JPWO2013014877A1 - Reliability calculation device, translation reliability calculation utilization method, and translation engine program - Google Patents

Abstract

Even if statistical machine translation is not premised on machine translation, the reliability of the translation result can be appropriately obtained, and if the reliability is low, the user on the input side can be prompted to re-input appropriately. To. The translation reliability is determined based on the distance between the translation unit 501 that performs machine translation on the input sentence, the reverse translation unit 502 that performs reverse translation processing on the translation result, and the input sentence and the reverse translation result. Translation reliability calculation means 503 to be calculated and reliability presentation for presenting at least one of the input sentence, translation result and reverse translation result to the user together with the reliability corresponding to the input sentence, translation result or reverse translation result Means 504, and when the translation reliability is less than or equal to a predetermined threshold, the reliability presentation means 504 receives at least the input sentence or the reverse translation result together with the translation reliability and the side on which the first language is input. Presented to the user.

Description

  The present invention relates to a reliability calculation apparatus, a reliability calculation usage method, and a translation engine program for determining the reliability of machine translation and speech translation.

  In recent years, machine translation technology for converting character information such as text written in a first language into a different second language has been actively used. In addition, speech translation technology that uses speech instead of text as input and uses speech recognition results as input for machine translation has been put into practical use.

  If such a machine translation or speech translation system is used, people with different languages can communicate with each other, so there is high expectation for this technology. On the other hand, speech recognition and machine translation techniques are not perfect, and results may include misrecognition and mistranslation. If the result including misrecognition and mistranslation is presented to the other party, communication may be disrupted, but it is generally difficult for the user who is the speaker himself to judge the validity of the final translation result. is there.

  For this reason, attempts have been made to assign numerical values such as confidence and reliability from the system side to the speech recognition results and translation results. When reliability is given to speech recognition results and translation results, communication failures can be avoided by rejecting low-confidence utterances / inputs, or a speaker can rephrase the low-reliability part to bring it closer to the correct result. It becomes possible.

  For example, Patent Document 1 discloses a method of using reliability (RSCM) based on a sum of ranks in a machine translation output candidate selection apparatus. This method is based on the top hypothesis of each N-best based on the rank of the hypothesis that is each translation candidate among the top N translation candidates (N-best) output by Statistical Machine Translation (SMT). The posterior probability of the word is calculated and the reliability is obtained.

  Further, in Patent Document 2, when machine translation from a reference text in a first language into a text in a second language is performed, the translated text in the second language is converted into a folded translation text in the first language. A data conversion appropriate evaluation method is described in which the similarity between the reference text before translation and the data in the first language after reverse conversion is converted into an appropriate conversion value.

  Patent Document 3 discloses the N best hypothesis output from a speech recognition module in a machine translation system that performs speech recognition and machine translation into a second language for input speech in a first language. And re-scoring each of N × K translation candidates based on the K translation candidates for each of the N best hypotheses output from the statistical machine translation module and the information associated therewith. A configuration including a rescoring module that outputs a translation that achieves the highest score as the best translation is described.

JP 2006-039616 A JP 2006-252323 A JP 2006-053683 A

  However, the method described in Patent Document 1 is based on the premise that statistical machine translation is used as a translation process. For this reason, it is not applicable when using grammatical rule type translation that performs translation processing on a grammar base (rule base). This is because the concept of likelihood is not clear in grammatical rule translation. Currently, it is generally considered that grammatical rule-based translation can achieve more accurate translation than statistical machine translation.

  Further, the method described in Patent Document 1 gives reliability only to the processing result of machine translation. Therefore, when there is an unclear part in the input-side sentence in machine translation, it is not possible to prompt the user who has the input-side language to appropriately input again. In addition, when performing machine translation after performing speech recognition on a sentence that is input in speech, if there is a possibility that it will eventually be translated correctly if part of the original sentence is changed, It is difficult for the user on the input side to determine whether or not to re-input the original text only with the reliability of the translation processing result.

  If the method described in Patent Document 2 is used, the reliability of machine translation can be obtained without assuming statistical machine translation as machine translation. However, since the method described in Patent Document 2 gives reliability only to the processing result of machine translation, as in Patent Document 1, it is appropriately re-input to a user having an input language. There is a problem that cannot be urged.

  In addition, the method described in Patent Document 3 combines N × K translation candidate candidates by combining feature amounts (for example, likelihood information reflecting each accuracy) obtained from the speech recognition module and the machine translation module. Since a score is assigned to each of these, it may be possible to obtain a translation result with higher reliability in consideration of not only the result of translation processing but also the result of speech recognition processing. However, the method described in Patent Document 3 assumes statistical machine translation as machine translation. Further, it is only assumed that the re-scored score is used to select the best translation from a plurality of translation candidates. For this reason, it is difficult to know where the problem is in the evaluation that is calculated by unifying by modeling, and it is not possible to prompt the user to input again appropriately based on the reliability.

  Therefore, the present invention can appropriately obtain the reliability of the translation result without assuming statistical machine translation as the machine translation, and appropriately re-input to the input side user when the reliability is low. An object of the present invention is to provide a reliability calculation device, a translation reliability calculation usage method, and a translation engine program. Here, in the case where the reliability for the translation result is input by a method other than text data such as speech, the reliability is the first for the sentence or sentence data (for example, speech data) in the input first language. This refers to the reliability of the final processing result of machine translation performed through language recognition processing. That is, it means the reliability considering not only the processing result of the machine translation process but also the processing result of the preceding recognition process.

  The reliability calculation apparatus according to the present invention includes a translation unit that performs translation processing on an input sentence into a second language that is different from the first language that is the language of the input sentence, and translation processing that is performed by the translation unit A reverse translation means for performing a reverse translation process to the first language which is the language of the input sentence, and a reverse translation result obtained as a result of the reverse translation process by the input sentence and the reverse translation means. A translation reliability calculation means for calculating a translation reliability based on the distance between the input sentence, the translation result, and the reverse translation result, at least one of the input sentence, the translation result, and the reverse translation result. A reliability presenting means for presenting to the user together with the degree of reliability, and the reliability presenting means first outputs at least the input sentence or the reverse translation result together with the translation reliability when the translation reliability is below a predetermined threshold. Entered the language of Characterized in that it presented to the user.

  The translation reliability calculation and utilization method according to the present invention performs a translation process on an input sentence into a second language that is different from the first language that is the language of the input sentence, and results of the translation process The translation result is reverse translated into the first language, which is the language of the input sentence, and the translation reliability is determined based on the distance between the input sentence and the reverse translation result obtained as a result of the reverse translation process. When calculating and presenting at least one of the input sentence, translation result and reverse translation result to the user who has input the first language together with the reliability corresponding to the input sentence, translation result or reverse translation result In addition, when the translation reliability is equal to or lower than a predetermined threshold, at least the input sentence or the reverse translation result is presented to the user who has input the first language together with the translation reliability.

  Further, the translation engine program according to the present invention includes a computer, a translation process for converting an input sentence described in a first language into a second language that is a language different from the first language, and a result of the translation process. Translation that calculates translation reliability based on the distance between the reverse translation process that converts the obtained translation result into the first language that is the language of the input sentence, and the reverse translation result that is obtained as a result of the reverse translation process A reliability calculation process and a reliability presentation process for presenting at least one of the input sentence, the translation result, and the reverse translation result together with the reliability corresponding to the input sentence, the translation result, or the reverse translation result to the user. In the reliability presentation process, when the translation reliability is below a predetermined threshold, at least the input sentence or the reverse translation result is presented to the user who has input the first language together with the translation reliability. The features.

  According to the present invention, even if statistical machine translation is not premised on machine translation, the reliability of the translation result can be appropriately obtained, and when the reliability is low, the input user is appropriately re-input Can be encouraged.

It is the block diagram which showed the example of a structure of the reliability calculation apparatus of 1st Embodiment. It is a flowchart which shows an example of the processing flow of the reliability calculation process in the reliability calculation apparatus of 1st Embodiment. It is a block diagram which shows the other structural example of the reliability calculation apparatus of 1st Embodiment. It is a block diagram which shows the other structural example of the reliability calculation apparatus of 1st Embodiment. It is a block diagram which shows the structural example of the machine translation apparatus of a 1st Example. It is a flowchart which shows an example of the processing flow of the translation process in a 1st Example. It is explanatory drawing which shows the example of calculation of reliability. It is a block diagram which shows the structural example of the reliability calculation apparatus of 2nd Embodiment. It is a flowchart which shows an example of the processing flow of the reliability calculation process in the reliability calculation apparatus of 2nd Embodiment. It is a block diagram which shows the other structural example of the reliability calculation apparatus of 2nd Embodiment. It is a block diagram which shows the structural example of the speech translation system of a 2nd Example. It is a flowchart which shows an example of the processing flow of the translation process in a 2nd Example. It is explanatory drawing which shows the example of calculation of reliability. It is a block diagram which shows the outline | summary of this invention.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

Embodiment 1. FIG.
FIG. 1 is a block diagram showing a configuration example of a reliability calculation apparatus according to the first embodiment of the present invention. The reliability calculation apparatus 10 shown in FIG. 1 inputs text or a character string, performs machine translation, calculates the reliability, and outputs it.

  The reliability calculation device 10 includes a machine translation unit 13, a reverse translation unit 14, and a translation reliability calculation unit 15. In addition, the reliability calculation apparatus 10 includes an input unit 11 and an output unit 12 directly or indirectly. Note that “indirectly” means, for example, that other devices connected via a network may be provided.

  The reliability calculation device 10 may be, for example, a general-purpose computer system that includes a CPU, a RAM, a ROM, and a nonvolatile storage device (not shown). That is, in the reliability calculation device 10, an information processing device such as a CPU reads an OS and a reliability calculation program stored in a RAM, a ROM, or a non-volatile storage device, and executes predetermined processing including reliability calculation processing. It may be configured to. Thereby, the reliability of a translation result can be output. In addition, the reliability calculation apparatus 10 does not need to be one computer system, and may be configured by a plurality of computers.

  The input unit 11 is a keyboard or the like for inputting a sentence described in a first language before translation, that is, an original sentence, to the reliability calculation apparatus 10. Note that the input unit 11 may input instruction information indicating in which language the original sentence is translated. The output unit 12 is a display or the like that displays the reliability output from the reliability calculation device 10 to the user.

  The machine translation unit 13 translates the input original sentence into a translated second language sentence, that is, a translated sentence. The technique used as the translation technique may be statistical machine translation, grammatical rule translation, or example-based translation. Since the translation technique itself is well known as a known technique, a detailed description thereof is omitted here.

  The reverse translation unit 14 receives the translation translated by the machine translation unit 13 and translates the input translation into the original first language. The translation method here may be a method different from the translation method in the machine translation unit 13, but it is generally easier to use the same method.

  The translation reliability calculation unit 15 receives the original text described in the first language before translation and the reverse translation result translated into the first language again by the reverse translation unit 14, and reverses the input original text. Find the distance from the translation result and use it as the reliability. Hereinafter, the reliability calculated by the translation reliability calculation unit 15 (specifically, the reliability with respect to the translation processing result of the machine translation unit 13) is referred to as translation reliability. In the present embodiment, the translation reliability calculated here is directly used as the reliability for the translation result.

  The translation reliability calculation unit 15 may obtain an evaluation value based on a predetermined scale representing the distance between the original text and the reverse translation result, that is, the closeness of both sentences, for example, using a BLEU score using the shared N-gram number. Good. The BLEU score is a scale used when statistical machine translation is evaluated, and a word chain (N-gram) shared by both of how much a sentence of a certain translation result can reproduce a correct sentence. The number is calculated using the number of 0 and output in the range of 0-1. An example of a BLEU score calculation formula is shown below as Formula (1).

In the formula (1), BP represents a penalty longitudinal, N represents represents the maximum length of the N-gram, _{w n} denotes a coupling weight (which may be 1 / N), _{P n} is modified N-gram accuracy Represents. The modified N-gram accuracy is obtained by dividing the maximum number of shared N-grams by the total number of N-grams. The BLEU score has a problem that it becomes low accuracy when the sentence is short, but accuracy can be improved by combining with other measures such as reliability of speech recognition as described later. Moreover, although the example which uses a BLEU score as distance was shown here, the calculation method of distance is not restricted to this method.

  It is also possible to obtain the distance between the original text and the reverse translation result in units of words or word strings. In such a case, after obtaining the distance in units of words or word strings, it is preferable to normalize the whole sentence. Thus, by obtaining the distance, that is, the translation reliability in units of words or word strings, it becomes possible to present a low reliability section in one sentence. As a method of extending the translation reliability to the sentence unit, the obtained word unit reliability may be normalized for the whole sentence, or the whole sentence reliability may be regarded as the simultaneous probability of the reliability of each word unit. Good.

The following equation (2) is obtained by normalizing the word posterior probabilities for the entire sentence when the translation reliability of each word w _i constituting the sentence {w ⁿ _l } is CM _MT (w _i ), and the translation reliability CM. It is a calculation example in the case of extending _MT in sentence units.

Also, the following equation (3) is a calculation example in the case where the reliability of each word is regarded as a probability event, and the translation reliability CM _MT is expanded in sentence units with the reliability of the entire sentence as the joint probability.

  In addition, since it can be said that the reliability by reverse translation represents the reproducibility by translation, assigning the reliability has the same meaning as assigning to the reverse translation result or to the original text before translation. That is, the low confidence interval can be presented not only for the translation result but also for the input original text or reverse translation result.

  Next, the operation of this embodiment will be described. Below, the reliability calculation process in the reliability calculation apparatus 10 of this embodiment is demonstrated with reference to FIG. FIG. 2 is a flowchart illustrating an example of a processing flow of the reliability calculation process in the reliability calculation apparatus 10 of the present embodiment. As shown in FIG. 2, first, the reliability calculation device 10 inputs a text that is an original sentence to be translated (step S <b> 100). Here, the input of the sentence described in the first language, that is, the original sentence is received via the input unit 11, and the received original sentence is output to the machine translation unit 13 and the translation reliability calculation unit 15.

  Next, the machine translation unit 13 translates the input original text (step S101). The machine translation unit 13 performs a translation process on the input original sentence from the first language that is the original language to the second language that is the other language specified, and reversely translates the translation result To the unit 14.

  Next, the reverse translation unit 14 performs reverse translation on the translation that is the translation result of the machine translation unit 13 (step S102). The reverse translation unit 14 executes a translation process from the second language, which is the translated language, to the first language, which is the original language, again on the translation that is the translation result input from the machine translation unit 13. The reverse translation result, which is the translation result, is output to the translation reliability calculation unit 15.

  Next, the translation reliability calculation unit 15 calculates the reliability (step S103). The translation reliability calculation unit 15 includes the original text described in the first language before translation input via the input unit 11 and the reverse translation result translated into the first language again by the reverse translation unit 14. Is input, and the distance between the two is calculated as the reliability. In addition, although it is suitable to obtain | require the distance as a translation reliability using shared N-gram, it is not restricted to this.

  Finally, the reliability calculation device 10 outputs the result (step S104). The reliability calculation device 10 outputs the translation reliability obtained by the translation reliability calculation unit 15 via the output unit 12 such as a display and displays it to the user. At this time, the reliability calculation device 10 may output the reverse translation result together with the translation reliability.

  In the above embodiment, the distance between the original sentence and the reverse translation result is calculated when calculating the distance between the original sentence and the reverse translation result. At this time, the original sentence and the reverse translation result may be developed synonymously. For example, a synonym expansion unit 161 and a synonym dictionary 162 may be further provided as in the reliability calculation device 10b illustrated in FIG. FIG. 3 is a block diagram illustrating another configuration example of the reliability calculation apparatus according to the present embodiment.

  The synonym dictionary 162 stores words or word strings having different pronunciations and notations but having the same meaning in association with each other. The synonym expansion unit 161 expands each word or word string of the original text or the reverse translation result into a plurality of different expressions using the synonym dictionary 162.

  In such a case, the translation reliability calculation unit 15 calculates distances for a plurality of candidates that have been synonymously expanded by the synonym expansion unit 161, and finally calculates the distance from the closest candidate. You may select as translation reliability. By adopting such a configuration, for example, even if a word has the same meaning but has the same meaning but different notation as a result of reverse translation, it can be determined that the distance is close if the meaning is the same. Therefore, it is possible to more appropriately calculate the reliability by preventing the distance from being unduly increased.

  Further, for example, as shown in FIG. 4, a reliability presentation unit 17 is further provided, and the reverse translation result obtained by the reverse translation unit 14 and the translation reliability obtained by the reliability submission unit 17 of the reliability calculation device 10. The reliability obtained by the calculation unit 15 may be presented together. FIG. 4 is a block diagram illustrating another configuration example of the reliability calculation apparatus according to the present embodiment.

  As described above, according to the present embodiment, the distance between the original sentence and the reverse translation result obtained by reversely translating the translated sentence that is the translation processing result into the original language is calculated as the translation reliability. The reliability can be appropriately calculated even by using a translation method other than statistical machine translation such as type translation. Thus, by obtaining an appropriate degree of reliability, it is possible to reduce communication failures as a result.

  Also, by outputting the reverse translation result together with the translation reliability, for example, when the output destination is low reliability, the user who has input the original text can present the result together with the reverse translation result to the user. Even if you can't understand the second language you want to translate, you can check the result in the speaker's language, and as a result, you can estimate the cause and location of the low reliability, Paraphrasing can be performed.

  Next, this embodiment will be described using a more specific example. Hereinafter, a first embodiment of the reliability calculation apparatus according to the present invention will be described using the example of the machine translation apparatus 100 shown in FIG. The machine translation apparatus 100 illustrated in FIG. 5 includes an input unit 11, a reliability calculation unit 10, a result determination unit 110, and an output unit 12.

  The machine translation device 100 is realized by, for example, a general-purpose computer system including a CPU, a RAM, a ROM, and a nonvolatile storage device. That is, the CPU may be configured to read a machine translation program in which the operations of the reliability calculation unit 10 and the result determination unit 110 are described in a program language, and execute machine translation processing according to the read program. The machine translation apparatus 100 does not have to be a single computer system, and may be configured by a plurality of computer systems. In such a configuration, the machine translation apparatus 100 operates as, for example, one translation engine, and provides a highly accurate translation result as a result by rejecting a low-reliability result.

  The input unit 11 receives text in the first language input to the apparatus and outputs the text to the reliability calculation unit 10. The input unit 11 may be a keyboard as a text input device, for example.

  The reliability calculation unit 10 receives text input from the input unit 11 and performs a translation process and a reliability calculation process. The reliability calculation unit 10 corresponds to the reliability calculation device 10 of FIG. Therefore, only the operation different from that in FIG. 1 will be described below, and the description of the same operation as the configuration in FIG. 1 will be omitted. The reliability calculation unit 10 in this example may correspond to the reliability calculation device 10b in FIG. 3 or the reliability calculation device 10c in FIG.

  The machine translation unit 13 translates the first language into the second language, that is, the translation, not only to the reverse translation unit 14 but also to the result determination unit 110. The translation reliability calculation unit 15 outputs the calculated reliability to the result determination unit 110.

  The result determination unit 110 determines whether the translation of the translation result input from the machine translation unit 13 should be output according to the reliability input from the translation reliability calculation unit 15. For example, if the input reliability is greater than or equal to a predetermined threshold value, the translation result is output, and if not, it is determined that the reliability is low and the result is rejected without outputting the translation result.

  The output unit 12 outputs a translation that is a translation result obtained from the result determination unit 110. For example, the translated text, that is, the text written in the second language may be displayed on the screen, or may be read out as a voice using a speaker incorporating a voice synthesizer for outputting the text. Moreover, the output part 12 does not need to output anything when the result determination part 110 rejects a translation result, and outputs a reverse translation result and reliability with the message which shows that it rejected. Also good.

  In the example illustrated in FIG. 5, the reliability calculation unit 10 includes the machine translation unit 13, but the configuration uses an existing machine translation unit (such as a search engine) included in the machine translation apparatus 100. It is also possible to do. In such a case, the reverse translation unit 14 may perform reverse translation using a translation engine provided in the machine translation apparatus 100.

  Next, the overall flow of the translation processing according to the present embodiment will be described with reference to the flowchart shown in FIG. FIG. 6 is a flowchart showing an example of the processing flow of translation processing in the first embodiment of the present invention. As shown in FIG. 6, first, the machine translation apparatus 100 inputs text in a first language (step S200). Specifically, the machine translation apparatus 100 accepts text in the first language input via the input unit 11 and outputs the text to the reliability calculation unit 10.

  Next, the machine translation device 100 translates the input text (step S201). Specifically, the machine translation device 100 performs translation processing into the second language by the machine translation unit 13 on the text of the first language input from the user via the input unit 11. The translation result is output to the reverse translation unit 14 and the result determination unit 110.

  Next, the machine translation device 100 performs reverse translation (step S202). Specifically, the machine translation apparatus 100 performs reverse translation processing on the translation result obtained by the machine translation unit 13 to the first language again by the reverse translation unit 14.

  Next, the machine translation device 100 calculates the reliability (step S203). Specifically, in the translation reliability calculation unit 15 of the machine translation apparatus 100, the original text described in the first language before translation input via the input unit 11 and the first in the reverse translation unit 14 again. The reverse translation result translated into the language is input, and the distance between the two is calculated as the reliability.

  FIG. 7 is an explanatory diagram illustrating an example of calculating reliability in the present embodiment. For example, as shown in FIG. 7A, for an input sentence “what would you like to drink” in English as the first language, the translation result in Japanese as the second language is “what. Suppose you want to drink (nani wo nomitaidesuka) "and the reverse translation result is" What would you like to drink? " At this time, the reverse translation result completely reproduces the input sentence, and the reliability of the input sentence and the reverse translation result is 1.0 when the BLEU score is calculated in 1-gram.

  On the other hand, referring to the example of FIG. 7B, the translation result into the second language Chinese is shown for the input sentence “do you have chairs for children” in the first language English. Suppose that the reverse translation result is “Is there a lecture of a child?”. Here, mistranslation has occurred, and when the reliability is calculated in the same manner, it becomes 0.0.

  At this time, as shown in the configuration of FIG. 3, the synonym expansion unit 161 may expand each word or word string of the original text or the reverse translation result into a plurality of candidates using the synonym dictionary 162. In such a case, the translation reliability calculation unit 15 calculates the distance for each combination of a plurality of synonym expanded candidates, and selects the closest one as the final translation reliability. Also good. For example, on the source side of Fig. 7 (b), "chairs" is converted to "sofas", "stools", etc., or "do you have" is converted to "are there", etc., and the source text is expanded into multiple candidates. May be. Similarly, in the reverse translation result, “lecture” may be converted into “class” or the like and expanded into a plurality of candidates. The distances may be calculated for all combinations of the plurality of candidates developed in this manner, and the closest one may be selected as the reliability. In addition, the expansion to a plurality of candidates may be any one. By performing synonym expansion, it is possible to calculate highly reliable reliability even when the meaning is the same at the time of reverse translation but the notation has changed.

  Next, the machine translation apparatus 100 performs a result determination (step S204). Specifically, the machine translation apparatus 100 determines whether the result determination unit 110 should output the translation of the translation result notified from the machine translation unit 13 according to the reliability notified from the translation reliability calculation unit 15. Judgment. For example, in the case of FIG. 7A, since the reliability is as high as 1.0, it is determined that acceptance, that is, output is possible. Further, in the case of FIG. 7B, since the reliability is as low as 0.0, rejection, that is, output is impossible.

  Next, the machine translation apparatus 100 outputs the result (step S205). Specifically, the machine translation apparatus 100 outputs the translation result obtained as a result of the determination by the result determination unit 110 via the output unit 12 and displays the result on a display, for example.

  Thus, by rejecting the result with low reliability, it is possible to realize translation with high accuracy as a result. In any case, the translation result may be output and the reliability may be displayed at the same time. Moreover, when rejected with low reliability as shown in the configuration of FIG. 4, the reverse translation result may be displayed together with the reliability to prompt the user to input again.

  In addition, when the calculation of reliability is calculated | required per word or word string, you may show reliability as FIG.7 (c) and (d), for example. For example, as shown in FIG. 7 (c), the input sentence “what would you like to drink” in the first language, English, the translation result in Japanese, the second language, is “What drink Nani wo nomitaidesuka ”and the reverse translation result is“ What would you like to drink? ”. At this time, since the reverse translation result completely reproduces the input sentence, the reliability is 1.0 for all words. On the other hand, referring to the example of FIG. 7 (d), the translation result in Japanese, which is the second language, for the input sentence “which will you take” in the first language, “Which is you? ”Do you take (anata ha dochira ga kakarimasuka)” and the reverse translation result is “Which is your take?” Here, mistranslation occurs, but it is easy for the user who sees the word reproduced by reverse translation to rephrase the low reliability part that was not reproduced by displaying it with high reliability. become.

  In the examples of FIGS. 7C and 7D, for simplicity, the reliability is set to 1.0 for 1-gram where the surface layers match, and 0.0 for non-matching, The reliability value is not limited to this. For example, it is possible to consider not only the surface layer but also the combination with the part of speech and the reading information, the difference in utilization, and the like. For example, 0.8 may be used if only the surface layer is matched, 0.9 may be used if the parts of speech are also matched, and only the difference in utilization (you and your in the example of FIG. 7 (d)). If so, a score of 0.3 is also possible.

Embodiment 2. FIG.
Next, a second embodiment of the present invention will be described. In the present embodiment, more appropriate reliability is calculated by assuming input other than text, such as voice recognition, and integrating the reliability. FIG. 8 is a block diagram illustrating a configuration example of the reliability calculation apparatus according to the second embodiment of this invention. The reliability calculation device 10d illustrated in FIG. 8 receives the word or word string of the speech recognition result and the reliability thereof, and outputs a reliability obtained by integrating the reliability of the speech recognition result and the reliability of the translation result.

  The reliability calculation apparatus 10d illustrated in FIG. 8 includes a machine translation unit 13, a reverse translation unit 14, a translation reliability calculation unit 15, and a reliability integration unit 21. In addition, the reliability calculation device 10d may include a speech recognition result input unit 210 and an output unit 12 directly or indirectly. Here, the machine translation unit 13, the reverse translation unit 14, and the translation reliability calculation unit 15 are the same as those in the first embodiment shown in FIG. Below, only a different part from 1st Embodiment is demonstrated.

  The speech recognition result input unit 210 inputs a speech recognition result for a sentence or sentence to be translated input by speech and its reliability.

  The machine translation unit 13 and the translation reliability calculation unit 15 receive the speech recognition result from the speech recognition result input unit 210 as the original text.

  The reliability integration unit 21 inputs the speech recognition reliability, which is the reliability of the speech recognition result, from the speech recognition result input unit 210, and further inputs the translation reliability from the translation reliability calculation unit 15. Integrate reliability and translation reliability. The reliability integration unit 21 may integrate the speech recognition reliability and the translation reliability by linear interpolation, for example, to obtain an integrated reliability.

For example, if the integrated reliability is CM, the speech recognition reliability is CM _ASR , and the translation reliability is CM _MT , the integrated reliability CM is expressed by the following equation (4) using the linear interpolation coefficient α. The

CM = (1−α) × CM _ASR + α × CM _MT (4)

As the speech recognition reliability CM _ASR , for example, there is a method using a word posteriori probability (word posteriori probability) calculated by the speech recognition device for each word. Note that the method of providing the voice recognition reliability is not limited to this. Since the method for calculating the word posterior probability is well known as a known technique, a detailed description thereof is omitted here.

By the way, in general, the translation reliability CM _MT is given in sentence units, whereas the speech recognition reliability CM _ASR is often given in word units like word posterior probabilities. In such a case, the speech recognition reliability CM _ASR may be extended in sentence units. As a method of extending the speech recognition reliability CM _ASR in sentence units, the obtained word posterior probabilities may be normalized for the entire sentence, or the posterior probabilities for the entire sentence may be regarded as the simultaneous probabilities for each word posterior probability. .

The following equation (5) is obtained by normalizing the word posterior probabilities for the entire sentence, where P (w _i | X) is the posterior probability of each word w _i constituting the sentence {w ⁿ _l }, and the speech recognition reliability. It is a calculation example in the case of extending CM _ASR in sentence units.

Further, the following formula (6) is a calculation example in the case where the posterior probability of the whole sentence is regarded as the simultaneous probability of each word posterior probability and the speech recognition reliability CM _ASR is expanded in sentence units.

Conversely, in order to set the machine translation reliability CM _MT in units of words or word strings, the shared N-gram and the speech recognition N-gram language model or the voice recognition reliability CM _ASR are combined in units of words or word strings. This is possible. For example, for a word that is not a shared N-gram, that is, a word that has not been reproduced by reverse translation, a certain penalty value is added to the speech recognition reliability CM _ASR to obtain an integrated reliability CM in units of words, Normalize the whole sentence. Thus, by obtaining the integrated reliability in units of words, it becomes possible to present a low reliability section within a sentence.

  The output unit 12 is a display or the like that displays the reliability (here, integrated reliability) output from the reliability calculation device 10d to the user. The output unit 12 may display the reliability obtained for each sentence or each word along with the translation result. In the case of this embodiment, as the reliability to be presented, the reliability of the recognition result (speech recognition reliability, etc.) may be used when assuming other modules such as translation reliability and speech recognition in the input unit. Good. These may be used in combination as the integrated reliability.

  In this embodiment, it is assumed that the speech recognition result is input as the original text, but the configuration placed in the preceding stage is not limited to the speech recognition. For example, a similar configuration is possible by inputting a character input result (or character recognition result) using an OCR (Optical Character Reader) or a touch panel and its reliability.

  Also in the present embodiment, as in the first embodiment, the synonym expansion unit 161 and the synonym dictionary 162 are provided and synonym expansion is performed, so that more appropriate reliability calculation can be performed. At this time, it is possible to further increase efficiency by performing synonym expansion only on words or word strings that are commonly held by the speech recognition dictionary used for speech recognition and the translation dictionary used for machine translation. That is, it is possible to reduce the amount of processing by preventing useless expansion to unknown words and expanding synonyms only within a translatable range.

  Further, for example, as in the reliability calculation device 10e shown in FIG. 10, the reliability presentation unit 17 is further provided, the reverse translation result obtained by the reverse translation unit 14 in the reliability submission unit 17, and the reliability integration unit 21. You may make it show together with the reliability obtained by. FIG. 10 is a block diagram illustrating another configuration example of the reliability calculation apparatus according to the present embodiment. The reliability presented by the reliability presentation unit 17 does not need to be in units of sentences (or the entire utterance), and may be presented in units of words, for example.

  Next, the operation of this embodiment will be described. Below, the reliability calculation process in the reliability calculation apparatus 10d of this embodiment is demonstrated with reference to FIG. FIG. 9 is a flowchart illustrating an example of a processing flow of reliability calculation processing in the reliability calculation device 10d of the present embodiment. As shown in FIG. 9, first, the reliability calculation device 10d inputs a speech recognition result as an original sentence to be translated and its reliability (step S300). Here, the input of the speech recognition result is accepted as a sentence written in the first language, that is, the original text via the speech recognition result input section 210, and the received original text is sent to the machine translation section 13 and the translation reliability calculation section 15. Output. At the same time, an input of the voice recognition reliability of the voice recognition result is received, and the received voice recognition reliability is output to the reliability integration unit 21.

  Next, the machine translation unit 13 translates the input original text (step S301). The machine translation unit 13 performs a translation process on the input original sentence from the first language that is the original language to the second language that is the other language specified, and reversely translates the translation result To the unit 14.

  Next, the reverse translation unit 14 performs reverse translation on the translation that is the translation result of the machine translation unit 13 (step S302). The reverse translation unit 14 executes a translation process from the second language, which is the translated language, to the first language, which is the original language, again on the translation that is the translation result input from the machine translation unit 13. The reverse translation result, which is the translation result, is output to the translation reliability calculation unit 15.

  Next, the translation reliability calculation unit 15 calculates the reliability (step S303). The translation reliability calculation unit 15 uses the original text described in the first language before translation input via the speech result input unit 210 and the reverse translation translated into the first language again by the reverse translation unit 14. The result is input, and the distance between the two is calculated as the translation reliability. In the present embodiment, the translation reliability calculated here is simply handled as the reliability for the translation processing result.

  Next, the reliability integration unit 21 integrates the speech recognition reliability, which is the reliability of the original text recognition process, and the translation reliability (step S304). The reliability integration unit 21 integrates the speech recognition reliability input from the speech recognition result input unit 210 and the translation reliability input from the translation reliability calculation unit 15 by, for example, linear interpolation. Note that the reliability to be integrated may be a sentence unit or a word or a word string unit.

  Finally, the reliability calculation device 10d outputs the result (step S305). The reliability calculation device 10d outputs the integrated reliability obtained by the reliability integration unit 21 via the output unit 12 such as a display and displays it to the user. At this time, the reliability calculation apparatus 10d may output the reverse translation result together with the integrated reliability. In addition to the integrated reliability, the speech recognition reliability and the translation result reliability used to calculate the integrated reliability may be output as additional information.

  As described above, according to the present embodiment, when input other than text, such as speech recognition, is performed, integration with the reliability of recognition processing in the first language for the input original sentence is more appropriate. Reliable reliability can be calculated.

  In addition, if the reliability is low, presenting the reverse translation result and reliability together, even if the user who entered the original text cannot understand the second language to be translated, As a result, it is possible to check the result and to estimate the cause and location of the low reliability, and to rephrase or rephrase the portion. Furthermore, in the present embodiment, since reliability can be given in units of words, by emphasizing and presenting a sentence expression for the reverse translation corresponding to the word that has become low reliability, The result can be confirmed more easily, and as a result, the cause of the low reliability and the location can be estimated more easily, and the portion can be rephrased and rephrased.

  Next, this embodiment will be described using a more specific example. Hereinafter, an example of the speech translation system 200 shown in FIG. 11 will be described as a second embodiment of the reliability calculation apparatus according to the present invention. The speech translation system 200 illustrated in FIG. 10 includes a speech input unit 211, a speech recognition unit 212, a reliability calculation unit 10d, a result determination unit 220, and an output unit 12.

  The speech translation system 200 is realized by, for example, a general-purpose computer system that includes a CPU, a RAM, a ROM, and a nonvolatile storage device. That is, the CPU may be configured to read a speech translation program in which the operations of the speech recognition unit 212, the reliability calculation unit 10d, and the result determination unit 220 are described in a program language, and execute the speech translation process according to the read program. . Note that the speech translation system 200 does not have to be a single computer system, and may be configured by a plurality of computer systems. In such a configuration, the speech translation system 200 receives the speech of the first language that is the original sentence to be translated, translates the sentence or sentence represented by the speech into the second language, and trusts the translation result. Output with degrees.

  The voice input unit 211 receives a first language voice input to the system and outputs the first language voice to the voice recognition unit 211. For example, a microphone as an input device may be used.

  The speech recognition unit 210 performs speech recognition processing on the speech input via the speech input unit 211, and outputs the speech recognition result and the speech recognition reliability that is the reliability to the reliability calculation unit 10d.

  The reliability calculation unit 10d accepts the speech recognition result from the speech recognition unit 210 and the input of the reliability as the original text to be translated and the reliability of the original text, and the translation process, the calculation process of the translation reliability and the integrated reliability, I do. The reliability calculation unit 10d corresponds to the reliability calculation device 10d shown in FIG. Therefore, only operations different from those in FIG. 8 will be described below, and description of the same operations as those in the configuration in FIG. 8 will be omitted. In addition, the reliability calculation unit 10d of this example has a configuration equivalent to the reliability calculation device 10e of FIG. 10, and further includes a synonym expansion unit 161 and a synonym dictionary 162 as shown in FIG. Also good.

  The machine translation unit 13 translates the first language into the second language, that is, the translated sentence, not only to the reverse translation unit 14 but also to the result determination unit 220. The reliability integration unit 21 outputs the integrated reliability to the result determination unit 220. The result determination unit 220 determines whether the translation of the translation result input from the machine translation unit 13 should be output according to the reliability input from the reliability integration unit 21, that is, the integrated reliability.

  The output unit 12 outputs a translation that is a translation result obtained from the result determination unit 220. For example, the translated text, that is, the text written in the second language may be displayed on the screen, or may be read out by voice using a speaker incorporating a voice synthesizer for outputting the text as voice. Further, the output unit 12 does not need to output anything when the result determination unit 220 rejects the translation result, and outputs the reverse translation result and the reliability together with a message indicating that the result is rejected. Also good.

  Next, the overall flow of the translation processing according to the present embodiment will be described with reference to the flowchart shown in FIG. FIG. 12 is a flowchart showing an example of a processing flow of translation processing in the second embodiment of the present invention. As shown in FIG. 12, first, the speech translation system 200 inputs speech in the first language (step S400). Specifically, the speech translation system 200 accepts the first language speech input via the speech input unit 211 and outputs it to the speech recognition unit 212.

  Next, the speech translation system 200 recognizes the input speech (step S401). Specifically, in the speech translation system 200, the speech recognition unit 212 performs speech recognition processing on the speech input from the speech input unit 211, and obtains speech recognition results and speech recognition reliability that is reliability. The obtained voice recognition result and the voice recognition reliability are output to the reliability calculation unit 10d. Note that a speech recognition processing method for inputting speech and outputting text as a speech recognition result, and a method for calculating the reliability of the speech recognition result at that time are well known in the art. Therefore, detailed description is omitted here. In the speech reliability calculation unit 10d, the speech recognition result is output to at least the machine translation unit 13 and the translation reliability calculation unit 15, the obtained speech recognition result is output, and the speech recognition result is output to the reliability integration unit 21. Is done.

  Next, the speech translation system 200 translates the speech recognition result (step S402). Specifically, the speech translation system 200 translates the text of the speech recognition result input from the speech recognition unit 212 by the machine translation unit 13, and outputs the translation result to the reverse translation unit 14 and the result determination unit 220.

  Next, the speech translation system 200 performs reverse translation (step S403). Specifically, the speech translation system 200 performs reverse translation processing on the first language again by the reverse translation unit 14 on the translation result obtained by the machine translation unit 13.

  Next, the speech translation system 200 calculates translation reliability (step S404). Specifically, in the translation reliability calculation unit 15 of the speech translation system 200, the original sentence described in the first language before translation, which is the speech recognition result by the speech recognition unit 212, and the reverse translation unit 14 again The reverse translation result translated into one language is input, and the distance between the two is calculated as the translation reliability. The translation reliability calculated here is the reliability for the translation process.

  Next, the speech translation system 200 integrates reliability (step S405). Specifically, in the speech translation system 200, the reliability integration unit 21 integrates the input speech recognition reliability and the translation reliability by, for example, linear interpolation to obtain a final reliability for the translation result. .

  FIG. 13 is an explanatory diagram illustrating a calculation example of reliability in the present embodiment. For example, as shown in FIG. 13A, for an input sentence “what would you like to drink” in English, which is the first language, the speech recognition result is “what would you like to drink”. Since the speech recognition reliability is completely correct, it is assumed that it is 0.99 in units of sentences (entire utterances). Moreover, the translation result into Japanese, which is the second language, using the speech recognition result as the original text is “What do you want to drink (nani wo nomitaidesuka)” and the reverse translation result is “What would you like to drink? ”. At this time, the translation reliability was 1.0 because the original text whose reverse translation result was the speech recognition result was completely reproduced. In such a case, when these speech recognition reliability and translation reliability are linearly interpolated with a constant weight, the integrated reliability is 0.99.

  On the other hand, referring to the example of FIG. 13B, the speech recognition result is “you can check into wars before” for the input sentence “you can check in two hours before the departure time” in the first language, English. It is assumed that the speech recognition reliability is 0.79 because it includes an error. In addition, the translation result into Japanese, which is the second language, using the speech recognition result as the original text is “can check war before departure (shuppatsu jikoku no mae ni sensou wo check dekimasu)” and its reverse translation Assume that the result is "I can check a war before the departure time." At this time, the translation reliability is 0.67 because the reverse translation result includes an error, unlike the original text in which the speech recognition result is obtained. In such a case, when these speech recognition reliability and translation reliability are linearly interpolated with a constant weight, the integrated reliability is 0.77.

  Further, referring to the example of FIG. 13C, the speech recognition result is “I'll walk to the bank” for the input sentence “I'll watch over your bags” in the first language, English. It is assumed that the voice recognition reliability is 0.64. Further, it is assumed that the translation result into the second language, Chinese, using the speech recognition result as the original text is as shown in the figure, and the reverse translation result is “I'll walk to the bank”. At this time, the translation reliability was 1.0, which is a complete reproduction of the original text whose reverse translation result is the speech recognition result. In such a case, if these speech recognition reliability and translation reliability are linearly interpolated with a constant weight, the integrated reliability is 0.71.

  Further, referring to the example of FIG. 13D, the speech recognition result is “do you have chairs for children” for the input sentence “um do you have chairs for children” in the first language, English. Suppose that the voice recognition reliability is 0.94. Also, assume that the result of translation into Chinese, which is the second language, with the speech recognition result as the original text is as shown in the figure, and the reverse translation result is “Is there a lecture of a child?”. At this time, the translation reliability was set to 0.0 because the reverse translation result was different from the original sentence that was the speech recognition result and contained an error. In such a case, when these speech recognition reliability and translation reliability are linearly interpolated with a constant weight, the integrated reliability is 0.75.

  Next, the speech translation system 200 performs a result determination (step S406). Specifically, the speech translation system 200 determines whether the result determination unit 220 should output the translation of the translation result input from the machine translation unit 13 according to the integrated reliability input from the reliability integration unit 21. Judgment.

  For example, in the example of FIG. 13 (a), both the speech recognition result and the machine translation result are correct, and the integrated reliability (denoted as speech translation reliability in the figure) is as high as 0.99. judge. On the other hand, in the example of FIG. 13B, both the speech recognition result and the machine translation result include an error, and the integrated reliability is not sufficiently high as 0.77 (specifically, it is lower than a predetermined threshold). ), It is determined that it is rejected, that is, output is impossible.

  In the example of FIG. 13C, the machine translation result is correct but the speech recognition result includes an error. In the example of FIG. 13D, the speech recognition is almost correct but the translation result includes an error. Also in the example, since the integrated reliability is not sufficiently high as 0.77, it is determined that rejection, that is, output is impossible.

  Next, the speech translation system 200 outputs the result (step S407). Specifically, the speech translation system 200 outputs the translation result obtained as a result of the determination by the result determination unit 220 via the output unit 12, thereby displaying the result on a display or the like, for example.

  For example, referring to the example of FIG. 13D, although the input speech is “um do you have chairs for children”, the reverse translation result is “Is there a lecture of a child?” It is a degree. At this time, for example, when the reliability presentation unit 17 as shown in FIG. 10 is provided, the reliability submission unit 17 receives the reverse translation result obtained by the reverse translation unit 14 and the reliability obtained by the reliability integration unit 21. By presenting together with the degree, it is possible to cause the speaker side, that is, the English side user to estimate the cause of the low reliability and its location. As a result, the part can be reworded or rephrased. In other words, even if the user who is the speaker does not understand the translation result of the target Chinese by confirming the reverse translation result in the original language, “chairs” is “lecture” Can be easily guessed. For this reason, in other words, there is a high possibility that high reliability, that is, communication failure can be reduced by paraphrasing this part.

  At this time, the reliability presented by the reliability presentation unit 17 does not need to be in units of sentences (or entire utterances), and may be presented in units of words, for example. This may make it easier to identify the paraphrase. In that case, the integration of the speech recognition reliability and the translation reliability in the reliability integration unit 21 may be first performed for each word and then expanded for each sentence.

  For example, in the example of FIG. 13D, if the reliability of the “chairs” or “do you have” part is presented, the speaker side user can confirm that there is a problem in the part, It is easy to take measures such as rephrasing the part with another word or expression.

  As described above, in this embodiment, in a speech translation system using speech as an input, it is possible to output a translation result with high reliability by appropriately integrating and presenting the reliability of speech recognition and machine translation. . Further, it is possible for the user to clarify the necessity of re-input and the re-input part. As a result, communication failures can be reduced.

  Next, the outline of the present invention will be described. FIG. 14 is a block diagram showing an outline of the present invention. As shown in FIG. 14, the reliability calculation apparatus according to the present invention includes a translation unit 501, a reverse translation unit 502, a translation reliability calculation unit 503, and a reliability presentation unit 504.

  The translation unit 501 (for example, the machine translation unit 13) performs a translation process on the input sentence into a second language that is different from the first language that is the language of the input sentence.

  The reverse translation unit 502 (for example, the reverse translation unit 14) performs reverse translation processing on the translation result obtained as a result of translation processing by the translation unit 501 into the first language that is the language of the input sentence.

  The translation reliability calculation unit 503 (for example, the translation reliability calculation unit 15) calculates the translation reliability based on the distance between the input sentence and the reverse translation result obtained as a result of the reverse translation processing by the reverse translation unit 502.

  The reliability presentation unit 504 (for example, the reliability presentation unit 17 and the output unit 12) corresponds to at least one of the input sentence, the translation result, and the reverse translation result to the input sentence, the translation result, or the reverse translation result. Present to the user along with the reliability. At that time, if the translation reliability is equal to or less than a predetermined threshold, at least the input sentence or the reverse translation result is presented to the user who has input the first language together with the translation reliability.

  The “reliability corresponding to the (presented) word or word string” refers to, for example, when the original text or the reverse translation result is presented, the word or word string is compared with the original text and the reverse translation result. It is the integrated reliability that combines the calculated translation reliability and the translation reliability and the recognition reliability. Further, for example, when presenting a translation result, the translation reliability or the integrated reliability calculated for the original sentence word or word string that was used when the word or word string was translated.

  In addition, the reliability calculation apparatus according to the present invention converts a text, which is a result of performing recognition processing on a sentence or sentence data input by a method other than text input such as voice input, as an input sentence to the translation unit 501. In the case of the reliability calculation device, the integrated reliability calculation means for calculating the integrated reliability using the reliability of the recognition process for the input sentence and the translation reliability calculated by the translation reliability calculation means 503. (For example, the reliability integration unit 21) may be provided. In such a configuration, when the integrated reliability is equal to or lower than a predetermined threshold, the reliability presentation unit 504 outputs at least the input sentence or the reverse translation result together with the integrated reliability or the translation reliability as the first language. You may show to the user of the input side.

  Moreover, the translation reliability calculation means 503 may obtain | require the distance of an input sentence and the reverse translation result by a reverse translation means using shared N-gram.

  Moreover, the translation reliability calculation means 503 may obtain the translation reliability in units of words or word strings and may obtain the translation reliability in units of sentences by normalization. In such a configuration, when the integrated reliability or the translation reliability is equal to or lower than a predetermined threshold, the reliability presenting means 504 outputs at least the input sentence or the reverse translation result for each of the input sentence or the reverse translation result. You may show to the user who input the 1st language with the reliability corresponding to a word or a word string. Further, the reliability degree presentation unit 504 corresponds to each word or word string of the input sentence, the translation result, or the reverse translation result when presenting at least one of the input sentence, the translation result, and the reverse translation result to the user. The low confidence interval in one sentence may be emphasized and presented based on the confidence level.

  The reliability calculation device according to the present invention further includes synonym expansion means (for example, synonym expansion unit 161 and synonym dictionary 162) that performs synonym expansion on at least one of the input sentence and the reverse translation result. You may have. In such a configuration, the translation reliability calculation unit 503 may calculate the distance for each combination of the input sentence after the synonym expansion and the reverse translation result, and calculate the translation reliability based on the calculated distance. Further, the synonym expansion means may perform synonym expansion only for words or word strings that are commonly held in the recognition dictionary used in the recognition process and the translation dictionary used in the translation process and the reverse translation process. .

  Although part or all of said embodiment and its modification example may be described also as the following additional remarks, it is not necessarily limited to the following.

(Supplementary note 1) For the input sentence, a translation unit that performs translation processing into a second language that is different from the first language that is the language of the input sentence, and a translation result obtained as a result of the translation process On the other hand, based on the distance between the reverse translation unit that performs the reverse translation process to the first language that is the language of the input sentence, and the reverse translation result obtained as a result of the reverse translation process, the translation reliability And a reliability of presenting at least one or more of the input sentence, the translation result and the reverse translation result together with a reliability corresponding to the input sentence, the translation result or the reverse translation result to the user. A presentation unit, and when the translation reliability is less than or equal to a predetermined threshold, the reliability presentation unit inputs at least the input sentence or the reverse translation result together with the translation reliability to the first language. Presented to users of Yoriyukido calculation device.

(Supplementary Note 2) A reliability calculation device that uses a text that is a result of recognition processing for sentence or sentence data in a first language input by a method other than text input, as an input sentence to the translation unit. Using the reliability of the recognition process for the input sentence and the translation reliability calculated by the translation reliability calculation unit, the integrated reliability calculation unit for calculating the integrated reliability, the reliability presentation unit is When the integrated reliability is equal to or lower than a predetermined threshold, at least the input sentence or the reverse translation result is presented to the user who has input the first language together with the integrated reliability or the translation reliability. The reliability calculation apparatus described in 1.

(Additional remark 3) The reliability calculation apparatus of Additional remark 1 or Additional remark 2 which a translation reliability calculation part calculates | requires the distance of an input sentence and the reverse translation result by a reverse translation part using shared N-gram.

(Supplementary Note 4) The translation reliability calculation unit obtains the translation reliability in units of words or word strings and obtains the translation reliability in units of sentences by normalization, and the reliability presentation unit has the integrated reliability or the translation reliability. If it is less than or equal to a predetermined threshold, at least the input sentence or reverse translation result is sent to the user who has input the first language together with the reliability corresponding to each word or word string of the input sentence or reverse translation result. The reliability calculation apparatus according to any one of Supplementary Note 1 to Supplementary Note 3 to be presented.

(Additional remark 5) When a reliability presentation part shows a 1st language to a user at least 1 or more among an input sentence, a translation result, and a back translation result, each of the said input sentence, a translation result, or a back translation result The reliability calculation apparatus according to appendix 4, wherein the low reliability interval in one sentence is emphasized and presented based on the reliability corresponding to the word or the word string.

(Supplementary Note 6) A synonym expansion unit that performs synonym expansion on at least one of the input sentence and the reverse translation result is provided, and the translation reliability calculation unit is configured to display the input sentence after the synonym expansion and the reverse translation result The reliability calculation apparatus according to any one of appendix 1 to appendix 5, which calculates a distance for each combination and calculates a translation reliability based on the calculated distance.

(Additional remark 7) It is the reliability calculation apparatus which makes the text which is the result of having performed the recognition process with respect to the sentence input by methods other than text input as an input sentence to a translation part, Comprising: The reliability calculation apparatus according to appendix 6, wherein synonym expansion is performed only for words or word strings that are commonly held in a recognition dictionary used for recognition processing and a translation dictionary used for translation processing and reverse translation processing.

  This application claims the priority on the basis of the JP Patent application 2011-165708 for which it applied on July 28, 2011, and takes in those the indications of all here.

  Although the present invention has been described with reference to the embodiments, the present invention is not limited to the above-described embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

Industrial applicability

  The present invention is not limited to a translation device that performs a translation process on a sentence or sentence input in a text data format, but also a speech input or a sentence or sentence data input using an OCR or a touch panel, and a recognition process. The present invention can also be suitably applied to a translation apparatus that performs translation processing after passing through.

10, 10b, 10c, 10d, 10e, 10f Reliability calculation device, reliability calculation unit 11 input unit 12 output unit 13 machine translation unit 14 reverse translation unit 15 translation reliability calculation unit 161 synonym expansion unit 162 synonym dictionary 17 Reliability presentation unit 100 machine translation device 110, 220 result determination unit 200 speech translation system 210 speech recognition result input unit 211 speech input unit 212 speech recognition unit 21 reliability integration unit 501 translation unit 502 reverse translation unit 503 translation reliability calculation unit 504 Reliability presentation means

Claims (9)

Translation means for performing a translation process on the input sentence to a second language that is different from the first language that is the language of the input sentence;
Reverse translation means for performing reverse translation processing to the first language that is the language of the input sentence, with respect to the translation result obtained as a result of the translation processing;
A translation reliability calculation means for calculating a translation reliability based on the distance between the input sentence and the reverse translation result obtained as a result of the reverse translation process;
A reliability degree presentation unit that presents at least one or more of the input sentence, the translation result, and the reverse translation result together with a reliability corresponding to the input sentence, the translation result, or the reverse translation result;
When the translation reliability is equal to or lower than a predetermined threshold, the reliability presentation means presents at least an input sentence or a reverse translation result to the user who has input the first language together with the translation reliability. The reliability calculation apparatus characterized by this. A reliability calculation device that uses a text that is a result of recognition processing for a sentence or sentence data in a first language that is input by a method other than text input, as an input sentence to the translation means,
Using the reliability of the recognition process for the input sentence and the translation reliability calculated by the translation reliability calculation means, comprising integrated reliability calculation means for calculating the integrated reliability;
When the integrated reliability is equal to or less than a predetermined threshold, the reliability presenting means provides at least an input sentence or a reverse translation result as a user who has input the first language together with the integrated reliability or the translation reliability. The reliability calculation device according to claim 1. The reliability calculation device according to claim 1, wherein the translation reliability calculation unit obtains a distance between the input sentence and the reverse translation result by the reverse translation unit using a shared N-gram. The translation reliability calculation means calculates the translation reliability in units of words or word strings, and calculates the translation reliability in units of sentences by normalization,
When the integrated reliability or the translation reliability is equal to or less than a predetermined threshold, the reliability presenting means corresponds at least the input sentence or the reverse translation result to each word or word string of the input sentence or the reverse translation result. The reliability calculation apparatus according to any one of claims 1 to 3, wherein the reliability is presented to a user who inputs the first language together with the reliability. When presenting at least one or more of the input sentence, the translation result, and the reverse translation result to the user in the first language, the reliability degree presentation means each word or word string of the input sentence, the translation result, or the reverse translation result. The reliability calculation apparatus according to claim 4, wherein a low reliability section in one sentence is emphasized and presented based on the reliability corresponding to. Synonym expansion means for performing synonym expansion on at least one of the input sentence and the reverse translation result,
The translation reliability calculation means calculates a distance for each combination of the input sentence after the synonym expansion and the reverse translation result, and calculates the translation reliability based on the calculated distance. The reliability calculation apparatus according to item 1. A reliability calculation apparatus that uses a text that is a result of recognition processing for a sentence input by a method other than text input as an input sentence to the translation means,
The synonym expansion means expands synonyms only for words or word strings that are commonly held in a recognition dictionary used for recognition processing and a translation dictionary used for translation processing and reverse translation processing. Reliability calculation device. The input sentence is translated into a second language that is different from the first language that is the language of the input sentence,
For the translation result obtained as a result of the translation process, perform a reverse translation process to the first language that is the language of the input sentence,
Based on the distance between the input sentence and the reverse translation result obtained as a result of the reverse translation process, the translation reliability is calculated,
When presenting to the user at least one or more of the input sentence, translation result and reverse translation result together with the reliability corresponding to the input sentence, translation result or reverse translation result, the translation reliability is below a predetermined threshold A translation reliability calculation and utilization method, wherein at least the input sentence or the reverse translation result is presented to the user who has input the first language together with the translation reliability. On the computer,
A translation process for converting an input sentence written in the first language into a second language, which is a language different from the first language,
A reverse translation process for converting a translation result obtained as a result of the translation process into a first language that is a language of an input sentence;
A translation reliability calculation process for calculating a translation reliability based on a distance between the input sentence and a reverse translation result obtained as a result of the reverse translation process; and at least one of the input sentence, the translation result, and the reverse translation result The above process is executed to provide the user with the reliability corresponding to the input sentence, the translation result, or the reliability corresponding to the reverse translation result,
In the reliability presentation process, when the translation reliability is equal to or lower than a predetermined threshold, at least an input sentence or a reverse translation result is presented to a user who inputs the first language together with the translation reliability. Translation engine program.

Images