Translation support device and program

Abstract

To provide a translation support device capable of performing translation processing adapted to conventional styles, and a program.SOLUTION: A translation support device is configured to: acquire a translation memory; calculate the degree of match between a target source sentence contained in a target sentence to be processed and a source sentence contained in the translation memory; extract translation pairs whose match degrees fall within a predetermined range and perform fine-tuning of a pre-trained language model using the extracted translation pairs; and subject each target source sentence contained in the target sentence to be processed to machine translation processing using the fine-tuned language model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a translation support device and a program. [Background technology]

[0002] In recent years, translation processing using large-scale language models has been widely used and researched. Non-Patent Document 1 shows that using an in-context method, large-scale language models can translate by learning the terminology and style characteristics of a specific domain without fine-tuning. In particular, it has been shown that incorporating fuzzy matches from translation memories (referring to cases where the match rate that can be used as a translation from a translation memory is below a threshold) into the machine translation process of large-scale language models can improve accuracy and consistency. It has also been shown that they can adapt to the style and terminology of domain-specific sentences, and can sometimes demonstrate accuracy that exceeds that of conventional neural network machine translation systems. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Yasmin Moslem, et al., Adaptive Machine Translation with Large Language Models arXiv:2301.13294 (May 2023) Summary of the Invention [Problem to be solved by the invention]

[0004] However, when using the method disclosed in Non-Patent Document 1, the translation results of documents that have conventional formats, writing styles, and terminology, such as patent documents, may not match the conventional formats, etc.

[0005] for example, [Original text 1] “The information processing apparatus according to claim 2, therefore when the load on the second storage device is less than a third threshold, the processor writes, to the second storage device, data to be written to the first storage device or the second storage device, and when the load on the second storage device is greater than the third threshold, the processor writes the data to the first storage device.” When translated using the method in Reference 1, [Translation 1] "The information processing device described in claim 2 is characterized in that, when the determination unit determines that the load on the second storage device is lower than a third threshold, the data to be written to the first storage device or the second storage device is written to the second storage device, and, when the determination unit determines that the load on the second storage device is higher than the third threshold, the data is written to the first storage device." As can be seen by comparing it with Reference Translation 1, the result was not appropriate in terms of the style of translation of the patent claims.

[0006] On the other hand, when translating manually, [Reference Translation 1] (human translation) "The information processing device according to claim 2, wherein the information processing device writes the data to be written to the second storage device when the determination unit determines that the load on the second storage device is lower than a third threshold, and writes the data to be written to the first storage device when the determination unit determines that the load on the second storage device is higher than the third threshold." As such, translations are currently being carried out in accordance with conventional writing styles and formats.

[0007] The present invention has been made in view of the above circumstances, and one of its objects is to provide a translation support device and a program that are capable of performing translation processing that is compatible with commonly used formats. [Means for solving the problem]

[0008] One aspect of the present invention for solving the problems of the above-mentioned conventional examples is a translation support device that uses a large-scale language model, including: an acquisition means for acquiring a translation memory related to a predetermined translation project, the translation memory storing translations that associate original sentences and translations related to the translation project; a calculation means for calculating the degree of correspondence between a target original sentence included in a target sentence to be processed and the original sentences stored in the acquired translation memory; a fine-tuning processing means for extracting translations where the calculated degree of correspondence falls within a predetermined range, and using the extracted translations to perform fine-tuning processing of a language model that has been machine-trained in advance; a machine translation means for performing machine translation processing of each of the target original sentences included in the target sentence to be processed, using the fine-tuned language model; and a means for outputting the results of the machine translation processing.

[0009] Furthermore, the machine translation means may perform translation using the fine-tuning model for each target original sentence included in the target sentence to be processed, by referring to at least a portion of the parallel translations included in the translation memory. [Effects of the Invention]

[0010] According to the present invention, by performing a fine-tuning process using a translation memory that stores past translation results and then executing a machine translation process, it is possible to perform translation processing that conforms to commonly used formats, etc. Furthermore, by having the machine translation means perform translation using a fine-tuning model for each target original sentence included in the target sentence to be processed, by referring to at least a portion of the parallel translations stored in the translation memory, it is possible to achieve even higher accuracy through a synergistic effect. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a block diagram illustrating an example of the configuration of a translation support device according to an embodiment of the present invention. [Figure 2] 1 is a functional block diagram illustrating an example of a translation support device according to an embodiment of the present invention. [Figure 3] FIG. 2 is a flowchart illustrating an example of the operation of the translation support device according to the embodiment of the present invention. [Figure 4] 1 is an explanatory diagram comparing the processing results of a translation support device according to an embodiment of the present invention with a conventional example; DETAILED DESCRIPTION OF THE INVENTION

[0012] An embodiment of the present invention will be described with reference to the drawings. A translation support device 1 according to an embodiment of the present invention is realized using a general computer device including a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, and a communication unit 15, as shown in FIG.

[0013] The translation support device 1 is also communicably connected to a translation memory system that holds a translation memory that records bilingual translations of original texts (texts written in a first language) and translations (texts obtained by translating the original text into a second language different from the first language) as a result of previous translations. The translation support device 1 may function as this translation memory system, or may be realized by an external server device 2 connected via a communication unit 15.

[0014] In the example of this embodiment, the translation memory system may be a translation memory system that is widely used. Such a translation memory system holds a translation memory for each past translation project (for example, a translation project that switches to a translation task that requires standardization of formats, etc., for each client, project type, etc., in addition to a combination of first and second languages). This translation memory records, for each translation project, a bilingual document that associates sentences (hereinafter referred to as source sentences) included in documents translated in the translation project with the translated sentences.

[0015] The control unit 11 includes a program control device such as a processor, and operates according to a program stored in the storage unit 12. By processing according to this program, the control unit 11 of this embodiment acquires a translation memory that records bilingual translations of original texts and translations resulting from previous translations, and calculates the degree of correspondence between the target original text contained in the target text to be translated and the original text contained in the acquired translation memory.

[0016] Next, the control unit 11 extracts from the translation memory translations whose calculated degree of match falls within a predetermined range, and uses the extracted translations to perform fine-tuning of a language model that has been machine-learned in advance.The control unit 11 then uses the fine-tuned language model to perform machine translation of each of the target original sentences contained in the target sentence to be processed.Detailed operation of the control unit 11 will be described later.

[0017] The storage unit 12 includes a memory device and a disk device, and stores programs to be executed by the control unit 11. The operation unit 13 includes a keyboard, a mouse, etc., and receives user operations and outputs information indicating the content of the operations to the control unit 11.

[0018] The display unit 14 is a display or the like, and displays and outputs information in accordance with instructions input from the control unit 11. The communication unit 15 is a network interface or the like, and outputs information received via a network to the control unit 11. The communication unit 15 also sends information via the network in accordance with instructions input from the control unit 11.

[0019] Next, the operation of the control unit 11 of this embodiment will be described. The control unit 11 of this embodiment operates in accordance with a program stored in the storage unit 12, thereby realizing a configuration that functionally includes an acquisition unit 21, a coincidence calculation unit 22, a bilingual extraction unit 23, a fine-tuning processing unit 24, a machine translation unit 25, and an output unit 26, as shown in FIG.

[0020] Here, the acquisition unit 21 receives a translation project specification from the user (specification of a past translation result that has a common set of first and second languages ​​and a common format with the translation target, etc.), and acquires a translation memory related to the translation project from the translation memory system.

[0021] The matching degree calculation unit 22 divides the target document to be translated into predetermined units (hereinafter referred to as the target original sentence; it may be a segment unit widely used in translation memory systems). Once the calculation unit 22 has divided the target original sentence into predetermined units (hereinafter referred to as the target original sentence, which may be a segment unit widely used in translation memory systems), the calculation unit 22 calculates the degree of matching between the target original sentence and each original sentence included in the translation memory acquired by the acquisition unit 21 (the original sentence included in the acquired translation memory will be referred to as the comparison original sentence, hereinafter).

[0022] Here, the degree of agreement is the degree of agreement between the target text and the comparison text, e.g. The ratio of the number of words commonly used in the target and comparison texts to the total number of words in the target text (morpheme matching) The percentage of common word pairs that appear in the same order (word order) (morpheme order) There are various methods for calculating the degree of similarity between a pair of sentences, and detailed explanations will be omitted, but in the following, the degree of similarity will be expressed as a percentage ranging from "0" (complete mismatch) to "100" (complete match).

[0023] The parallel translation extraction unit 23 extracts parallel translations including a comparison source text whose degree of coincidence calculated by the coincidence calculation unit 22 is within a predetermined range. Specifically, the parallel translation extraction unit 23 extracts parallel translations including a comparison source text whose degree of coincidence with one of the target source texts in the target document is within the predetermined range.

[0024] The fine tuning processing unit 24 executes fine tuning processing of a language model that has been machine-learned in advance, using the parallel translations extracted by the parallel translation extraction unit 23. Here, for example, the language model may be GPT (Generative Pre-trained Transformer)-3.5, and the fine tuning processing may utilize a fine tuning API provided by OpenAI.

[0025] The machine translation unit 25 sequentially performs machine translation processing on the target original sentences included in the target document using the language model after fine-tuning processing. Specifically, when the machine translation unit 25 is a fine-tuned GPT-3.5, the machine translation unit 25 inputs a prompt to translate the target original sentence to the fine-tuned GPT-3.5, which is the language model.

[0026] In this case, the machine translation unit 25 may perform translation using a language model, with reference to at least a portion of the parallel translations stored in the translation memory, for each of the target original sentences included in the target sentence to be processed. In this example, the machine translation unit 25 may extract a portion of the parallel translations stored in the translation memory acquired by the acquisition unit 21 based on the degree of agreement with the target original sentence (which may be the degree of agreement calculated by the agreement calculation unit 22), and perform translation using a language model with reference to the extracted parallel translations. In this example, the machine translation unit 25 displays the extracted parallel translations in response to a prompt to translate the target original sentence, and instructs the user to translate the target original sentence with reference to the parallel translations (processing similar to so-called in-context translation).

[0027] In response to the prompt, the machine translation unit 25 outputs the translation result output by the language model, such as fine-tuned GPT-3.5, to the output unit 26. Then, the output unit 26 outputs the result of the machine translation by the machine translation unit 25 to a file or to the outside.

[0028] [Operation] The translation support device 1 according to this embodiment basically has the above configuration and operates as follows:

[0029] In the following example, an external server (hereinafter referred to as the language model server) that provides fine-tuning services based on GPT-3.5 is used as the language model. In this example, the translation support device 1 receives from the user a translation project specification and a target document to be translated, and starts the process illustrated in FIG. 3. Here, the translation project specification can be, for example, "translation of claims in an English patent application document whose main IPC (International Patent Classification) is G06F into claims in a Japanese patent application document."

[0030] First, the translation support device 1 acquires a translation memory related to a specified translation project from the translation memory system (S11).

[0031] The translation support device 1 also calculates the degree of agreement between each target original sentence obtained by dividing the target document to be translated into segments and the corresponding parallel original sentence (comparison original sentence) contained in the translation memory acquired in step S11 (S12), and extracts parallel translations including the comparison original sentence whose degree of agreement is within a predetermined range (for example, between a lower limit of 65% and an upper limit of 95%) (S13).

[0032] The calculation of the degree of match in step S12 may be performed by matching morphemes and calculating the ratio of the number of words commonly used in the subject and comparison texts to the total number of words in the subject text. The range of the degree of match here may be one in which only the lower limit is set.

[0033] The translation support device 1 repeatedly executes the processes of steps S12 and S13 for each target original text, and extracts translations including comparison original texts whose degree of matching with any of the target original texts is within the predetermined range.

[0034] The translation support device 1 uploads the translations extracted in step S13 to an external language model server and requests fine tuning (S14). The translation support device 1 waits for the completion of fine tuning before proceeding to the next process, sequentially selecting target sentences included in the target document and performing the next process. That is, the translation support device 1 acquires, as a reference translation, one of the translations including the comparison sentence extracted in step S13 (e.g., the one with the highest degree of similarity), which is a comparison sentence that has a relatively high degree of similarity with the selected target sentence (S15). The translation support device 1 then instructs the external language model server after the fine tuning request to translate the selected target sentence with reference to the acquired reference translation (S16), and acquires and records the response (translation) (S17).

[0035] In step S15, if no comparison source text is found that matches the target source text relatively well (for example, if the highest match between the target source text and each comparison source text does not exceed a predetermined lower threshold, such as 50%), the translation support device 1 may instruct the external language model server after the fine-tuning request to simply translate the selected target source text without obtaining a reference translation, and may obtain and record the response.

[0036] The translation support device 1 repeatedly executes the processes of steps S15 to S17 for each target original sentence to obtain a translation corresponding to each target original sentence. Then, when the translation support device 1 has recorded translations for all target original sentences, it outputs the recorded translations (S18).

[0037] [If there is a matching translation] Furthermore, in the processing up to this point, if a translation containing a comparison source text whose degree of similarity with the target source text is equal to or greater than a predetermined threshold value (for example, a comparison source text whose degree of similarity with the target source text is 95% or greater) is extracted in step S15, the translation support device 1 may record the translation contained in the translation, instead of issuing a translation instruction in step S16, and return to the process of selecting the next target source text.

[0038] [Special Processing] In one example of this embodiment, the translation support device 1 may perform a provisional machine translation process on the target original text, excluding reference tags contained in the target original text, and then use a language model to refer to the results of the provisional machine translation process to generate a machine translation result that includes the reference tags.

[0039] Specifically, the reference tag may be the following tag (a tag string inserted into a sentence): タグが付される)、下付き文字(タグが付される)や、下線(タグが付される)等、種々の例がある。本実施の形態の翻訳支援装置1は、このように対象原文に参照タグが含まれる場合には、当該参照タグを除いて仮に、言語モデルを用いた(つまり生成AIにより)機械翻訳処理を行う。

[0040] またこの例の翻訳支援装置1は、参照タグを含む対象原文全体を、当該所定のパターンを含む部分をそのまま対応する訳文に翻訳する機械翻訳システムに入力して、対応する訳文を得る。

[0041] 一例として、所定のパターンが上記参照タグであるときには、このような機械翻訳システムとして、言語モデルを用いるものではなく、参照タグを含めて翻訳が可能なニューラルネットワーク翻訳システム(NMT)などを用いる。このような翻訳システムは広く知られたものがあるので、ここでの詳しい説明は省略する。

[0042] 本実施の形態の翻訳支援装置1は、図3に例示した処理のステップS16において、外部の言語モデルサーバに送信しようとする対象原文に、上記参照タグがあるか否かを判断する。ここで対象原文に、上記参照タグがあれば翻訳支援装置1は、対象原文から当該参照タグを除いた補正対象原文を作成し、この補正対象原文を、取得した参照対訳を参照して翻訳するよう外部の言語モデルサーバに指示し、その応答である訳文(以下補正訳文と呼ぶ)を得る(ステップS17)。

[0043] 一方、翻訳支援装置1は、補正前の対象原文を別途、参照タグ等の上記参照タグを含めて翻訳が可能な翻訳システムに送出し、その応答として訳文(参考訳文と呼ぶ)を得る。翻訳支援装置1は、次に、言語モデルを用いた生成AI(上記外部の言語モデルサーバでよい)に対して、参考訳文を参照して、上記参照タグを含めて補正訳文を修正するよう指示する。そして翻訳支援装置1は、当該指示に応答して生成AIが出力する、修正後の訳文を得て、対象原文を翻訳した訳文として記録する。

[0044] なお、外部の言語モデルサーバに送信しようとする対象原文に、上記参照タグがないと判断したときには、翻訳支援装置1は、図3の例と同様に、ステップS15で取得した参照対訳を参照して、選択した対象原文の翻訳を行うよう、外部の言語モデルサーバに指示し(ステップS16)、その応答(訳文)を取得して(ステップS17)、その訳文を記録する。

[0045] 本実施の形態のこの例によると、ファインチューニングを施した、言語モデルを用いた翻訳システムが参照タグ等に対応していなくても、当該翻訳システムが生成した、慣用された様式等に適合した翻訳結果に、参照タグ等を取り込んだ翻訳を反映させて得ることができる。このとき、参照タグに対応する翻訳システムは、本実施の形態のように、慣用された様式等に適合した翻訳が得られるものでなくてもよい。

[0046] [実施形態の効果]本実施の形態の例によると、慣用された様式等に適合した翻訳処理を行うことが可能となる。具体的に、技術分野として筆頭IPC(国際特許分類)のメイン分類がG06Fである特許文献から得られる原文と参考訳との組を50文対用意して翻訳の対象とし、(1)GPT3.5をそのまま利用した結果、(2)非特許文献1のインコンテキスト法による結果、(3)予め同種の特許文献に係る対訳を記録した翻訳メモリを取得して、上述の例のようにファインチューニングを施した本実施の形態の翻訳支援装置1を用いて翻訳した結果と、(3)処理の対象とする対象文に含まれる対象原文のそれぞれについて、翻訳メモリに含まれる対訳の少なくとも一部を参照して、前記ファインチューニングモデルを用いた翻訳とを比較した。なお、比較にあたってはBLEU(Bilingual Evaluation Understudy)とRIBES(Rank-based Intuitive Bilingual Evaluation Score)の指標を用いた。これらは、機械翻訳の精度を評価するための指標で、機械翻訳されたテキストと人間が翻訳した参照テキストとの間の類似性を測定することで、翻訳の質を評価するものである。

[0047] ここでBLEUは、単語のn-gramの重なりを計算することにより、機械翻訳の出力と一つ以上の参照翻訳との間の類似度を測定する。BLEUスコア(%)は0から100までの値を取り、100に近いほど翻訳の質が高いことを示す。BLEUは、文脈や文の流れを考慮しないという問題点があるが、機械翻訳の性能比較に広く使用されている。

[0048] またRIBESは、翻訳された文の単語の順序を参照翻訳と比較することで、機械翻訳の自然さと流暢さを評価する指標である。RIBESスコア(%)もBLEU同様0から100までの値を取り、100に近いほど翻訳の質が高いことを示す。RIBESは、単語の順序の正確さに基づいて計算され、特に日本語などの語順が英語と大きく異なる言語に対する翻訳評価に有効である。RIBESは、順序の一致に基づいてスコアを算出するため、BLEUよりも請求項などの文構造をより良く考慮することができる。

[0049] 比較例(1):GPT3.5を用いた例では、BLEU:34、 RIBES:57 (N=50)のスコアを得た。比較例(2):インコンテキスト法を利用した例では、BLEU:46、 RIBES:62 (N=50)のスコアを得た。さらに、実施例(3):ファインチューニングによる本実施の形態の翻訳支援装置1については、BLEU:46、 RIBES:84 (N=50)のスコアを得た。実施例(4):ファインチューニングにさらにインコンテキスト法を加えた本実施の形態の翻訳支援装置1については、BLEU:70、 RIBES:85 (N=50)のスコアを得た。

[0050] これらの結果をまとめると、図4に示されるように、実施例(2)のインコンテキスト法ではBLEUを増加する傾向が強く、一方実施例(3)のファインチューニング法ではBLEUのみならずどちらかというとRIBESを増加する傾向が顕著であった。(1)及び(2)とはことなり、以下のように日本語の特許請求項特有の文体を再現することができた。原文サンプル1:The arithmetic processing device according to claim 1, further comprising: an update information generating unit that generates, based on the statistical information acquired by the statistical information acquiring unit, first update information for updating the decimal point position of the fixed-point number output by the first converting unit and second update information for updating the decimal point position of the fixed-point number output by the second converting unit, wherein the first converting unit sets the decimal point position of the fixed-point number to be output based on the first update information, and the second converting unit sets the decimal point position of the fixed-point number to be output based on the second update information.に対する実施例(3)の訳文は、訳文2のようであった。[訳文2]「前記統計情報取得部が取得した統計情報に基づいて、前記第1変換器が出力する固定小数点数の小数点位置を更新する第1の更新情報と、前記第2変換器が出力する固定小数点数の小数点位置を更新する第2の更新情報とを生成する更新情報生成部をさらに備え、前記第1変換器は、前記第1の更新情報に基づいて、出力する固定小数点数の小数点位置を設定し、前記第2変換器は、前記第2の更新情報に基づいて、出力する固定小数点数の小数点位置を設定する、ことを特徴とする請求項1に記載の演算処理装置。」

[0051] この訳文2では、日本語の文体(語順)が参照訳と概ね一致した結果、RIBESが大きく向上したと考えられる。また実施例(4)では、ファインチューニングにさらにインコンテキスト法を加えた。この実施例では、RIBESの増加に加えて、BLEUの増加も顕著であり、単純な足し算以上の相乗効果がみられることが分かった。

符号の説明

[0052] 1 翻訳支援装置、2 サーバ装置、11 制御部、12 記憶部、13 操作部、14 表示部、15 通信部、21 取得部、22 一致度演算部、23 対訳抽出部、24 ファインチューニング処理部、25 機械翻訳部、26 出力部。

Claims

1. An acquisition means for acquiring a translation memory related to a predetermined translation project, the translation memory storing bilingual translations that associate original texts and translations related to the translation project; a calculation means for calculating the degree of agreement between a target original sentence included in a target sentence to be processed and an original sentence included in the acquired translation memory; a fine-tuning processing means for extracting translations whose calculated degree of coincidence falls within a predetermined range, and for executing a fine-tuning process using the extracted translations to generate a language model that has been machine-learned using a large-scale language model; a machine translation means for performing machine translation processing of each of the target original sentences included in the target sentence to be processed using the fine-tuned language model based on a large-scale language model; means for outputting the results of the machine translation processing; A translation support device comprising:

2. 2. The translation support device according to claim 1, The machine translation means performs translation using the fine-tuning model for each target original sentence included in the target sentence to be processed, by referring to at least a portion of the parallel translations included in the translation memory.

3. 3. The translation support device according to claim 2, The machine translation means extracts a portion of the parallel translations contained in the translation memory based on the degree of similarity with the target original text when translating using the language model, and performs translation using the language model by referring to the extracted parallel translations.

4. 2. The translation support device according to claim 1, The machine translation means In the machine translation process of the target original text, the machine translation process is performed excluding reference tags contained in the target original text; A translation support device that performs machine translation processing by using a large-scale language model to generate machine translation results including reference tags from new machine translation results including reference tags obtained by neural network machine translation and the machine translation processing excluding the reference tags.

5. 5. A translation support device according to claim 1, A translation support device in which the range of the degree of coincidence used by the fine-tuning processing means to extract the parallel translations is determined by a lower limit value and an upper limit value.

6. Computer, An acquisition means for acquiring a translation memory related to a predetermined translation project, the translation memory storing bilingual translations that associate original texts and translations related to the translation project; a calculation means for calculating the degree of agreement between a target original sentence included in a target sentence to be processed and an original sentence included in the acquired translation memory; a fine-tuning processing means for extracting translations whose calculated degree of coincidence falls within a predetermined range, and for executing fine-tuning processing of a language model that has been machine-learned in advance using the extracted translations; a machine translation means for performing machine translation processing on each of the target original sentences included in the target sentence to be processed using the fine-tuned language model; means for outputting the results of the machine translation processing; A program that functions as a

Images