JP3766406B2 - Machine translation device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
  The present invention relates to a machine translation device that translates a first language sentence into a second language sentence.
[0002]
[Prior art]
In recent years, with the globalization of society, interest in machine translation technology that automatically translates a first language sentence (original text) into a second language sentence has increased, and it has been installed in translation services and computers provided on the Internet. Many machine translation software products that perform translation processing have appeared.
[0003]
By the way, the most important point for machine translation is the accuracy of translation. Every year, the accuracy of translation has been improved by improving computer performance, improving translation technology, and accumulating translation knowledge. However, the expression of natural language is extremely diverse, so simply improving translation technology. In other words, it is impossible to achieve satisfactory translation for users only by accumulating translation knowledge.
[0004]
Therefore, in order to overcome the above-mentioned problems in translation accuracy, a pattern translation technique has been proposed in which translation processing is performed under a completely new framework that has a different viewpoint from the conventional framework.
[0005]
One of them is to extract a required sentence from a basic example database that stores bilingual examples and past translation results, and create a perforated example sentence with a part of the sentence as a variable and register it in the variable example database. In addition, a translation method has been realized in which the input first language sentence and the example sentence in the variable example database are collated to perform translation processing (Patent Document 1).
[0006]
The other one is a translation device that enables selection of a more desirable pattern when there are a plurality of translation patterns having the same constant part (a part that is not specified as a variable in a perforated example sentence) ( Patent Document 2). Specifically, this translation device measures information about the structure of the variable part of the first language input sentence by giving information on the phrase structure to the variable part, and calculates the desired translation from the measured value. This is a translation method for selecting a pattern. For example, if the variable equivalent part is a verb phrase, a noun phrase, a simple noun phrase, or a noun phrase with a combined modifier clause, the structure is equivalent to the structure of the input sentence. This is a method of selecting a translation pattern that conditions
[0007]
However, in the machine translation apparatus using the pattern translation technology as described above, as the user proceeds to accumulate the translation patterns, one input sentence may be applied to a plurality of translation patterns depending on how the user gives variables. Inevitable. The reason is that information that can be specified as a variable condition by a general user is limited as described below.
[0008]
<Example of information that can be specified by the user>
This includes grammatical category information of variable parts, phrase structure information, character strings of surface words, semantic information (selected from specific options), and the like.
[0009]
Here, the semantic information can be specified by the user because it belongs to a large category that can be easily classified, such as “people”, “location”, “motion”, and so on. It is a heavy burden to specify correct information, and the user cannot select and specify appropriate information.
[0010]
Next, an example in which a plurality of translation patterns are applied to one input sentence will be described. P1 and P2 indicate translation patterns for English-Japanese translation, and P3 and P4 indicate translation patterns for Japanese-English translation. E is an English pattern and J is a Japanese pattern. In these patterns, $ 1, $ 2, and $ 3 indicate variable portions. Further, here, the variable conditions are omitted for the sake of simplicity of explanation, but all are examples in which noun phrases are used as variables.
[0011]
<Variable range is different, but one input sentence applies to multiple patterns>
(P1) E: $ 1 introduces you to $ 2.
J: $ 1 is an introduction to $ 2.
[0012]
(Original text: This book introduces you to UNIX.)
(P2) E: $ 1 introduces $ 2 to $ 3.
J: $ 1 introduces $ 3 to $ 2.
[0013]
(Original text: Our center introduces foreign students to japanese culture.)
<Example of registering multiple identical patterns with the same constant part>
(P3) J: $ 1 is officially $ 2.
[0014]
E: The official name for $ 1 is $ 2.
(Original text: China is officially called the People's Republic of China.)
(P4) J: $ 1 is officially $ 2.
[0015]
E: $ 1 is technically called $ 2.
(Original text: This substance is formally called a lipophilic alkaloid.) Therefore, when multiple translation patterns are listed as candidates as described above, they are selected according to the priority among these candidates. For example, the adjustment is made according to the following priority standards.
[0016]
(1) Priority is given to long constant strings.
[0017]
(2) If not determined in (1) above, priority is given to those with a small number of constant character strings.
[0018]
(3) If not determined in (2) above, the longest constants are compared and the longer one is given priority.
[0019]
(4) If not determined in (3) above, priority is given to those registered later in time series.
[0020]
[Patent Document 1]
Japanese Patent Laid-Open No. 06-68134
[0021]
[Patent Document 2]
Japanese Patent Laid-Open No. 06-290210
[0022]
[Problems to be solved by the invention]
Therefore, in the translation method and translation apparatus using the pattern translation technology as described above, when a plurality of translation pattern candidates are listed for one input original text, all of them satisfy the application condition. There is a problem that a desired translation pattern cannot be selected.
[0023]
In addition, when a user registers a translation pattern, many similar translation patterns exist if the user does not register while always paying attention to past registration contents, and the burden on the user increases more and more.
[0024]
  The present invention has been made in view of the above circumstances, and enhances the use efficiency of translation patterns, and selects a translation pattern most suitable for the original text to realize highly accurate translation.EquipmentThe purpose is to provide.
[0025]
  Another object of the present invention is to provide machine translation that greatly reduces the burden on the user when translating a first language sentence into a second language sentence using a translation pattern.EquipmentIt is to provide.
[0026]
[Means for Solving the Problems]
(1) In order to solve the above-described problem, the machine translation device according to the present invention includes variable information and a fixed expression that can be changed in addition to knowledge information necessary for translating the first language sentence into the second language sentence. A translation dictionary part in which a plurality of translation patterns each associated with a character string pattern, a condition, and a reference original text that are composed of a constant part that is a format, and an input that is the first language sentence input from the input part Control means for judging an original sentence and a translation processing instruction and outputting a translation processing start instruction, and translation processing of each word constituting the input original sentence based on the knowledge information based on the translation processing instruction outputted from the control means Means for acquiring various information necessary for the above, means for comparing and collating the input original text with the character strings of the respective translation patterns, extracting translation pattern candidates, and each variable portion of each of the extracted translation pattern candidates Above If there is a variable condition check means for checking whether the condition of the corresponding translation pattern is met based on the seed information and a plurality of translation patterns that match the condition by the variable condition check means, the variable equivalent part of the input source text And a similarity calculation means for selecting the translation pattern having the highest similarity, and applying the selected translation pattern to the second language. It is the structure which translates into a sentence.
[0027]
With the above configuration, the present invention searches for translation pattern candidates by comparing and referring to the character string of the input original text and the character strings of the source text side constant parts of the plurality of translation patterns in the pattern dictionary section. If there are a plurality of translation pattern candidates as a result of this search, it is determined whether or not the variable portion of each translation pattern candidate meets the variable conditions in the pattern dictionary portion. When it is determined that it matches, use the reference source text of each translation pattern in the pattern dictionary section for the input source text, calculate the similarity between translation pattern application candidates, and translate the input source text using the translation pattern with the highest similarity Therefore, even when a plurality of translation pattern candidates exist, a more appropriate translation pattern can be selected. Further, when selecting an optimal translation pattern, the user does not need to select the entire past registered pattern in the conventional manner, and the burden on the user can be greatly reduced. Note that the translation pattern candidate search process and the variable condition matching process described above can be processed in parallel.
[0028]
When an instruction to confirm the translation result input from the input unit is detected with respect to the translation result obtained by applying the translation pattern, or when the cursor moves from the translation result sentence to the later translation target sentence A confirmation detecting means for detecting that the translation has been confirmed, and a pattern registration means for storing the input original text in association with the reference original text of the applied translation pattern when it is determined that the translation is confirmed by the confirmation detecting means. If it is added, the reference source text of each translation pattern is accumulated each time the input source text is repeatedly translated, so that the most appropriate translation pattern can be searched for the translation of the input source text after the next, enabling more accurate translation processing. Become.
[0031]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[0032]
FIG. 1 is a block diagram showing a first embodiment of a machine translation apparatus according to the present invention.
[0033]
This machine translation apparatus is inputted from an input unit 1 as an input means for inputting a first language sentence to be translated (hereinafter referred to as an input original sentence and a translation target sentence) and various commands, and the input unit 1. A processing control unit 2 that judges an input original sentence and various commands and outputs a command necessary for each component, and a translation dictionary that stores various rules and a dictionary for translating the first language sentence into the second language sentence A translation processing unit 4 that translates a first language sentence into a second language sentence using various rules and dictionaries stored in the translation dictionary unit 3 when a translation command is received from the unit 3 and the processing control unit 2 , And the output unit 5.
[0034]
When the processing control unit 2 and the translation processing unit 4 are configured by a CPU, the processing control unit 2 records a program for centrally controlling each component, and the translation processing unit 4 performs translation processing. It is also possible to execute a series of processes by using the recording medium 7 for recording the translation processing program that defines the processing procedure relating to the above.
[0035]
The input unit 1 generally uses a keyboard, a mouse, and the like. In addition to inputting various commands as described above, the input unit 1 is input by inputting an original text created by an input operation of the keyboard, or by specifying a specific area in a document using a mouse. In addition to inputting input text to be selected, input of read data by OCR (Optical character Rerder) that reads a printed or handwritten document, input of input text stored on a floppy disk, magnetic tape, magnetic disk, etc., and the Internet There are input source text input from above, or input source text obtained by converting a conversation text input from a microphone into a natural language character string using dictation software (Speech Dictation Software). That is, the input unit 1 includes not only general input original text input but also input original text input in various different input forms.
[0036]
The process control unit 2 is provided with an instruction determination unit 2a that determines the instruction contents of various commands input from the input unit 1, and a process control unit 2b that outputs a required process command according to a determination result by the instruction determination unit 2a. ing.
[0037]
The translation dictionary unit 3 includes a vocabulary unit 3a, a morpheme analysis rule unit 3b, a syntax / semantic analysis rule unit 3c, a conversion rule unit 3d having rules related to language conversion, a syntax generation rule unit 3e, a morpheme generation rule unit 3f, The pattern dictionary unit 3g, the concept dictionary unit 3h, and other rules and dictionaries necessary for translation processing are stored. In addition, it can be said that the other rules 3a to 3f other than the pattern dictionary unit 3g and the concept dictionary unit 3h are knowledge information used for translating the input original sentence into the second language sentence.
[0038]
The translation processing unit 4 uses the vocabulary unit 3a and the morphological analysis rule unit 3b in the translation dictionary unit 4 for the input original sentence, and includes various parts of speech, utilization, semantic information, and other attributes of the words constituting the input original sentence Pattern matching extraction means for sequentially matching the original pattern of each translation pattern in the pattern dictionary unit 3g and extracting a candidate of the translation pattern that matches the constant character string in the original pattern 4a, variable condition check means 4b for checking whether a variable portion existing in the translation pattern candidate extracted by the pattern matching extraction means 4a meets a predetermined translation pattern condition, an input original sentence and each translation pattern candidate The similarity calculation means 4c for calculating the similarity by comparing with the reference original text of the translation, and the translation parameter having the highest similarity calculated by the similarity calculation means 4c. Select over emissions candidate, and target generating means 4d for performing translation processing using knowledge information translation dictionary unit 3 is provided.
[0039]
The output unit 5 has a function of outputting a translation result that is an output of the translation processing unit 4 or displaying a response from the processing control unit 2 to various instructions input from the input unit 1. Usually, display means such as various displays are used. In addition, for example, printing means such as a printer, writing registration means to a floppy disk, magnetic tape, magnetic disk, transmission means for transmitting to other media, etc. There are various output forms desired by the user.
[0040]
FIG. 2 is a diagram showing an example of translation patterns stored in the pattern dictionary unit 3g used in the present embodiment.
[0041]
Each translation pattern of the pattern dictionary unit 3g stores five types of information as one unit as described below. That is, each translation pattern is
1. Source side pattern
2. Translation side pattern
3. Source-side conditions
4). Translation condition
5). Reference text
It is constituted by.
[0042]
The $ 1, $ 2,... Inserted in the source sentence pattern and the translated sentence pattern indicate variable parts, and the other character strings indicate constant parts. By changing the numbers “1” and “2” next to “$” for each variable, the variable portion of the original pattern corresponds to which variable portion in the translated pattern.
[0043]
On the other hand, the source-side condition is a part that specifies a condition to be satisfied by the character string corresponding to the variable when collating with the input original text. In the example shown in FIG. Specifies what the is. In FIG. 2, NP is a noun phrase and VP is a verb phrase. In addition, for example, a surface character string of a word in the original sentence, a part of speech, semantic information, and the like can be designated. The translation-side condition is a part that specifies under what conditions the phrase of the variable part is to be output when the translation is generated. In the example shown in FIG. 2, the verb phrase of the second variable is a clause in the translation. It is specified to be generated in the form of In addition, for example, number information, tense, and the number of responses to other phrases can be specified. Note that what is specified in the condition is not limited in the present invention, and can be freely set within a range not departing from the gist.
[0044]
Further, the reference original text is a portion that stores an input original text registered when the translation pattern is registered.
[0045]
Next, in the machine translation apparatus as described above, an example of interactive processing with the user in the processing control unit 2 will be described with reference to FIG. Note that interactive processing with the user can be executed by a hardware logic circuit configuration, but here, an example in which a series of processing is executed according to a program stored in the recording medium 6 will be described.
[0046]
When the operation of the apparatus starts, the processing control unit 2 determines whether or not an input original is input from the input unit 1 according to a program stored in the recording medium 6 (S11), and it is determined that the input original is input. In such a case, after temporarily storing in an appropriate memory not shown, a display command is issued to the output unit 5 to display the input original (S12).
[0047]
Here, after displaying the input original text or when it is determined in step S11 that the input original text has not been input, the processing control section 2 determines whether or not a translation instruction command has been input from the input section 1 ( S13) When a translation instruction is input, the input original text temporarily stored in the memory is sent to the translation processing unit 4 and a translation processing start command is sent. Based on the translation processing start command from the processing control unit 2, the translation processing unit 4 executes a translation process for translating the input original sentence into the second language sentence using the rules and dictionary of the translation dictionary unit 3 (S14). The translation result is displayed on the output unit 5 (S15). A series of processing examples of translation processing by the translation processing unit 4 will be described later.
[0048]
If it is determined in step S13 that a translation instruction command has not been input, it is subsequently determined whether or not a command for starting pattern registration has been input from the input unit 1 (S16), and a pattern registration start command has been input. If it is, the pattern registration unit 8 to be described later is instructed to start the pattern registration process. The pattern registration unit 8 executes pattern registration processing in accordance with a user instruction (S17).
[0049]
If an instruction to end the entire process is input from the input unit 1, all the processes are ended (S18). Further, when an instruction command other than the above instruction commands is input from the input unit 1, the process control unit 2 determines the content of the instruction command (S19), and performs processing according to the instruction command (S20). ).
[0050]
Accordingly, the process control unit 2 functionally includes steps S11, S13, S16, S18, S19, and the like that determine various instructions input from the input unit 1 among the above processes. The other step processing corresponds to the processing control means 2b.
[0051]
The determination of the end instruction command is performed before the determination of other instruction instructions, but it may be performed after the determination of other instruction instructions, and the other processing order is not particularly limited.
[0052]
  Next, of the series of processes as described above, the process control unit 2 determines that the instruction is a translation process instruction (S13), and the translation process (S14) of the translation process unit 4 when a translation process start command is sent. This will be described with reference to FIG. FIG. 4 is a diagram showing a series of processing examples from when an input source sentence (first language sentence) is inputted from the input unit 1 to translation into a second language sentence and output.The
[0053]
When the translation processing unit 4 receives a translation processing start command together with the input original text from the processing control unit 2, the translation processing unit 4 executes the pattern matching extraction unit 4a according to the translation processing program shown in FIG. This pattern matching extraction means 4a uses the vocabulary part 3a and the morpheme analysis rule part 3b in the translation dictionary part 4 for the input original sentence, performs morpheme analysis / dictionary processing, and the part of speech, utilization, meaning of the input original sentence Various information necessary for information and other translation processing is acquired (S21). Therefore, in this process, candidates for all parts of speech, utilization, semantic information, etc. of the words constituting the input original sentence can be listed.
[0054]
After listing the part of speech of the word constituting the input original text and candidates for utilization, the pattern dictionary unit 3g is used to sequentially match the original text side pattern of each translation pattern against the input original text. All of the translation pattern candidates that match the constant character string in the pattern are extracted, and the numbers [for example, N0.1, N0.3,...] Of these extracted translation patterns are extracted as pattern candidate storage registers Pcand [n] (not shown). ) Are sequentially arranged and recorded (S22). If no translation pattern candidate exists as a result of collating the original sentence side pattern of each translation pattern against the input original sentence, the pattern candidate storage register Pcand [0] is empty. At this stage, the suitability evaluation of the condition of the variable part is not performed. Note that these steps S21 and S22 correspond to a pattern matching extraction function and a pattern matching step. Thereafter, 0 is set to the array counter n indicating the array order of the extracted translation patterns (S23).
[0055]
Thereafter, the process proceeds to step S24, and it is determined whether or not there is an unanalyzed pattern candidate from the stored contents of the pattern candidate storage register Pcand [n] for storing the translation pattern candidate. Here, when the counter n = 0 and the corresponding area of the storage register Pcand [n] is empty, that is, when there is no translation pattern candidate, the translation pattern is determined using the syntax / semantic analysis rule unit 3c. A normal syntax / semantic analysis process that is not used is executed (S25), and a general translation process in steps S36 to S38 to be described later is executed. These steps S25 and S36 to S38 correspond to a translation generation processing means, a translation generation processing function, and a translation generation processing step.
[0056]
The translation processing unit 4 executes the variable condition checking unit 4b when n is not 0 and the register Pcand [n] is not empty in step S24. That is, if the register Pcand [n] is not empty, the variable condition checking unit 4b determines that there is an unanalyzed translation pattern candidate, and the variables shown in FIG. 2 are as many as the number of variable parts existing in the translation pattern candidate. Since it is necessary to check the compatibility of conditions and the like, first, 1 is set to the variable number m (S26), and it is determined whether or not variable analysis has been performed sequentially from the first variable portion (S27). That is, in step S27, it is determined whether or not the compatibility check has been analyzed for the m-th variable. If it has not been analyzed, the syntax / semantic analysis rule part 3c is related to the input original text string corresponding to the variable. Is used to execute syntax / semantic analysis processing (S28), and acquire the structure, semantic information and other various grammatical information of the variable equivalent portion of the input original sentence. Then, the obtained various information described above is collated with the description content of the original text condition of the variable, and it is determined whether or not the condition is met (S29). If the original condition of the variable is met, the variable number m is incremented by +1 (S30), and the process returns to step S27 to repeat the same process for the next variable part.
[0057]
By the way, if it is determined in step S27 described above that there is no unanalyzed variable for a specific translation pattern, that is, if conditions such as conditions have already been checked for all variable parts, the constant part and the variable part The structures are connected to form one structure sentence (S31).
[0058]
FIG. 5 is a structural connection diagram in which the structures of the constant portion and the variable portion are connected. In this figure, the structure of the variable part and the constant part is connected in the order of occurrence in the input original text, “(pattern attribute V1)” is equivalent to the first variable, and “(pattern attribute V2)” is the second. “(Pattern attribute C1)” indicates the first constant element, and “(pattern attribute C2)” indicates the second constant element.
[0059]
After creating a linked structure in which the structures of the variable part and the constant part are linked as described above, +1 is incremented to the counter n representing the sequence order of translation pattern candidates (S32), and the process returns to step S24 to perform the same processing. Run repeatedly.
[0060]
Also, in step S29, when the obtained various information is compared with the description content of the original condition of the variable, and it is determined that the variable condition is not met, the translation pattern is incompatible. Then, the process proceeds to the next translation pattern candidate process (after S24) through step S32 as described above.
[0061]
The series of processing steps S24 and S26 to S32 correspond to a variable condition check function or a variable condition check step.
[0062]
As described above, the translation processing unit 4 completes the variable condition check for all the translation pattern candidates, the structural connection between the variable part and the constant part, etc., and determines in step S24 that no unanalyzed pattern remains. If so, the similarity calculation means 4c is executed.
[0063]
In this similarity calculation means 4c, when it is determined in step S24 that there is no unanalyzed variable portion, the process proceeds to the next step S33, and for one input original, there is a translation pattern candidate suitable for the variable condition check. Judge how many are left. As a result of this determination, there are cases where only one translation pattern remains and cases where a plurality of translation pattern candidates still remain. In the former case, the process proceeds to step S36, and the structure of the first language, which is the input source sentence, is converted to the structure of the second language based on the corresponding translation pattern. If there are a plurality of translation pattern candidates, The similarity between the input original text and the variable equivalent part of the reference original text (see FIG. 2) of each translation pattern candidate is calculated (S34: similarity calculation function, similarity calculation step). The similarity calculation process will be described later with reference to FIG.
[0064]
When the similarity is calculated as described above, the translation pattern candidate having the highest similarity is selected from the calculated similarities (S35), and the conversion process is started (S36). In this conversion process, as described above, the pattern dictionary unit 3g and the conversion rule unit 3d are used to replace the structure of the input original sentence with the structure of the second language. Specifically, if the input structure to the conversion process is a translation pattern application result, the translation side pattern information of the corresponding translation pattern in the pattern dictionary unit 3g is used, the constant part is replaced according to the translation side pattern, The process of restructuring the structure with reference to the order in the translation of the component is performed. FIG. 6 is a diagram showing an example of the replacement processing result. Therefore, at this stage, the structure of the variable part remains the analysis result of the first language. Subsequently, the partial structure of the variable portion is replaced with the structure of the second language by using the conversion rule unit 3d of the translation processing unit 3. FIG. 7 is a diagram illustrating an example of the replacement result.
[0065]
Thereafter, a process for generating a one-dimensional word string in the second language is performed on the conversion result using the syntax generation rule unit 3e (S37). Further, using the morpheme generation rule unit 3f, final word utilization is performed, and a translation of the second language for the input original is generated and output (S38). These steps SS35 to S38 correspond to a translation generation function or a translation generation step.
[0066]
Next, an example of similarity calculation processing between translation pattern candidates will be described with reference to FIG.
First, a counter n = 1 is set to set the first candidate among the translation pattern candidates as the target pattern (S41). This n is a variable representing the number of the target pattern in the candidate. After n = 1 is set in step S41, it is determined whether or not there is an unprocessed translation pattern candidate (S42). If it is determined that there is an unprocessed translation pattern candidate, the similarity of the translation pattern is determined. Is initialized to 0, and 1 is set to the counter m (not shown) of the variable of interest (S43). This means that the first variable is processed.
[0067]
Subsequently, it is determined whether or not an unprocessed variable remains (S44). If there is an unprocessed variable, a character string in the input original corresponding to the variable of interest and a reference to the translation pattern are referred to. The character string corresponding to the variable in the original text is compared and the similarity is calculated (S45). A specific method for calculating the similarity will be described later. As a result of the similarity calculation, the smaller the numerical value, the higher the similarity. The calculated numerical value of the similarity is added to the register D [n] for storing the similarity (S46). When the similarity with respect to one variable is calculated, +1 is incremented to the target variable counter m (S47), and the process returns to step S44 to perform the similarity calculation process with respect to the next variable. Then, when all the variables have been processed, it is determined that there is no unprocessed variable, that is, that the processing of the translation pattern candidate has been completed, the counter n is incremented by +1 (S48), the process proceeds to step S42, and the next unprocessed process is performed. Determine whether there is a translation pattern. If all the translation pattern candidate processes have been completed, it is determined that there is no unprocessed pattern, and the process proceeds to step S49. Here, the similarity of each translation pattern candidate stored in the register D [n] is compared with a translation pattern having the smallest numerical value as a translation pattern having a high similarity, and the process is passed to step S35 to perform all processes related to similarity calculation. finish.
[0068]
* Specific calculation method of similarity.
[0069]
In the similarity calculation process, specifically, the similarity is calculated using the concept dictionary unit 3 h of the translation dictionary unit 3. In the concept dictionary of the concept dictionary unit 3h, the relationships between individual concepts are constructed in a tree shape, for example, and the concepts are connected by a relationship name called an arc. Each concept is composed of the following elements.
[0070]
* Concept name Example: Country
* Concept ID Example: 001f229b
* Synonym set Example: {Japan, China, People's Republic of China, Korea, USA, USA, USA, ...}
For example, the concept name “country name” is assigned 001f229b as a concept ID, and its elements are “Japan”, “China”, “People's Republic of China”, “Korea”, “America”, “United States”, etc. Specific phrases are defined. As the concept ID, instead of defining the surface word itself, a set of headword IDs assigned to each registered word in the vocabulary part 3a may be defined. A concept ID may be added to each headword in the vocabulary part 3a so that the concept dictionary can be easily searched from individual words.
[0071]
Next, the main items of relation names that connect concepts are as follows:
* Upper / lower relationship (IS-A)
* There is an antitrust (ANT).
[0072]
Further, in the above example, noun items are given for the purpose of simplifying the explanation of the concept. However, it is also effective to associate an adjective concept, a verbal concept, etc. with the concept dictionary unit 3h. is there.
[0073]
FIG. 9 is a diagram schematically showing a part of the concept dictionary used in the present embodiment. In the figure, each circle represents an individual concept, and a line connecting these concepts is called an arc. Here, for convenience, only the concept names are listed, and some of the concepts are omitted. Further, in this schematic diagram, for the sake of convenience, an arc representing “upper / lower relationship” is taken as an example.
[0074]
In this embodiment, basically, a distance between two specific concepts in a tree structure is converted into a coefficient by associating it with an arc name and converted into a numerical value of similarity. Specifically, when tracing from a specific concept to another specific concept on the dendritic structure, “1” every time “IS-A” is traced, and “2” every time “ANT” is traced. A simple similarity calculation method of adding to the storage register D [n] is employed. Although not shown, it is also effective to use “ANT” to change the coefficient depending on the part of speech or to set it in detail using other arcs not mentioned here. Furthermore, since “IS-A” has a directionality, it is possible to make a coefficient while taking the directionality into consideration.
[0075]
In short, the configuration of the concept dictionary and the quantification based thereon are not limited in the present invention, and can be freely determined without departing from the spirit of the present invention. For example, a network structure such as a neural network or WordNet may be used instead of the tree structure as shown in FIG.
[0076]
* Specific examples of similarity calculation.
[0077]
(1) Now, a specific example will be described in which similarities are compared when “Korea is officially called the Republic of Korea” is input as the original Japanese-English translation.
[0078]
First, in a series of translation processes shown in FIG. 4, in step S22, a translation pattern candidate is searched, and a translation pattern No. in FIG. 2, No. Assume that two candidates 3 are extracted. These two patterns have the same constant range, and the character strings of the variable equivalent part in the input original text and the variable equivalent part of both patterns are the same. The portion corresponding to the first variable is “Korea”, and the portion corresponding to the second variable is “South Korea”.
[0079]
Therefore, the variable equivalent portion is subjected to variable analysis in step S28. These two variable equivalent parts are both noun phrases using the registered contents of the vocabulary part 3a, and the semantic information is analyzed as “organization, place”. Since these variable-corresponding portions all satisfy the original text condition of the corresponding pattern, similarity comparison processing is performed in step S34 after variable analysis. Comparison of similarities is performed for each variable equivalent for N0.2 and N0.3, and the similarity of each translation pattern is high by adding the similarities for each pattern. Determine whether.
[0080]
First, “Korea” corresponding to the first variable and the translation pattern No. The degree of similarity with “China”, which is a portion corresponding to the first variable in the reference original text of 2, is calculated.
[0081]
Therefore, the respective positions in the concept dictionary unit 3h of “Korea” and “China” are detected using the concept ID added to each headword in the vocabulary unit 3a, and the distance is quantified. As a result of this detection, it can be seen that these two variable-corresponding parts are both defined in the synonym set having the same ID, so 0 is added to the similarity storage register D [1]. That is, it can be said that “Korea” and “China” have the same position defined in the concept dictionary unit 3h.
[0082]
(2) Next, “Korea”, which is also equivalent to the first variable, and the translation pattern No. The degree of similarity with “this substance”, which is the portion corresponding to the first variable in the reference text of 3, is calculated. When this “substance” is analyzed using the vocabulary part 3a, the morphological analysis rule part 3b, and the syntax / semantic analysis rule part 4c, it is analyzed that the head of the phrase is “substance”. The similarity is calculated. The concept ID assigned to the “substance” is not shown in detail in FIG. 9, but matches the concept ID one level below the “specific object”. Therefore, the distance to the “country” where “Korea” is located is determined to be 7, and this 7 is added to the similarity storage register D [n].
[0083]
Subsequently, the similarity calculation processing for the portion corresponding to the second variable is started. Similarly, “Republic of Korea” and “People's Republic of China” are judged to have a similarity of 0, and this similarity of 0 is added to the similarity storage register D [1], and finally the similarity storage register D [1]. Becomes 0.
On the other hand, the degree of similarity between the “album” of the head of “lipophilic alkaloid” and “Korean Republic” corresponding to the second variable is calculated. This “alkaloid” is not shown in FIG. 9 as well, but belongs to the concept one level below that is connected to the concept to which the “substance” belongs in an IS-A relationship. Therefore, the similarity is determined to be 8, 8 is added to the similarity storage register D [2], and the final similarity storage register D [2] is 15.
[0084]
Therefore, with these similarity calculations, the translation pattern No. 1 which is the first candidate with a low similarity value is obtained. 2 is selected.
[0085]
Next, an example will be described in which an input original “The loss of electrical energy at this time is formally referred to as resistance loss” is input, and similarly, the input original is compared by a Japanese-English translation process.
[0086]
Similarly, for this input original, “loss of electrical energy at this time” in the portion corresponding to the first variable and “resistance loss” in the portion corresponding to the second variable are analyzed, and each translation pattern No. 2, No. Since the original text side condition of 3 is satisfied, both patterns remain as candidates. The head word corresponding to the first variable is analyzed as “loss”. Based on the concept ID assigned to this “loss”, “loss” is “China”, “substance” and the concept dictionary tree, respectively. The distance on the structure is quantified. “Loss” belongs to “action”, the distance to “China” is determined to be 7, and the distance to “substance” is determined to be 4.
[0087]
Next, the head word of the portion corresponding to the second variable is analyzed as “loss”, and since this is also given the concept ID of “action”, the distance from “China” is 8, “alkaloid” Is determined to be 5. Eventually, D [1] is 15 and D [2] is 9. Therefore, the translation pattern N0.3, which is the second candidate with a low similarity value, is selected.
[0088]
In the above example, the similarity comparison is performed using only the concept dictionary unit 3h. However, as a preprocessing, for example, a configuration may be used in which candidates are cut off using a result of syntax / semantic analysis. . In addition, the similarity of the phrase structure of the variable part, which is the result of parsing, may be scaled to arrange numerical values.
[0089]
(3) Next, the similarity comparison when “This historical museum introduces you to the history of the Renaissance” is input as an English-Japanese translation original will be described.
[0090]
Similarly to the above, the input original text is searched for the translation pattern, and the translation pattern No. 1 in FIG. 4, No. Two candidates of 5 are extracted. These two patterns are different from the example described above, and have different variable ranges. However, in this embodiment, even when the variable ranges are different, the similarity of each variable-corresponding portion is calculated, and finally the similarity is calculated by a method of addition input. In the calculated similarity, the translation pattern having a smaller numerical value has a higher similarity, so that, for example, “you” in the input original is translated pattern No. 4 matches the constant, and the translation pattern No. 5 matches the variable. Pattern No. As long as it matches the constant, 4 is treated as substantially the same as when the similarity is determined to be 0 when it is recognized as a variable because there is no addition.
[0091]
First, “This historical museum” which is a portion corresponding to the first variable of the input original and the pattern No. The degree of similarity with “This book” in the portion corresponding to the first variable of 4 is calculated. Here, as a result of the syntax / semantic analysis, the similarity between “museum” and “book” is calculated. “Museum” has the concept ID of “space organization” in FIG. 9, and “book” is not shown in FIG. 9, but the concept ID of the concept located two steps below “concrete” have. Therefore, the similarity is determined to be 8, and 8 is added to the register D [1].
[0092]
Next, “the history of the Renaissance” corresponding to the second variable of the input original text and the pattern No. The similarity of the reference original text “4” with “UNIX” is calculated. Similarly, the similarity between “history” and “UNIX” is calculated. Although “history” is not shown in FIG. 9, it has a concept ID two steps below “abstract”, and “UNIX” is a concept concept three steps below that is connected from another arc from “abstract”. Has an ID. Therefore, the similarity is determined to be 5, and 5 is added to the register D [1], so that the register D [1] finally becomes 13.
[0093]
Next, the translation pattern No. The similarity comparison processing with 5 is started. Similarly, the head “museum” of “This historical museum”, which is the portion corresponding to the first variable of the input original, and the translation pattern No. “Center” in reference text 5 is located in “spaced organization” in FIG. Therefore, at this stage, the similarity is determined to be 0, and 0 is set in the register D [1].
[0094]
Subsequently, “you”, which is the portion corresponding to the second variable of the input original text, and the translation pattern No. The head “students” of “foreign students” in the reference text of 5 is located in “human” in FIG. 9. Therefore, at this stage, the similarity is determined to be 0, and the register D [2] remains 0.
[0095]
Furthermore, “the history of the Renaissance” corresponding to the last third variable of the input original text and the translation pattern No. 5 starts calculating the degree of similarity between the heads “history” and “culture” of “Japanese culture”, which is the portion corresponding to the third variable. Although this “culture” is not shown in FIG. 9, it has a concept ID of a concept that is connected by another IS-A arc from the one-level concept of “history”. Therefore, the distance is 2, and finally, 2 is set in the register D [2]. Therefore, the second translation pattern No. 2 having a small similarity value. 5 is selected.
[0096]
Therefore, according to the embodiment as described above, the character string of the constant part of the input original text and the character string of the constant part in the source text side patterns of the plurality of translation patterns of the pattern dictionary unit 3g are compared and referenced, and the translation pattern Search for candidates. As a result, when there are a plurality of translation pattern candidates, it is determined whether or not the variable part of each translation pattern candidate meets the variable condition of the pattern dictionary unit 3g. Then, when it is determined that they match, the similarity between the translation pattern application candidates is calculated using the reference original text of each translation pattern of the pattern dictionary unit 3g for the input original text, and the input using the translation pattern with the highest similarity. Since the original text is translated, a more appropriate translation pattern can be selected even when a plurality of translation pattern candidates exist.
[0097]
In addition, when selecting an optimal translation pattern, the user does not need to select the entire registered pattern while conscious of the entire past registered pattern, and the burden on the user can be greatly reduced.
[0098]
FIG. 10 is a block diagram showing a second embodiment of the machine translation apparatus according to the present invention. In this figure, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description thereof will be given in the description of FIG.
[0099]
In this embodiment, after a specific translation pattern is selected, a translation is generated using the translation pattern, a cursor is set on the translation generation sentence, and if the user performs a confirming operation from the input unit 1, This is an example in which the confirmation operation is detected and the input original sentence to which the translation pattern is applied is automatically stored in the reference original sentence of the translation pattern.
[0100]
Specifically, the process control unit 2 is provided with a confirmation detection means 2c or a confirmation detection function and a confirmation detection step in addition to the instruction determination means 2a and the process control means 2b. In this machine translation apparatus, a translated sentence is generated using a translation pattern, and a cursor is set in the generated sentence. At this time, when a translated sentence confirmation instruction is input from the input unit 1, the confirmation detecting unit 2c, for example, In step S16 or S19 shown in FIG. 3, it is determined that there is a translation finalization instruction, and it is detected that the translation result is a translation lock to avoid rewriting the translation result in a later editing, either directly or by processing. A translation edit lock processing command is sent to the pattern registration unit 8 through the control means 2b (S20).
[0101]
Here, when receiving the translation editing lock processing command, the pattern registration unit 8 adds the corresponding input original text to the reference original text of the pattern dictionary section 3g if the translation text on which the cursor is already placed is a translation pattern application text. . In addition, when there are a plurality of translation pattern candidates as described above, the reference original text to be added is calculated for each of the reference original text of each translation pattern, and the input original text is added to the reference original text of the translation pattern having a higher similarity. Is stored.
[0102]
Further, the timing of storing the input original sentence as the reference original sentence is the time when it is determined as the translation edit lock processing instruction in step S19 shown in FIG. 3, but the corresponding input original sentence is determined by various timings and operations from the input unit 1. Can be stored in the reference original text of the pattern dictionary section 3g.
[0103]
Therefore, according to the embodiment as described above, since the reference original text of each translation pattern is accumulated every time the translation of the input original text is repeated, the optimum translation pattern can be searched for the translation of the input original text after the next, More accurate translation processing can be realized.
[0104]
The pattern registration unit 8 is provided separately from the processing control unit 2 and the translation processing unit 4, but may be provided in either the processing control unit 2 or the translation processing unit 4.
[0105]
FIG. 10 is a block diagram showing a third embodiment of the machine translation apparatus according to the present invention. In this figure, the same parts as those in FIG. 1 are denoted by the same reference numerals, and detailed description thereof will be given in the description of FIG.
[0106]
In this embodiment, as in the second embodiment described above, an input original sentence is automatically acquired for a translated sentence sentence to which a translation pattern is applied, and stored in a reference original sentence of the corresponding translation pattern. However, in the second embodiment, the input original text is acquired when the user explicitly instructs the translation to be confirmed, whereas in the third embodiment, without an explicit instruction, This is an example in which it is determined whether or not an input original should be acquired from a user operation, and the input original is stored as a reference original.
[0107]
Specifically, when the user inputs a movement command to move the cursor during the translation process, the instruction determination unit 2a determines that the instruction is a cursor movement instruction in “Other instruction command” in step S19 shown in FIG. Then, it notifies the fixed detection means 2c. Upon receipt of the determination notification of the cursor movement instruction, the confirmation detection unit 2c determines whether the translated sentence generation sentence immediately before the downward movement of the cursor is a translation pattern application sentence. “Downward movement of cursor = determination of translation pattern application sentence is determined, and an input original sentence acquisition command is sent to the pattern registration unit 8 via the processing control means 2b.
[0108]
Here, the pattern registration unit 8 additionally stores the original text immediately before the cursor is moved downward in the reference original text area of the translation pattern applied to the text.
[0109]
Therefore, according to such an embodiment, variations of the input source text that may be automatically applied to individual translation patterns without user explicit action are stored and used for subsequent translation processing. Thus, each time the translation is repeated, the appropriate selection accuracy of the translation pattern is increased, and as a result, a highly accurate translation process can be realized.
[0110]
The pattern registration unit 8 is provided separately from the processing control unit 2 and the translation processing unit 4, but may be provided in either the processing control unit 2 or the translation processing unit 4.
[0111]
The invention of the present application is not limited to the above-described embodiment, and various modifications can be made without departing from the scope of the invention.
[0112]
In addition, the embodiments can be implemented in combination as much as possible, and in that case, the effect of the combination can be obtained. Further, each of the above embodiments includes various higher-level and lower-level inventions, and various inventions can be extracted by appropriately combining a plurality of disclosed constituent elements. For example, when an invention is extracted because some constituent elements can be omitted from all the constituent elements described in the means for solving the problem, the omitted part is used when the extracted invention is implemented. Is appropriately supplemented by well-known conventional techniques.
[0113]
【The invention's effect】
  As described above, according to the present invention, the use of a reference original text can increase the efficiency of use of a translation pattern, and the machine translation that can realize a highly accurate translation by selecting a faithful translation pattern based on the original text.EquipmentCan be provided.
  The present invention also provides machine translation that can greatly reduce the burden on the user when translating a first language sentence into a second language sentence using a translation pattern.EquipmentCan be provided.
[0114]
In addition, the present invention can provide a machine translation system, a program, and a machine translation method that can greatly reduce the burden on the user when translating a first language sentence into a second language sentence using a translation pattern.
[Brief description of the drawings]
FIG. 1 is a configuration diagram showing an embodiment of a machine translation apparatus according to the present invention.
FIG. 2 is a data array diagram of a pattern dictionary unit stored in the translation dictionary unit shown in FIG.
FIG. 3 is a diagram for explaining a flow of a series of processes of a process control unit shown in FIG.
FIG. 4 is a diagram for explaining a flow of a series of processes of a translation processing unit shown in FIG.
FIG. 5 is a structural connection diagram of a constant part and a variable part of an input source text.
FIG. 6 is a diagram illustrating conversion processing for a constant part.
FIG. 7 is a view for explaining variable part conversion processing;
FIG. 8 is a diagram for explaining a flow of a series of processes in the similarity calculation unit of the translation processing unit shown in FIG. 1;
FIG. 9 is a diagram illustrating an example of a concept dictionary unit stored in a translation dictionary unit.
FIG. 10 is a block diagram showing another embodiment of the machine translation apparatus according to the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Input part, 2 ... Process control part, 2a ... Instruction judgment means, 2b ... Process control means, 2c ... Confirmation detection means, 3 ... Translation dictionary part, 3g ... Pattern dictionary part, 3h ... Concept dictionary part, 4 ... Translation processing unit, 4a ... pattern matching extraction unit, 4b ... variable condition check means, 4c ... similarity calculation means, 4d ... translation generation processing means, 5 ... output part, 6, 7 ... recording medium, 8 ... registration pattern part.

Claims (5)

In addition to the knowledge information necessary for translating the first language sentence into the second language sentence, each of the character string pattern, condition, and reference source text composed of variable parts that can be changed and constant parts that are fixed expressions A translation dictionary part that stores a plurality of associated translation patterns;
Control means for judging a translation processing instruction input from the input unit and outputting a translation processing start instruction;
Based on the translation processing command output from the control means, means for acquiring various information necessary for translation processing of each word constituting the input original text that is the first language based on the knowledge information;
Means for comparing and collating the input original text and the character strings of the respective translation patterns, and extracting translation pattern candidates;
Variable condition checking means for checking whether each variable part of each extracted translation pattern candidate meets the conditions of the corresponding translation pattern based on the various information;
When there are a plurality of translation patterns that meet the conditions by this variable condition checking means, the similarity between the variable equivalent part of the input original text and the variable equivalent part of the reference original text of each translation pattern is calculated, and the highest similarity is obtained. A similarity calculation means for selecting a translation pattern,
A machine translation device that applies the selected translation pattern and translates it into the second language sentence. The machine translation device according to claim 1,
The similarity calculation means provides a concept dictionary that defines a connection relationship between concepts in the translation dictionary unit, specifies an equivalent concept of the concept dictionary corresponding to a unit word of each variable equivalent part of the input original sentence, A machine translation apparatus characterized in that a coupling distance of an equivalent concept is digitized, and a similarity degree of a variable equivalent part of a reference original text of each translation pattern is calculated from the numeric value. The machine translation apparatus according to claim 1 or 2 ,
A confirmation detection means for detecting an instruction to confirm the translation input from the input unit with respect to a translation result obtained by applying the translation pattern, and when the translation detection is determined by the confirmation detection means, the input original text The machine translation apparatus further comprising: a pattern registration unit that stores the information in association with a reference original of the applied translation pattern. The machine translation device according to claim 3 ,
The machine translation apparatus, wherein the confirmation detection means detects a translation edit lock processing instruction for avoiding a translation result being rewritten by later editing. The machine translation apparatus according to claim 1 or 2 ,
Confirmation detection means for detecting that the translation has been confirmed when the cursor moves from the sentence of the translation result to the translation target sentence behind, and if the translation detection is determined by this confirmation detection means, the translation result before the movement of the cursor A machine translation apparatus, further comprising: a pattern registration unit that stores an input original corresponding to the sentence in association with a reference original of the applied translation pattern.

Images