JP2021140281A - Translation assistance system and program - Google Patents

Abstract

To provide a translation assistance system and a program which can easily assist in translation by a combination of field-specific machine translation means.SOLUTION: The translation assistance system for assisting in work of translating a document in a first language to a document in a second language uses a plurality of machine translation engines each of which can translate sentences in the first language to sentences in the second language by machine translation, to divide the document in the first language into prescribed segment units and generates translation results of the plurality of machine translation engines with respect to each of the segments obtained by the division. The translation assistance system evaluates each of the generated translation results from the plurality of machine translation engines and selects and outputs one to be employed of the translation results on the basis of evaluation results with respect to each of the segments.SELECTED DRAWING: Figure 1

Description

The present invention relates to a translation support system and a program.

Conventionally, a technique for supporting translation by combining a translation support system and machine translation has been known. For example, Patent Document 1 discloses a translation support system for supporting machine translation from a source language sentence to a target language sentence. The system disclosed in Patent Document 1 includes an error database for storing words or word combinations contained in a plurality of original language sentences in which machine translation from the original language sentence to the target language sentence is not performed correctly, and a plurality of systems. It is provided with a controlled original language sentence database for storing the original language sentence of the above and the controlled original language sentence which is a controlled original language sentence expressed in a format corresponding to the plurality of original language sentences and satisfying a predetermined condition. Then, in this system, it is determined whether or not the input original language sentence is machine-translatable, and the input original language sentence determined that the input original language sentence is not machine-translatable is output.

Japanese Unexamined Patent Publication No. 2018-169715

However, in the above-mentioned conventional technique, it is necessary to individually construct an error database and a controlled original language sentence database for each field of the original language sentence. Therefore, when supporting translation for various technical fields such as translation of patent application specifications, it is troublesome and difficult to construct these databases.

Further, as for the means of machine translation (MT), while MT specialized in a field such as a patent-only MT can easily achieve high accuracy, it is necessary to specify an MT engine according to the field of the original language sentence at the time of translation. However, the original language sentences to be translated do not always consist of sentences in a single field. Therefore, when a translator uses a translation support system to start translation, if he / she first tries to translate all at once, he / she must specify the MT engine every time the field changes (for each part of the original language sentence). Although the benefits of the above-mentioned high precision cannot be obtained, it is also difficult to specify the MT engine for each part.

The present invention has been made in view of the above circumstances, and one of its purposes is to provide a translation support system and a program capable of easily performing translation support by combining field-specific machine translation means. do.

One aspect of the present invention that solves the problems of the above conventional examples is a translation support system that supports the work of translating a document in a first language into a document in a second language, each of which is of the first language. A plurality of machine translation means capable of machine translating a sentence into a sentence in a second language, a dividing means for dividing a document in the first language into predetermined segment units, and each segment obtained by the division. A means for generating translation results by the plurality of machine translation means, an evaluation means for evaluating each of the translation results by the plurality of machine translation means, and the plurality of machine translations for each of the segments based on the evaluation results. It includes a selection means for selecting one translation result to be adopted from the translation results by the means, and an output means for outputting the translation result selected for each segment.

According to the present invention, it is possible to easily perform translation support by combining field-specific machine translation means.

It is a block diagram which shows the structural example of the translation support system which concerns on embodiment of this invention. It is a functional block diagram which shows the example of the translation support system which concerns on embodiment of this invention. It is explanatory drawing which shows the example of the information output by the translation support system which concerns on embodiment of this invention. It is a flow chart which shows the example of the operation flow of the translation support system which concerns on embodiment of this invention.

Embodiments of the present invention will be described with reference to the drawings. As illustrated in FIG. 1, the translation support system 1 according to the embodiment of the present invention is realized by a general computer and includes at least a control unit 11, a storage unit 12, an operation unit 13, and a display unit 14.

The control unit 11 is a program control device such as a CPU, and operates according to a program stored in the storage unit 12. In an example of this embodiment, the control unit 11 accepts the input of a sentence in the first language to be translated. The control unit 11 divides the input first language sentence into predetermined segment units. The control unit 11 executes a plurality of machine translation program modules, and by executing each of the plurality of machine translation program modules, the input first language sentence is divided into the segments obtained. Machine translate each unit into a sentence in a second language. Here, by executing the plurality of machine translation program modules, the control unit 11 can translate each segment of the sentence of the first language into the corresponding segment of the sentence of the second language. And.

Each machine translation program module uses different dictionaries, analysis rules, machine learning results, etc. to transform a common first language sentence segment into a second language sentence segment that is generally different from each other. It will be translated.

The control unit 11 evaluates the segment (translation result) of the sentence of the corresponding second language obtained by the processing of each machine translation program module for each segment obtained by the division. Based on the result of this evaluation, the control unit 11 selects one of the corresponding second language sentence segments obtained by the processing of each machine translation program module for each segment obtained by the above division. Then, the selected segment is output. The details of the processing of the control unit 11 will be described later.

The storage unit 12 is a storage device such as a memory or a disk device, and holds a program (including a machine translation program) executed by the control unit 11. The storage unit 12 also operates as a work memory for the control unit 11. The program stored in the storage unit 12 may be provided by being stored in a computer-readable and non-temporary recording medium, and may be copied in the storage unit 12.

The operation unit 13 includes a keyboard, a mouse, and the like, accepts a user's operation, and outputs information representing the content of the operation to the control unit 11. The display unit 14 is a display or the like, and displays and outputs information according to an instruction input from the control unit 11.

Next, the operation of the control unit 11 will be described. In the present embodiment, as illustrated in FIG. 2, the control unit 11 includes the original text receiving unit 21, the preprocessing unit 22, the segment dividing unit 23, the machine translation processing unit 24, and the translation evaluation unit 25. , A selection unit 26, a post-processing unit 27, and an output unit 28 are functionally included.

The original text receiving unit 21 accepts the input of the document in the first language to be translated and outputs it to the preprocessing unit 22 and the segment dividing unit 23. Specifically, the source text receiving unit 21 receives a document in the first language from another computer on the network, for example, via a network interface (not shown).

The preprocessing unit 22 executes a predetermined preprocessing on the document in the first language accepted by the original text receiving unit 21. This preprocessing may be predetermined for each type of document accepted by the original text receiving unit 21. As an example, preprocessing in translation of claims in a patent application document can be considered. In general, one claim can be divided into a plurality of segments by the segment division unit 23 described later. However, since the claims are traditionally described in one sentence, the following problems occur. Here specifically

A means to A and
B means to B and
X with.
As described above, an operation example of the preprocessing unit 22 will be described by taking as an example a claim including a component unit such as "A means A" and a subject unit such as "X having ...". If the claims in this example are translated into English,

X comprising: X comprising:
A means for doing A, and
B means for doing B.
It is expected that the translation will be output as follows.

However, each component part and the subject part are divided by the segment division part 23 described later, and the information of other segments is not referred to at the time of machine translation of each segment. If you perform the translation process as it is and concatenate without doing it, even if each segment is translated correctly,

A means for doing A,
B means for doing B,
X? ..
(The last "?" Indicates that the corresponding word cannot be found)
It will be translated as.

Therefore, in the preprocessing unit 22, the subject unit is transposed after the translation in response to the translation of the claim, and the subject unit is input from the original text receiving unit 21 so that the subject unit is placed at the beginning or after the translation. Regarding the sentence related to the claim in the document, the subject part and the component part are recognized respectively. Specifically, the preprocessing unit 22 uses words such as “comprise” and “include” included in the subject part as clues (such words are hereinafter referred to as clues). , The subject part and the subject phrase (“X” in the above example) included in the subject part are specified.

Further, the preprocessing unit 22 extracts elements other than the subject unit and which are divided by ", (comma)", "; (semicolon)", etc., as component units.

Then, the preprocessing unit 22 records each part of the claim sentence in association with range identification information indicating whether the part is a “subject part” or a “component part”. This range identification information is used later in the post-processing unit 27.

The segment division unit 23 divides the document in the first language accepted by the original text reception unit 21 into segments Si (i = 1, 2, ...) According to a predetermined rule. As an example, the segment division unit 23 detects a predetermined segment break phrase, symbol, or the like, and divides the document before and after the detected position to divide the document into segments Si (i = 1, 2, ...). To get. Here, as a segment delimiter, a symbol, etc.,
・ Period (end of sentence)
・ Colon ・ Line feed ・ "," that appears after ", and"
・ "," Appearing after ", or"
And so on.

So, for example, the original example:
A liquid container for supplying a liquid to a liquid ejection apparatus, the liquid container comprising: [LF]
a cubic bag in which a liquid storage portion for containing the liquid is provided, [LF]
a liquid outlet for leading out the liquid in the bag to the liquid ejection apparatus, [LF]
When the three dimensional cartesian coordinates is defined with a X-direction, a Y-direction, and a Z-direction, the X direction being defined as a direction along a direction from one edge side of the bag toward the other edge side, [LF]
a spacer member is arranged at a position of the sides of the bag, where the sides are parallel to the YZ plane. [LF]
The segment division unit 23 divides the segment into segments as follows. In the above example, the line feed position is set to "[LF]" in order to clearly indicate the line feed.

・ Segment 1: A liquid container for supplying a liquid to a liquid ejection apparatus, the liquid container comprising:
・ Segment 2: a cubic bag in which a liquid storage portion for containing the liquid is provided,
・ Segment 3: a liquid outlet for leading out the liquid in the bag to the liquid ejection apparatus,
・ Segment 4: where in when the three dimensional cartesian coordinates is defined with a X-direction, a Y-direction, and a Z-direction,
・ Segment 5: the X direction being defined as a direction along a direction from one edge side of the bag toward the other edge side,
・ Segment 6: a spacer member is arranged at a position of the sides of the bag, where the sides are parallel to the YZ plane.

Further, the segment division unit 23 refers to the range identification information generated by the preprocessing unit 22, and the identification information for identifying each segment (a serial number for each segment may be used) and the segment identified by the identification information are obtained. , Information indicating whether or not the subject part is included and whether or not the component part is included is recorded in association with each other.

Specifically, the segment division unit 23 has the subject unit as the next segment of the segment including the number representing the item number of the claim. Further, the segment division unit 23 is a segment immediately before a segment containing "where in" at the beginning of a predetermined word, for example, if the first language is English, from the segment next to the segment (referred to as the subject segment) as the subject portion. Up to is a segment of the component part (component segment), and the segment after wherein is a segment of the component explanation part. The segment division unit 23 records the identification information that identifies each segment in association with information indicating whether each segment is a subject segment, a component segment, or a component description segment.

Further, when the end of the segment matches a predetermined pattern, the segment division unit 23 may replace the portion matching the pattern with a predetermined character string corresponding to the pattern. For example, the trailing ", (comma)" or ", and" may be replaced with ". (Period)". This process causes the machine translation engine to translate the segment as if it were a single sentence.

The segment division unit 23 sequentially outputs the segments obtained by dividing the document in the first language accepted by the original text receiving unit 21 to the machine translation processing unit 24.

The machine translation processing unit 24 translates each of the segments Si (i = 1, 2, ...) Input from the segment division unit 23 by a plurality of translation engines Ej (j = 1, 2, ...), And translates each segment. The translation result Tj (Si) into the second language by the engine Ej is obtained.

Here, it is assumed that the machine translation program modules of the translation engine Ej used by the machine translation processing unit 24 use different dictionaries, and use different translation methods such as analysis rules and machine learning results. Therefore, a segment of a sentence in a first language is generally translated into a segment of a sentence in a second language that is different from each other.

As an example, the translation engine Ej here is
-A general-purpose neural translator (general-purpose NT) that is machine-learned using Japanese-English bilingual sentences (sentence pairs) extracted from patent documents and general documents as training data.
-A neural translator for patent documents (Patent NT) that machine-learned Japanese-English sentence pairs extracted from patent documents as training data, and
-Based on general-purpose NT or patent NT, an adaptation engine (AT) that additionally machine-learns Japanese-English sentence pairs extracted from translated sentences in a specific field, and
Suppose that

In this example, the three machine translation program modules, general-purpose NT, patent NT, and AT, output the results of translating each of the above segments.

The translation evaluation unit 25 accepts each translation result Tj (Si) for each segment Si of the plurality of translation engines Ej included in the machine translation processing unit 24, and evaluates each of the translation results related to the segment Si by a predetermined method.

As an example, this evaluation is performed as follows. The translation evaluation unit 25 reverse-translates each of the translation results Tj (Si) related to the segment Si into a sentence in the first language to obtain an evaluation sentence Si'. Here, the machine translation engine used by the translation evaluation unit 25 for reverse translation uses a machine translation engine (a translation engine that has been machine-learned using the same learning data as at the time of translation) that outputs the translation result to be reverse-translated. And. That is, in the above example, the sentence in the second language, which is the translation result translated by the general-purpose NT, is also back-translated by the general-purpose NT to become the sentence in the first language. In addition, the sentence in the second language, which is the translation result translated by the patent NT, is also reverse-translated by the patent NT into the sentence in the first language, and so on. I will do it.

The translation evaluation unit 25 compares the segment Si before translation with the evaluation sentence Si'obtained by reverse translation of the translation result, and BLEU (BLEU), which is an evaluation scale for machine translation, is used for each translation result Tj (Si). Bilingual Evaluation Understudy), RIBES (Rank-based Intuitive Bilingual Evaluation Score), or both evaluation values are calculated. When both evaluation values are calculated, the square root of the average of the squares and the like may be further calculated to obtain one evaluation value.

The selection unit 26 refers to the evaluation value obtained for each translation result Tj (Si) related to the segment Si, and selects the translation result Tj (Si) having the highest evaluation value for each segment Si.

The post-processing unit 27 executes a predetermined post-processing on the translation result selected by the selection unit 26. Specifically, the post-processing unit 27 refers to the information recorded by the pre-processing unit 22 in association with the segment identification information, and immediately before the translation result selected for the segment that is supposed to include the subject unit. , Moves to the front of the translation result for the segment that is supposed to contain the component part.

Further, the post-processing unit 27 changes the end representation of the translation result of at least one segment according to a predetermined rule. This rule may be set for each field of the document to be translated. For example, in translating a patent document, the post-processing unit 27 includes the particle “to,” indicating parallelism for the segment including the component unit among the segments obtained from the sentence according to one “claim”. .. Further, the post-processing unit 27 postfixes the word of the subject part extracted from the sentence of the claim to the translation result of the last segment among the segments obtained from the sentence of one "claim". do.

As a specific example, the translation result selected by the selection unit 26 for each of the segments obtained from the above original text example is
Segment 1: A liquid container for supplying a liquid to a liquid injector, which comprises:
-Segment 2: A cube-shaped bag provided with a liquid storage portion for storing the liquid inside.
-Segment 3: A liquid lead-out unit for leading the liquid in the bag to the liquid injection device.
-Segment 4: When a three-dimensional coordinate system is defined in the X, Y, and Z directions,
Segment 5: The X direction is defined as a direction along the direction from one end to the other end of the bag.
Segment 6: A spacer member arranged on a side surface portion of the bag whose side surface is parallel to the YZ plane.
When, the post-processing unit 27 uses the word "liquid container" (and the transition word "prepared") of the subject unit.
-Segment 1: A liquid container "is" for supplying a liquid to a liquid injector.
-Segment 2: A cube-shaped bag "to," provided with a liquid storage portion for storing the liquid inside.
-Segment 3: A liquid lead-out unit "with," for leading out the liquid in the bag to the liquid injection device.
-Segment 4: When a three-dimensional coordinate system is defined in the X, Y, and Z directions,
Segment 5: The X direction is defined as a direction along the direction from one end to the other end of the bag.
Segment 6: A liquid container having a spacer member "arranged on a side surface portion of the bag and whose side surface is parallel to the YZ plane."
Replace and change the words in each segment. In addition, although the changed part is specified by adding the key brackets ("") here, such specification is not always necessary in reality.

The output unit 28 outputs the translation result of each segment post-processed by the post-processing unit 27. In an example of the present embodiment, the output unit 28 may output the original text and the translation result in a format of XLIFF (XML Localization Interchange File Format) so as to be comparable.

Specifically, the output unit 28 inserts the original text of the segment into the source (<source>) tag portion of the XLIFF for each segment obtained from the original text, and post-processes the target (<target>) tag portion. The translation result of the corresponding segment post-processed by the unit 27 is inserted and output.

The output in the XLIFF format thus obtained can be displayed and referenced using a general XLIFF viewer.

As an example, the output unit 28 includes the original text and the translation result for each segment in the source tag and the target tag, respectively, in the XLIFF format file as described above, and the evaluation value for each segment (translation selected by the selection unit 26). (Evaluation value related to the result) may be recorded in association with each other and used for display processing.

Further, the output unit 28 may include the result of the reverse translation (sentence in the first language) generated by the translation evaluation unit 25 in the XLIFF format file in association with the translation result of each segment and perform the display processing. .. As shown in FIG. 3, the display of the XLIFF format file in this example shows the original text (L), the translation result (T), the evaluation value (V), and the reverse translation result (B) for each segment, as illustrated in FIG. And will be included. In the example of FIG. 3, an example of displaying the result of reverse translation and the original text in comparison with each other for the selected segment is shown. At this time, the information that identifies the translation engine that outputs the adopted translation result may also be output.

[motion]
The translation support system 1 of the present embodiment basically has the above configuration and operates as follows. In an example of this embodiment, the translation support system 1 accepts an input of the original text in the XLIFF format in which the document of the first language which is the original text is described in the source tag.

The translation support system 1 accepts data that is the original text of the translation, as illustrated in FIG. 4 (S1). Specifically, this data may be PDF (Portable Document Format) data including text. In the example below, the first language is English and the second language targeted for translation is Japanese.

The translation support system 1 transmits this PDF data to an external system to generate XLIFF format data. In this example, the text included in the PDF is extracted from the generated XLIFF format data and included in the source tag.

The translation support system 1 extracts the text of the source tag from the generated XLIFF format data. The text extracted here is that of the patent document as in the above-mentioned original text example, and one of the claims is

A liquid container for supplying a liquid to a liquid ejection apparatus, the liquid container comprising: [LF]
a cubic bag in which a liquid storage portion for containing the liquid is provided, [LF]
a liquid outlet for leading out the liquid in the bag to the liquid ejection apparatus, [LF]
When the three dimensional cartesian coordinates is defined with a X-direction, a Y-direction, and a Z-direction, the X direction being defined as a direction along a direction from one edge side of the bag toward the other edge side, [LF]
a spacer member is arranged at a position of the sides of the bag, where the sides are parallel to the YZ plane. [LF]
And so on.

The translation support system 1 executes a predetermined preprocessing on this text. Specifically, assuming that the noun "the liquid container" immediately before the clue phrase "comprising" is the subject word, the part containing this subject word (the part separated by a line break (LF),: (colon), etc., In the above example, the part of "A liquid container for supplying a liquid to a liquid ejection apparatus, the liquid container comprising") is recognized as the subject part. Further, the translation support system 1 extracts and stores "comprising" as a transition word.

In addition, from the clauses other than the subject part, the clauses to be the component parts are specified according to a predetermined rule. Here, it is assumed that a part other than the above-mentioned subject part and separated by a line feed is recognized as a component part.

Further, the translation support system 1 divides the text into segments according to a predetermined rule (S2). Here, for example, a line feed and a segment are divided at the position of ", (comma)" next to ", and", and a predetermined pattern at the end of the sentence is replaced.
・ Segment 1: A liquid container for supplying a liquid to a liquid ejection apparatus, the liquid container comprising:
・ Segment 2: a cubic bag in which a liquid storage portion for containing the liquid is provided.
・ Segment 3: a liquid outlet for leading out the liquid in the bag to the liquid ejection apparatus.
・ Segment 4: where in when the three dimensional cartesian coordinates is defined with a X-direction, a Y-direction, and a Z-direction.
・ Segment 5: the X direction being defined as a direction along a direction from one edge side of the bag toward the other edge side.
・ Segment 6: a spacer member is arranged at a position of the sides of the bag, where the sides are parallel to the YZ plane.
And, it is assumed that it is divided. Here, the translation support system 1 specifies the segment 1 as the "subject part" and the remaining segments 2 to 6 as the "component part".

The translation support system 1 machine-translates each segment obtained here by a plurality of translation engines (S3). The translation engines here are, respectively.
-A general-purpose neural translator (general-purpose NT) that is machine-learned using Japanese-English bilingual sentences (sentence pairs) extracted from patent documents and general documents as training data.
-A neural translator for patent documents (Patent NT) that machine-learned Japanese-English sentence pairs extracted from patent documents as training data, and
-Based on the patent NT, an adaptation engine (AT) that additionally machine-learns Japanese-English sentence pairs extracted from translated sentences in a specific field, and
Suppose that

The translation support system 1 uses, for example, a translation by a general-purpose NT for segment 1.
"Equipped with a liquid container to supply the device that ejects the liquid:",
Translated text by Patent NT, which was machine-learned using Japanese-English sentence pairs extracted from patent documents as training data:
"A liquid container for supplying a liquid to a liquid injector, which comprises:"
Based on general-purpose NT, translated sentences by AT that additionally machine-learned Japanese-English sentence pairs extracted from translated sentences in a specific field,
"A liquid container for supplying a liquid to a liquid injector, which comprises:"
To get.

The translation support system 1 reverse-translates each of the obtained translation results into a sentence (evaluation sentence) in the first language using the machine translation engine that generated the corresponding translation result. For example, the above translation by general-purpose NT is reverse-translated by general-purpose NT,
Get the evaluation statement "A container of liquid which supplies the liquid to a liquid ejector?".

Similarly, the translated text by Patent NT is reverse-translated by Patent NT,
The translation "A container of liquid for supplying the liquid to a liquid ejection apparatus, comprising:" and the translation by AT are reverse-translated by AT.
Suppose you get the translation "A liquid container for supplying a liquid to a liquid ejection apparatus, comprising:".

The translation support system 1 obtains the evaluation values of BREU and RIBES by comparing these evaluation sentences with the corresponding original sentences (S4). The evaluation values are all normalized so that the maximum value is, for example, "100" (for example, when the maximum value of the evaluation value of BREU is "1", the value is multiplied by 100).

を求める。 Translation support system 1, the evaluation value of Breu (those that have been normalized) and S _B, the evaluation value of Ribes (also those that have been normalized) by using the S _R, one evaluation value S

To ask.

Here, as an example, it is assumed that the evaluation values of 90, 91, and 95 are obtained for each of the translation results of the general-purpose NT, the patent NT, and the AT.

The translation support system 1 refers to this evaluation value for the segment 1 and selects the translation result of the AT having the highest evaluation value for the segment 1 (S5).

Hereinafter, the same processing is performed for each segment, and the translation result having the highest evaluation value is selected from the plurality of translation results obtained by the plurality of machine translation engines for each segment. If there are a plurality of translation results with the highest evaluation values, one of the translation results is selected according to a predetermined rule. For example, a priority may be set for each translation engine, and a translation result by the translation engine having the highest priority may be selected from a plurality of translation results having the highest evaluation value.

Then, the translation support system 1 has a noun + "and" when the sentence end of the translation result selected for each segment of the segment, which is considered to be a component part, ends with "noun + punctuation mark". , ”(Parallel particle) is executed (one of the post-processing).

As a result, for example, the translation result selected for segment 2 is
"A cubic-shaped bag provided with a liquid storage portion for storing the liquid inside."
If, the end is "a cubic-shaped bag provided with a liquid storage portion for storing the liquid inside."
Is replaced as.

Further, in the translation support system 1, when the sentence end of the translation result selected for each segment of the segment, which is considered to be the component part, ends with "do" or "do." Replace with "" or "".

As a result, for example, the translation result of the segment that is the component part can be obtained.
"The X direction is a direction from one edge side of the bag to the other edge side, and the direction from the one edge side to the other edge side is a positive direction, and the positive direction is defined as the positive direction. The direction opposite to the direction of is the negative direction. "
If, the translation support system 1
"The X direction is a direction from one edge side of the bag to the other edge side, and the direction from the one edge side to the other edge side is a positive direction, and the positive direction is defined as the positive direction. The direction opposite to the direction of is the negative direction. "
Replace with.

Further, the translation support system 1 moves the translated word "liquid container" of the theme word "the liquid container" in the subject part to the last part of the last segment of the segments of the component part.

In the above example, the last segment of the component portion is segment 6, and the selected translation result is, for example, "a spacer member arranged on a side surface portion of the bag and whose side surface is parallel to the YZ plane."
It is assumed that it is. When the word immediately before the last kuten of this segment is an uninflected word (terminated by a noun), the translation support system 1 replaces the kuten with a comma, and translates the transition word + the subject word. Add a punctuation mark
"A liquid container provided with a spacer member, which is arranged on a side surface portion of the bag and whose side surface is parallel to the YZ plane."
And.

If the last segment of the component part is not an uninflected word, the punctuation mark is replaced with a comma, and the translation of the subject word + the punctuation mark is added. That is, this segment 6 is, for example,
"A spacer member is provided which is arranged on the side surface portion of the bag and whose side surface is parallel to the YZ plane."
If it is,
"A liquid container provided with a spacer member arranged on a side surface portion of the bag and whose side surface is parallel to the YZ plane."
And.

Further, in the translation support system 1 of the present embodiment, a process may be performed in which the translated word corresponding to the fixed expression is replaced with the translated word corresponding to the predetermined fixed expression.

The translation support system 1 then inserts the original text of the segment into the source tag part and inserts the post-processed translation result of the corresponding segment into the target tag part for each segment obtained from the original text. Obtain the data in the format and output this data (S6).

The XLIFF format data after inserting the translation result may be sent to the user of the translation support system 1 by e-mail or the like.

[Processing related to related adverbs]
In the translation support system 1 of the present embodiment, when the relational adverbial phrase after the relational adverb straddles the segment, the segment of the relational adverbial phrase is recognized and the translation result is adjusted.

Specifically, citing the description of claim 1 of the US Publication of Patent Publication No. 2019184710A1, the description of claim 1 states.
When three directions orthogonal to each other are defined as a D direction, a T direction, and a W direction,
the D direction being defined as a direction along a direction from the one edge portion side of the bag toward the other edge portion side, where a direction from the one edge portion side toward the other edge portion side is defined as a + D direction and an opposite direction to the + D direction is defined as a −D direction, and
the T direction being defined as a direction in which a dimension of an outer shape of the liquid container is smallest among the three directions,
Is included.

When this sentence is input to the translation support system 1 of the present embodiment as the original sentence, the translation support system 1 divides this sentence into segments and divides the sentence into segments.
Segment 1: XXXXX when three directions orthogonal to each other are defined as a D direction, a T direction, and a W direction,
Segment 2: the D direction being defined as a direction along a direction from the one edge portion side of the bag toward the other edge portion side, where a direction from the one edge portion side toward the other edge portion side is defined as a + D direction and an opposite direction to the + D direction is defined as a −D direction, and
Segment 3: the T direction being defined as a direction in which a dimension of an outer shape of the liquid container is smallest among the three directions,
To get.

Here, as one of the preprocessing, a predetermined word such as wherein is replaced with a symbol string associated in advance (a symbol string output by the translation engine as it is, hereinafter referred to as a replacement symbol string). Also, here, the state before replacing "," and ", and" at the end of the segment is shown.

In the translation support system 1, when the above substitution symbol string is included in a certain segment and a relation adverb (where or when) is included following the substitution symbol string, the segment is used as a relation adverb segment and the relation adverb Refer to the end of the segment after the segment in sequence.

When the translation support system 1 finds a segment ending in ", and" or ", or", the translation support system 1 associates the segment up to the next segment with the segment identification information that identifies each segment as a relational adverbial phrase segment. In addition, information indicating that it is a relational adverb segment and a relational adverbial phrase segment is recorded in association with each other.

After that, the translation support system 1 replaces "," and ", and" at the end of the segment with "." And causes a plurality of translation engines to machine-translate each segment. The translation support system 1 selects one of the translation results obtained by a plurality of translation engines for each segment by the method already described.

Then, in addition to the post-processing already described, the translation support system 1 of this example is a segment following the segment when information indicating that it is a related adverb segment is recorded in association with the identification information of one of the segments of the translation result. There is a segment in which the identification information of the segment is associated with the information that it is an adverbial phrase segment, and the identification information of the next segment is not associated with the information that it is an adverbial phrase segment (relational adverbial phrase segment). The last segment of, hereinafter referred to as the terminal relation adverbial phrase segment) is specified.

The translation support system 1 identifies the translation of the relational adverb included in the translation result of the segment (called the relational adverb segment) identified by the identification information in which the information indicating that it is the relational adverb segment is recorded, and determines the translation. It is deleted from the related adverb segment and the last word is adjusted, and the translation support system 1 moves the deleted translated word to the end of the translation result of the terminal relation adverb phrase segment.

In the example above, the translation results selected for each segment are initially
Segment 1 (relational adverb segment): XXXXX When the three directions orthogonal to each other are the D direction, the T direction, and the W direction.
Segment 2 (related adverbial phrase segment): The D direction is a direction from the one edge side of the bag toward the other edge side, and from the one edge side to the other edge side. The direction to go is the + D direction, and the direction opposite to the + D direction is the −D direction.
Segment 3 (end-related adverb segment): Of the three directions, the direction in which the outer dimensions of the liquid container are minimized is defined as the T direction.
It has become.

The translation support system 1 performs the post-processing already described for the translation result obtained here, and also executes the process of moving the translated words of the above-mentioned related adverbs.
Segment 1 (relational adverb segment): Here, the three directions orthogonal to each other are defined as the D direction, the T direction, and the W direction.
Segment 2 (related adverbial phrase segment): The D direction is a direction from the one edge side of the bag toward the other edge side, and from the one edge side to the other edge side. The direction to go is the + D direction, and the direction opposite to the + D direction is the -D direction.
Segment 3 (end-related adverb segment): When the T direction is the direction in which the outer dimensions of the liquid container are minimized among the three directions.
To get.

The original "here" is a translation of wherein, replacing the symbol string XXXXX corresponding to wherein, the end of each segment is replaced with "to," and the end of the end-related adverb segment is here. It has been replaced with "when" which is a translation of the relative adverb "when".

As described above, a more natural translation can be obtained by recognizing the translation range of the relational adverb phrase, etc. corresponding to the relational adverb, etc. when dividing into segments and performing machine translation, and moving the translation of the relational adverb. Is possible.

On the other hand, as a comparative example, in the conventional machine translation system that does not execute the recognition, end processing, post-processing, etc. of the related adverbial phrase in the pre-processing and segment division processing performed by the translation support system 1 of the present embodiment, the above English sentence As a translation,
Segment 1: Here, when the three orthogonal directions are the D direction, the T direction, and the W direction, the segment 2: the D direction is a direction from the one edge side of the bag toward the other edge side. Is the + D direction, and the direction opposite to the + D direction is the −D direction. Segment 3: The T direction is the direction in which the outer dimension of the liquid container is the smallest among the three directions. Therefore, the recognition of the related adverbial phrase is wrong.

As yet another comparative example, when all the texts of claim 1 of the above-mentioned US Patent Publication No. 2019184710A1 are concatenated into one line and machine-translated without dividing them into segments, the result is that one sentence is too long and translation omission occurs. Become.

[When the first language is Japanese]
Further, the translation support system 1 of the present embodiment may have different contents of pre-processing, segment division and post-processing depending on the first language. For example, in the translation support system 1, when the first language is Japanese, it can be used as a segment delimiter, a symbol, or the like at the time of segment division.
・ Kuten (end-of-sentence symbol)
-Use predetermined ones such as colons and line breaks that correspond to Japanese (first language).

As an example, in the translation support system 1, when the first language is Japanese and the claims of the Japanese patent gazette or the Japanese application specification are translated, the subject part and the subject are as follows. Specify the word and component part. In this case as well, rules for identification are set in advance for each first language. For example, in Japanese
-If there is "... in" or "in ..." at the end of the first segment of the claim, the noun phrase immediately before it is extracted by morphological analysis or word-separation and used as the subject word.
And so on.

When the first language is Japanese, the translation support system 1 has an expression at the end of each segment (when the beginning is the subject part, the second and subsequent segments). , It is specified as a component part when the condition that it is a parallel particle "to," or "...," is satisfied. Further, the translation support system 1 specifies a segment after a segment containing a predetermined word corresponding to a transition word such as "... prepared" as a component explanation segment.

Further, in this example, the translation support system 1 identifies a clue phrase such as "equipped with ..." which is related to the subject part, and acquires a predetermined transition word. For example, if the second language is English, the English clue (transition word) corresponding to the subject part is acquired. Specifically, for "providing ...", "comprising:" is associated and stored in advance. The translation support system 1 may acquire and use the transition word associated with the specified clue phrase.

Further, the translation support system 1 may have different contents of pre-processing, segment division and post-processing on condition that not only the first language but also the second language is used. For example, when the first language is Japanese and the second language is English, expressions such as "when ..." and "when ..." are in the part before this expression (including the case where it straddles a segment). The translation result of the expression, such as "when", should appear before the translation result. Therefore, in the translation support system 1 of this example, whether or not there is an expression of "... when" or "... case" at the end of the segment of the component explanation part below "provides" corresponding to the transition word of the original text. Find out. Then, when these expressions are present, the translation support system 1 deletes the expressions and replaces the expressions at the beginning and end of the component explanation segment below "with" corresponding to the transition word with "... when,". Alternatively, replace it with "... case". As a result, when and where translations are rewritten so that they appear in the corresponding segment.

Further, in the translation support system 1 of this example, when the subject word is not obtained, for the final segment of each claim, "features ...", "contains ...", "has ...", "consists of ...", ", Predetermined subject clues such as "..." The following expressions are extracted and used as the subject words. Then, the translation support system 1 places the result of translating the subject word into English immediately before the transition word "comprising:" in the first segment.

[Table 1] shows the results of translating the patent documents in the field of chemistry using the translation support system 1 of the present embodiment using the general-purpose NT, the patent NT, and the AT.

In this [Table 1], the translation support system 1 evaluates each translation engine for each segment (No.), the evaluation value (highest evaluation value) of the translation engine selected for each segment, and the selection. It shows the translation engine that generated the translation results. There are 64 segments in total, but some are omitted for the convenience of charts. For reference, the last row of the table shows the average rating of each translation engine and the average of the highest ratings.

As shown in the results of [Table 1], for example, when a general-purpose NT or a patent NT is used alone, sufficient translation cannot be obtained in segment 4 or the like. Further, when only the AT having a relatively high evaluation value in the segment 4 is used, the translation result having a lower evaluation value than the patent NT in the segments 1, 8 and 9 and the like is obtained.

According to the translation support system 1 of the present embodiment, the translation result by the translation engine having a high evaluation value is selected for each segment, so that the average evaluation value is also improved.

[Summary of features]
Further, the features of the embodiment of the present invention include the following points.
(1) In the translation support system used when translating a document in the first language into a second language, it is in a state of being learned by a plurality of different training datasets, and the document in the first language and the document in the second language are in a state of being learned. The document is provided with a plurality of neural net machine translation means capable of mutually machine translating the document, the document of the first language is divided into segments, and the segment is translated by the plurality of neural net machine translation means. In addition, each of the translation results is translated into the first language by the corresponding neural net machine translation means, and the original text of the first language and the translation result obtained by translating from the second language to the first language. Calculate the similarity with. Then, from the plurality of similarities obtained by the calculation, the translation result of the second language corresponding to the translation result having the highest similarity is extracted for each segment. Then, the original text of the first language and the extracted translation result are displayed for each segment.

(2) When displaying the translation result in the XLIFF editor, the score of the extracted translation result, the original text, and the translation result from the second language to the first language are displayed together with the translation result.

(3) When English or Chinese is the first language and Japanese is the second language, the claim start part is the start segment and the claim end is the end for multiple source tags of the XLIFF file. As a segment, a segment to be a subject part, a segment of at least one component part, and a segment of a component explanation part (if any) are specified by a predetermined clue phrase, and an expression corresponding to the subject clue phrase of the subject part is specified. Place the Japanese expression corresponding to the end of the component segment. If the relational adverbial phrase consists of multiple segments, the last segment of the relational adverbial phrase is specified by the corresponding clue phrase, the relational adverb is moved to the beginning of the segment, and the subject word is placed in the final segment of the claim. Performs the process of arranging. This makes the translation of the claims natural.

(4) When Japanese is the first language and English or Chinese is the second language, the claim start part is the start segment and the claim end is the end segment for multiple source tags of the XLIFF file. As, the segment of the subject part, the segment of at least one component part, and the segment of the component explanation part (if any) are specified, and when the relation adverbial phrase consists of a plurality of segments, the relation is made by the corresponding clue phrase. After identifying the last segment of the adverbial phrase, move the expression corresponding to the related adverb to the corresponding segment, and move the subject word and the clue phrase between the component segment and the component description segment to the first segment of the claim. Place the expression together.

1 Translation support system, 11 Control unit, 12 Storage unit, 13 Operation unit, 14 Display unit, 21 Original text reception unit, 22 Preprocessing unit, 23 Segment division unit, 24 Machine translation processing unit, 25 Translation evaluation unit, 26 Selection unit , 27 Post-processing unit, 28 Output unit.

Claims (7)

A translation support system that supports the work of translating a document in the first language into a document in the second language.
Multiple machine translation methods, each capable of machine translating a sentence in the first language into a sentence in the second language,
A dividing means for dividing a document in the first language into predetermined segment units, and
A means for generating a translation result by the plurality of machine translation means for each segment obtained by the division, and a means for generating a translation result by the plurality of machine translation means.
An evaluation means for evaluating each of the translation results by the plurality of machine translation means,
For each of the segments, a selection means for selecting one translation result to be adopted from the translation results by the plurality of machine translation means based on the evaluation result.
A translation support system including an output means for outputting a translation result selected for each segment. The translation support system according to claim 1.
Each of the plurality of machine translation means can machine translate a sentence in the first language into a sentence in the second language, and can machine translate a sentence in the second language into a sentence in the first language. And
The evaluation means retranslates each of the translation results by the plurality of machine translation means into a segment of the first language by the machine translation means obtained the translation result, and the retranslation result and the first original text. A translation support system that compares a language segment and evaluates the translation result based on the result of the comparison. The translation support system according to claim 1 or 2.
The output means is a translation support system that outputs information indicating the result of evaluation of the translation result by the evaluation means together with the translation result. The translation support system according to any one of claims 1 to 3.
The dividing means further preprocesses the document in the first language based on a predetermined rule predetermined for each first language before and after the division into the segment units. A translation support system that runs. The translation support system according to claim 4.
When the document in the first language relates to the claims of the patent application document, the subject word is specified in the preprocessing, and the arrangement of the subject word is predetermined for each of the second language. A translation support system that includes the process of moving to a segment. The translation support system according to claim 4 or 5.
When the document in the first language relates to the claims of the patent application document and the first language is English, in the preprocessing, if there is a segment containing the relation adverb, the relation adverb Identify the scope of the phrase,
The output means is a translation support system that executes post-processing that moves a word corresponding to the translated word of the related adverb to the beginning of the segment corresponding to the range of the related adverb phrase in the translation result selected for each segment. A program to help a computer translate a document in a first language into a document in a second language.
Multiple machine translation methods, each capable of machine translating a sentence in the first language into a sentence in the second language,
A dividing means for dividing a document in the first language into predetermined segment units, and
A means for generating a translation result by the plurality of machine translation means for each segment obtained by the division, and a means for generating a translation result by the plurality of machine translation means.
An evaluation means for evaluating each of the translation results by the plurality of machine translation means,
For each of the segments, a selection means for selecting one translation result to be adopted from the translation results by the plurality of machine translation means based on the evaluation result.
An output means that outputs the translation result selected for each segment, and
A program that functions as.

Images