine translation score for translating the specific phrase S of the source language A into the candidate phrase of the target language B via the phrase P of the intermediate language E Based on the above, the translation score of the candidate word / phrase of the target language B can be calculated. For example, the translation score calculation unit 714 can calculate the translation score of the candidate word / phrase of the target language B by the CKY algorithm based on the reliability score and the machine translation score.

  For example, from among a plurality of phrases in the intermediate language E, a phrase having a reliability score larger than a predetermined reliability score threshold may be selected as the trust phrase.

  Preferably, the translation score Pi of the i-th candidate word / phrase of the target language B is calculated by the above formula (4).

  The translation result selection unit 716 may select a target language B phrase as a translation result from the target language B candidate phrases based on the translation score calculated by the translation score calculation unit 714.

  In the embodiment of the present invention, the ambiguity of the intermediate language is analyzed by forming a model, and a phrase having no ambiguity or very little ambiguity in the intermediate language is selected as a bridge. In an embodiment of the present invention, a single phrase P in the intermediate language E can translate a specific phrase S in the source language A into a phrase in the target phrase, and a plurality of phrases P in the intermediate language E The specific phrase S in the source language A can be translated into the phrase in the target phrase.

  Although the principle of the present invention has been described based on the specific embodiments, it should be understood that all or a part of the apparatus according to the embodiments of the present invention or any step may be software, firmware, hardware or the like. It may be realized by any combination method. When implemented by software or firmware, first, the software or firmware program is installed from a storage medium or a network to a machine having a dedicated hardware structure (for example, the general-purpose machine 800 shown in FIG. 8). When the various programs are installed, various functions of the above-described units and subunits can be executed.

  As shown in FIG. 8, the central processing unit (CPU) 801 is a program stored in a read-only memory (ROM) 802 or a program loaded from a storage unit 808 into a random access memory (RAM) 803. Various processes are performed based on this. The RAM 803 may store data necessary when the CPU 801 executes various processes according to needs. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input / output interface 805 is also connected to the bus 804.

  The input / output interface 805 includes an input unit 806 (including a keyboard and a mouse), an output unit 807 (including a display device such as a CRT, an LCD, and a speaker), a storage unit 808 (including a hard disk), and a communication unit. 809 (including a network connection card such as a LAN card and a modem) is connected. The communication unit 809 performs communication processing via a network such as the Internet. A drive 810 may be connected to the input / output interface 805 according to needs. Further, according to needs, a removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is set in the drive 810, so that a computer program read from the medium can be installed in the storage unit 808. Good.

  When the above-described series of processing is realized by software, a program constituting the software may be installed from a network such as the Internet or a storage medium such as a removable medium 811.

  It should be understood by those skilled in the art that such a storage medium is a removable medium as shown in FIG. 8 in which a program is stored and distributed independently of the apparatus so as to provide the program to the user. It is not limited to 811. Examples of the removable medium 811 include a magnetic disk (including a floppy (registered trademark) disk), an optical disk (including a CD-ROM and a DVD), a magneto-optical disk (including MD (registered trademark)), and a semiconductor memory. including. Alternatively, the storage medium may be a hard disk included in the ROM 802 and the storage unit 808, in which a program is stored, and may be distributed to a user together with a device including them.

  The invention also relates to a program product comprising machine-readable instruction codes. When this instruction code is read and executed by a machine, the method according to the above-described embodiment can be executed.

  Furthermore, a storage medium that stores a program product including the above-described machine-readable instruction code is also included in the present disclosure. Such storage media include, but are not limited to, magnetic disks (floppy disks), optical disks, magneto-optical disks, memory cards, memory memory sticks, and the like.

  The method according to the above-described embodiments of the present invention is not limited to being performed according to the time sequence described in the specification or illustrated in the drawings, but may be performed in parallel or independently according to other time sequences. Therefore, the execution order of the method described in this specification or the drawings does not limit the technical scope of the present invention.

  Of course, each processing process of the above-described method of the present invention may be realized by a computer-executable program system stored in various machine-readable storage media.

  Another object of the present invention is to provide a storage medium storing the above-described executable program code directly or indirectly to a system or equipment, and a computer or central processing unit (CPU) in the system or equipment. May be realized by a method of reading and executing the above-described program code.

  In addition, as long as the system or equipment has a function of executing a program, the implementation method of the present invention is not limited to the program, and the program may be in any other format such as an object program or an interpreter execution program. Alternatively, it may be a script program provided to the operating system operation system.

  The machine-readable storage medium described above may be various storage devices and storage units, semiconductor devices, disunits such as optical, magnetic and magneto-optical displays, and any other usable information storage medium.

  In addition, the client computer downloads the computer program code according to the above-described embodiment of the present invention via a predetermined website connected to the Internet, installs it in the computer, and then executes the program, The present invention can also be realized.

  Moreover, the following additional remarks are disclosed regarding the embodiment including each of the above-described examples.

(Appendix 1)
A method for evaluating phrases in an intermediate language,
The intermediate language phrase establishes a first specific attribute for the source language;
The intermediate language phrase establishes a second specific attribute for the target language;
Calculating a confidence score of the intermediate language phrase based on the first specific attribute and the second specific attribute; and
Evaluating the intermediate language phrase based on the confidence score;
Wherein the intermediate language phrase is a bridge for translating a source language specific phrase into a target language phrase.

(Appendix 2)
The method according to appendix 1, wherein
The first specific attribute includes the intermediate language phrase including a first semantic range in the source language, and the second specific attribute includes the intermediate language phrase includes a second semantic range in the target language. Method.

(Appendix 3)
The method according to appendix 2, wherein
The first semantic range is the number of phrases corresponding to the intermediate language phrase in the source language, or a function of the number, and the second semantic range is the target language, The number of phrases corresponding to the intermediate language, or a method of the number.

(Appendix 4)
The method according to appendix 1, wherein
The first specific attribute further includes:
A method comprising: a translation probability of the intermediate language phrase to the source language specific phrase and / or a lexicalized translation probability of the intermediate language phrase to the source language specific phrase.

(Appendix 5)
The method according to appendix 1, wherein
Prior to the step in which the intermediate language phrase establishes a first specific attribute for the source language, the method further comprises:
Aligning a specific phrase of the source language with a phrase in the phrase database of the intermediate language to obtain a first phrase of the intermediate language;
Removing a portion that does not correspond to the specific phrase of the source language from the first phrase to obtain a first alignment phrase;
Aligning phrases in the target language phrase database with phrases in the intermediate language phrase database to obtain a second phrase in the intermediate language;
Removing a portion of the second language that does not correspond to the phrase in the target language phrase database to obtain a second alignment phrase; and
Including a phrase in a common set of the first alignment phrase and the second alignment phrase as an evaluation pending phrase of the intermediate language.

(Appendix 6)
The method according to appendix 5, wherein
Before the step of making the phrase in the common set of the first alignment phrase and the second alignment phrase to be an evaluation pending phrase of the intermediate language,
Determining whether the starting part of the first alignment word and the starting part of the second alignment word are stable, respectively; and
Removing the starting portion if the starting portion of the first alignment phrase and the starting portion of the second alignment phrase are unstable.

(Appendix 7)
The method according to appendix 6, wherein
Each of determining whether the start portion of the first alignment word and the start portion of the second alignment word are stable,
The beginning portion of the first alignment phrase and the beginning portion of the second alignment phrase each determine whether the semantic range in the source language exceeds a first threshold; and / or
A method wherein the starting portion of the first alignment phrase and the starting portion of the second alignment phrase each determine whether a semantic range in the target language exceeds a second threshold.

(Appendix 8)
The method according to appendix 1, wherein
A method characterized in that the first specific attribute and the second specific attribute are characterized, and the reliability score of the intermediate language phrase is calculated by a regression algorithm.

(Appendix 9)
The method according to appendix 8, wherein
The method, wherein the regression algorithm is an artificial neural network algorithm.

そのうち、Kは、活性化関数を表し、g_m(x)は、第m個の特徴の値を表し、w_mは、第m個の特徴の重みを表し、biasWは、バイアス重みを表し、biasVは、バイアス値を表す、方法。 (Appendix 10)
The method according to appendix 9, wherein
Calculate the reliability score f (x) by the following formula,

Where K represents an activation function, g _m (x) represents the value of the m th feature, w _m represents the weight of the m th feature, biasW represents the bias weight, biasV is a method that represents a bias value.

(Appendix 11)
A machine translation method,
Obtaining a reliability score of the intermediate language phrase by the method according to any one of appendices 1 to 10;
The candidate language of the target language based on a reliability score of the intermediate language phrase and a machine translation score for translating a specific word of the source language into the candidate language of the target language via the intermediate language phrase Calculate the translation score of; and
Selecting a target language phrase as a translation result from the candidate phrases of the target language based on the translation score.

(Appendix 12)
The method according to appendix 11, further comprising:
Calculating a translation score of the candidate phrase of the target language by a CKY algorithm based on the reliability score and the machine translation score.

(Appendix 13)
The method according to appendix 11, further comprising:
Selecting from among a plurality of phrases of the intermediate language, a phrase having the reliability score greater than a predetermined reliability score threshold as a confidence phrase.

そのうち、r_jは、前記中間言語の第j個の信頼語句の信頼性点数を表し、T_i ^jは、前記ソース言語の特定語句を前記中間言語の第j個の信頼語句を経由して前記ターゲット言語の第i個の候補語句に翻訳する機械翻訳点数を表す、方法。 (Appendix 14)
The method according to appendix 13, wherein
The translation score P _i of the i-th candidate word / phrase of the target language is calculated by the following formula:

Of these, r _j represents the reliability score of the jth trust word in the intermediate language, and T _i ^j passes the specific phrase in the source language via the jth trust word in the intermediate language. A method for representing a machine translation score to be translated into the i th candidate word in the target language.

(Appendix 15)
A device for evaluating phrases in an intermediate language,
A first specific attribute determining unit for determining a first specific attribute of the intermediate language for the source language;
A second specific attribute determining unit for determining a second specific attribute for the target language by the phrase of the intermediate language;
A reliability score calculation unit for calculating a reliability score of the phrase of the intermediate language based on the first specific attribute and the second specific attribute; and
An evaluation unit that evaluates the phrase of the intermediate language based on the reliability score;
Wherein the intermediate language phrase is a bridge for translating a source language specific phrase into a target language phrase.

(Appendix 16)
The apparatus according to appendix 15, wherein
The first specific attribute includes the intermediate language phrase including a first semantic range in the source language, and the second specific attribute includes the intermediate language phrase includes a second semantic range in the target language. apparatus.

(Appendix 17)
The apparatus according to appendix 16, wherein
The first semantic range is the number of phrases corresponding to the intermediate language phrase in the source language, or a function of the number, and the second semantic range is the target language, A device that is or is a function of the number of phrases corresponding to a phrase in an intermediate language.

(Appendix 18)
The apparatus according to any one of appendices 15 to 17, comprising:
The first specific attribute further includes:
An apparatus comprising: a translation probability of the intermediate language phrase to the source language specific phrase and / or a lexicalized translation probability of the intermediate language phrase to the source language specific phrase.

(Appendix 19)
The apparatus according to any one of appendices 15 to 17, further comprising:
A first alignment unit that performs alignment between the specific phrase of the source language and the phrase in the phrase database of the intermediate language to obtain the first phrase of the intermediate language;
A first removal unit that removes a portion of the first word that does not correspond to the specific word of the source language to obtain a first alignment word;
A second alignment unit that aligns a phrase in the phrase database of the target language and a phrase in the phrase database of the intermediate language to obtain a second phrase of the intermediate language;
A second removing unit that removes a portion of the target language that does not correspond to the phrase in the phrase database of the target language to obtain a second alignment phrase; and
An apparatus comprising: a common set determination unit that sets a word in a common set of the first alignment word and the second alignment word as a word of the intermediate language waiting for evaluation.

(Appendix 20)
A machine translation device,
The apparatus according to any one of appendices 15 to 19, wherein the apparatus obtains a reliability score of the intermediate language phrase;
The candidate language of the target language based on a reliability score of the intermediate language phrase and a machine translation score for translating a specific word of the source language into the candidate language of the target language via the intermediate language phrase A translation score calculation unit for calculating the translation score of;
A machine translation device, comprising: a translation result selection unit that selects a word in the target language as a translation result from the candidate words in the target language based on the translation score.

  The preferred embodiment of the present invention has been described above, but the present invention is not limited to this embodiment, and all modifications to the present invention belong to the technical scope of the present invention unless departing from the spirit of the present invention.

Claims (10)

A method for evaluating phrases in an intermediate language,
The intermediate language phrase establishes a first specific attribute for the source language;
The intermediate language phrase establishes a second specific attribute for the target language;
Calculating a confidence score of the intermediate language phrase based on the first specific attribute and the second specific attribute; and
Evaluating the intermediate language phrase based on the confidence score;
The intermediate language phrase is a bridge for translating a source language specific phrase into a target language phrase. The method of claim 1, comprising
The first specific attribute includes the phrase of the intermediate language including a first semantic range in the source language; and
The second specific attribute is a method, wherein the phrase of the intermediate language includes a second semantic range in the target language. A method according to claim 2, comprising
The first semantic range is the number of words or a function of the number corresponding to the intermediate language words in the source language; and
The method wherein the second semantic range is a number of words or a function of the number corresponding to the words of the intermediate language in the target language. The method of claim 1, comprising
The first specific attribute further includes:
Probability of translation of the intermediate language phrase into the source language specific phrase, and / or
A method comprising lexicalized translation probabilities of the intermediate language phrases to the source language specific phrases. The method of claim 1, comprising
Prior to the step in which the intermediate language phrase establishes a first specific attribute for the source language, the method further comprises:
Aligning a specific phrase of the source language with a phrase in the phrase database of the intermediate language to obtain a first phrase of the intermediate language;
Removing a portion of the first word that does not correspond to a specific word of the source language to obtain a first alignment word;
Aligning phrases in the target language phrase database with phrases in the intermediate language phrase database to obtain a second phrase in the intermediate language;
Removing a portion of the second language that does not correspond to the phrase in the target language phrase database to obtain a second alignment phrase; and
Including a phrase in a common set of the first alignment phrase and the second alignment phrase as an evaluation pending phrase of the intermediate language. A method according to claim 5, comprising
Prior to the step of making a phrase in a common set of the first alignment phrase and the second alignment phrase to be an evaluation pending phrase of the intermediate language, the method further comprises:
Determining whether the starting part of the first alignment word and the starting part of the second alignment word are stable, respectively; and
Removing the starting portion if the starting portion of the first alignment phrase and the starting portion of the second alignment phrase are unstable. The method according to claim 6, wherein
Each of determining whether the start portion of the first alignment word and the start portion of the second alignment word are stable,
The beginning portion of the first alignment phrase and the beginning portion of the second alignment phrase each determine whether the semantic range in the source language exceeds a first threshold; and / or
A method wherein the starting portion of the first alignment phrase and the starting portion of the second alignment phrase each determine whether a semantic range in the target language exceeds a second threshold. A machine translation method,
Obtaining a confidence score of the phrase of the intermediate language by the method according to any one of claims 1-7;
The candidate language of the target language based on a reliability score of the intermediate language phrase and a machine translation score for translating a specific word of the source language into the candidate language of the target language via the intermediate language phrase Calculate the translation score of; and
A machine translation method including selecting a target language phrase as a translation result from the candidate words of the target language based on the translation score. A device for evaluating phrases in an intermediate language,
A first specific attribute determining unit for determining a first specific attribute of the intermediate language for the source language;
A second specific attribute determining unit for determining a second specific attribute for the target language by the phrase of the intermediate language;
A reliability score calculation unit for calculating a reliability score of the phrase of the intermediate language based on the first specific attribute and the second specific attribute; and
An evaluation unit that evaluates the phrase of the intermediate language based on the reliability score;
Wherein the intermediate language phrase is a bridge for translating a source language specific phrase into a target language phrase. A machine translation device,
10. The apparatus of claim 9, wherein the apparatus obtains a reliability score for the intermediate language phrase;
The candidate language of the target language based on a reliability score of the intermediate language phrase and a machine translation score for translating a specific word of the source language into the candidate language of the target language via the intermediate language phrase A translation score calculation unit for calculating the translation score of;
A machine translation device, comprising: a translation result selection unit that selects a word in the target language as a translation result from the candidate words in the target language based on the translation score.

Images