Inference device, inference method, and program

Abstract

To perform translation using an original language input sentence and a target language sentence similar to the original language input sentence without using parallel translation data as a translation memory.SOLUTION: An inference device includes a retrieval part for extracting a plurality of similar translations similar to an original language input sentence from a set of target language sentences by performing language cross retrieval with the original language input sentence as a query, an inference part for generating a plurality of output candidate sentences by performing translation by using the original language input sentence and each similar translation in the plurality of similar translations, and a re-ranking part for performing re-ranking to the plurality of output candidate sentences.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a technique for performing machine translation using a neural network. [Background technology]

[0002] Conventional techniques relating to neural machine translation using a translation memory, which is a collection of high-quality bilingual data, include those disclosed in Non-Patent Documents 1 and 2, for example.

[0003] The techniques disclosed in Non-Patent Documents 1 and 2 extract target language sentences that are paired with source language sentences similar to the input sentence based on the similarity between the input sentence and the source language sentences in a translation memory, and then link the target language sentences with the input sentence to provide input to a translation model. This type of processing can improve translation accuracy without changing the architecture of neural machine translation. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Bram Bulte and Arda Tezcan. Neural fuzzy repair: Integrating fuzzy matches into neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 1800-1809, Florence, Italy, July 2019. Association for Computational Linguistics. [Non-patent document 2] Nabil Hossain, Marjan Ghazvininejad, and Luke Zettlemoyer. Simple and effective retrieve-edit-rerank text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 2532-2538, Online, July 2020. Association for Computational Linguistics. Summary of the Invention [Problem to be solved by the invention]

[0005] Depending on the language pair or the translation target domain, there may not be a sufficient amount of bilingual data. If there is not enough bilingual data to serve as a translation memory, the methods disclosed in Non-Patent Documents 1 and 2 cannot be applied.

[0006] The present invention has been made in consideration of the above points, and aims to provide a technology for performing translation using a source language input sentence and a target language sentence similar to the source language input sentence, without using bilingual data as a translation memory. [Means for solving the problem]

[0007] According to the disclosed technology, there is provided a search unit that extracts a plurality of similar translations to a source language input sentence from a set of target language sentences by performing a cross-language search using the source language input sentence as a query; an inference unit that generates a plurality of output candidate sentences by performing translation using the source language input sentence and each of the plurality of similar translations; a reranking unit that performs reranking on the plurality of output candidate sentences; A reasoning apparatus is provided, comprising: [Effects of the Invention]

[0008] According to the disclosed technology, it is possible to perform translation using a source language input sentence and a target language sentence similar to the source language input sentence, without using bilingual data as a translation memory. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 10 is a diagram showing how a pair of an input sentence and a target language sentence is input to a translation model. [Figure 2] FIG. 1 is a diagram illustrating an example of the configuration of a training device 100. [Figure 3] 10 is a flowchart illustrating an example of the operation of the training device 100. [Figure 4] FIG. 2 is a diagram illustrating an example of the configuration of an inference device 200. [Figure 5] 10 is a flowchart illustrating an example of the operation of the inference device 200. [Figure 6] FIG. 10 is a diagram showing the flow of an inference process. [Figure 7] FIG. 1 is a diagram showing the number of sentences for each of the training sentences, development sentences, test sentences, and translation memory (target language sentences) used in the experiment. [Figure 8] FIG. 10 is a diagram showing a comparison of translation accuracy between similar translation search methods and reranking methods. [Figure 9] FIG. 2 illustrates an example of a hardware configuration of the apparatus. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, an embodiment of the present invention (the present embodiment) will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0011] In the following, we will first explain related technologies and their problems, and then explain the technology according to this embodiment in detail. In the following, "translation memory" is used to mean "a set of pairs of source language sentences and target language sentences" or "a set of target language sentences." A "set of pairs of source language sentences and target language sentences" may also be called bilingual data. Furthermore, "extraction" means obtaining data as a result of a search process.

[0012] Furthermore, hereinafter, machine translation refers to machine translation using a neural network translation model. However, machine translation applicable to the technology according to the present invention is not limited to machine translation using a neural network translation model, and machine translation using a translation model other than a neural network may also be applied to the technology according to the present invention.

[0013] (Regarding related technologies) Non-Patent Document 1 discloses a technology called Neural Fuzzy Repair (NFR) that improves translation accuracy by incorporating a translation memory into machine translation. Figure 1 is a diagram showing how NFR searches a translation memory for a source language sentence similar to an input sentence, extracts the corresponding target language sentence, and inputs the pair of input and target language sentences into a translation model. The processing shown in Figure 1 is also performed in the present invention.

[0014] In the NFR disclosed in Non-Patent Document 1, an input source language sentence is used as a query, and similar sentences that are similar to the input sentence are extracted from a collection of source language sentences in a translation memory based on the edit distance. Similar sentences in a target language (called similar translations), which are translations of the similar sentences in the source language, are then concatenated with the input sentence and used as input to a translation model. The output from the translation model is obtained as a target language sentence, which is a translation of the input source language sentence.

[0015] The edit distance is defined as the minimum number of operations required to transform an original string into another string by "inserting, deleting, or substituting" one character. In NFR, the similarity sim(x,y) between string x and string y is expressed by the following formula (1):

[0016]

number

[0017] In similar sentence search using edit distance, it is necessary to calculate and compare the similarity between the source language sentence and the input sentence for all source language sentences contained in the translation memory. Therefore, when using a large-scale translation memory, the calculation cost becomes significantly high. Therefore, Non-Patent Document 1 describes a similarity measure, containment, provided by the Python library SetSimilaritySearch. max We use a method to calculate the edit distance for a set of similar sentence candidates searched using

[0018] In this embodiment, this method is expressed as SetSimilaritySearch+EditDistance(sss+ed).

[0019] NFR simply uses translation memory to extend the input to the translation model, so translation memory can be incorporated into machine translation without changing the architecture of the machine translation. Therefore, it is highly compatible with various existing translation models and highly portable in terms of implementation.

[0020] However, this method requires that the search for similar sentences for the input sentence must be a bilingual sentence pair. Depending on the language pair or target domain, it may not be possible to prepare a sufficient amount of bilingual data, and in such cases NFR cannot be applied.

[0021] Furthermore, due to the constraints imposed by the maximum input sentence length for machine translation, the number of similar translations available for NFR is limited to a few sentences at most. Even if many useful similar translations are obtained, NFR cannot utilize all of them.

[0022] Non-Patent Document 2 discloses a method called Retrieve-Edit-Rerank, which utilizes a large number of similar translations.

[0023] In Retrieve-Edit-Rerank, similar to NFR, multiple different similar translations for an input sentence are extracted from a translation memory based on edit distance, and for each of the multiple similar translations, the input sentence and the similar translation are linked together to form an input for a translation model. Then, multiple different output candidate sentences are obtained from the translation model, and the output candidate sentences are reranked based on log-likelihood. Non-Patent Document 2 experimentally shows that translation accuracy is improved by selecting the output candidate sentence with the highest log-likelihood.

[0024] (About the assignment) Related techniques such as Neural Fuzzy Repair and Retrieve-Edit-Rerank require a sufficient amount of bilingual data for the language pair or target domain to be translated as a translation memory. Therefore, there is a problem that these techniques cannot be applied to language pairs or target domains for which a sufficient amount of bilingual data does not exist.

[0025] In addition, Retrieve-Edit-Rerank compares multiple output candidate sentences based on the log-likelihood output by the translation model, which means that if the lengths of the output candidate sentences are different, longer sentences will be at a disadvantage.

[0026] Furthermore, the log-likelihood used in Retrieve-Edit-Rerank is the probability of outputting an output sentence for an input that combines an input sentence and its similar translation, and therefore does not necessarily represent the translation equivalence or semantic similarity between the input sentence and the output sentence.

[0027] (Overview of the technology according to the present embodiment) In this embodiment, a collection of monolingual data in the target language (target language sentences) is used as the translation memory. During inference, an inference device 200 (described later) first uses a multilingual sentence embedding generation model to extract multiple target language sentences (output candidate sentences) similar to the input sentence from the translation memory (collection of target language sentences) by cross-language search. This eliminates the need to prepare bilingual data as the translation memory. All that is required is a collection of target language sentences in the domain to be translated.

[0028] Note that bilingual data may be used as the translation memory. In this case, a set of target language sentences in the bilingual data is used as the target of the cross-language search.

[0029] Next, inference device 200 reranks the multiple output candidate sentences. In this embodiment, the score function based on the logarithmic likelihood used in Retrieve-Edit-Rerank is normalized by sentence length, and an improved score function is used that takes into account the semantic similarity with the input sentence.

[0030] In this embodiment, the reranking may also use a score function based on the logarithmic likelihood used in Retrieve-Edit-Rerank. Also, only one of the two improvements mentioned above, "normalization based on sentence length" and "taking into consideration semantic similarity with the input sentence", may be applied.

[0031] The technology according to this embodiment can improve translation accuracy compared to Retrieve-Edit-Rerank.

[0032] (Similarity measure based on sentence embeddings) Here, the similarity measure based on sentence embedding used in this embodiment will be described.

[0033] A sentence embedding maps a sentence to a high-dimensional real vector. The conversion of sentence x into a sentence embedding is denoted as E(x). The similarity sim(x,y) between sentence x and sentence y using the sentence embedding is defined as follows:

[0034]

number

[0035] mSBERT is a multilingual version of SBERT, an English version of SBERT trained on the NLI (Natural Language Inference) dataset and achieving high accuracy in the STS (Semantic Textual Similarity) task, which has been distilled to generate language-independent sentence embeddings. LaBSE is a multilingual sentence embedding generation model trained on bilingual texts and has achieved high accuracy in the BUCC (Building and Using Comparable Corpora) task, which performs bilingual text search.

[0036] In this embodiment, in the cross-language search, the sentence embedding E(s) obtained by converting the source language input sentence s using the multilingual sentence embedding generation model is used as a query to search for similar translations in the monolingual data in the target language, and k similar translations t'1, t'2, ..., t' k For example, the top k similar translations based on the similarity score calculated by the above formula (2) are extracted.

[0037] More specifically, for example, the sentence embedding E(s) obtained by converting the source language input sentence s using a multilingual sentence embedding generation model is used as a query, and a vector neighborhood search is performed on the monolingual data in the target language to find k similar translations t'1, t'2,..., t'k The technique disclosed in FAISS (Jeff Johnson, Matthijs Douze, and Herve Jegou. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, Vol. vol.7, pp. 535-547, July 2021) may be used for vector neighborhood search.

[0038] The configurations and operations of the training device 100 and the inference device 200 in this embodiment will be described in detail below. Note that the training device 100 and the inference device 200 may be configured as a single device. In addition, in the following description of the device operation, explanations of tokenization and subword division, which are generally performed in translating sentences using a translation model, will be omitted.

[0039] (Configuration and Operation Examples of Training Device 100) Fig. 2 shows an example of the configuration of a training device 100 according to this embodiment. As shown in Fig. 2, the training device 100 includes a training data DB 110, a translation memory DB 120, a search unit 130, a training unit 140, and a translation model DB 150. Note that DB is an abbreviation for database. "DB" may also be called a "storage unit" or a "storage unit."

[0040] As mentioned above, in this embodiment, the term "translation memory" is used to mean "a collection of target language sentences" or "a collection of pairs of source language sentences and target language sentences." However, the term is not limited to this, and may also be used to mean "a database, memory unit, storage unit, or memory device" that stores "a collection of target language sentences" or "a collection of pairs of source language sentences and target language sentences."

[0041] The training data DB 110 stores training data consisting of a set of pairs of source language sentences s and target language sentences t. The translation memory DB 120 stores a set of target language sentences (monolingual data in the target language) as a translation memory.

[0042] In this embodiment, the widely used Transformer model is used as the translation model. The Transformer model has an encoder and a decoder, and a source language sentence is input to the encoder, and a translated sentence is output from the decoder. Note that a translation model other than the Transformer model may also be used as the translation model.

[0043] An example of the operation of the training device 100 will be described with reference to the flowchart in Fig. 3. For convenience of explanation, the description will be given assuming that "pairs of source language sentences and target language sentences" in the training data are processed one by one, but processing may also be performed for multiple pairs (batches).

[0044] In S101, a source language sentence s in training data read from the training data DB 110 is input to the search unit 130, and a target language sentence t is input to the training unit 140. The search unit 130 has a multilingual sentence embedding generation model, and the training unit 140 has a translation model to be trained.

[0045] In S102, the search unit 130 performs a cross-language search using the multilingual sentence embedding generation model with the source language input sentence s as a query against the translation memory (a set of target language sentences) in the translation memory DB 120, and retrieves k target language similar sentences (referred to as similar translations) t'1, t'2, ..., t' k Specifically, in cross-language search, vector neighborhood search is used.

[0046] In S103, the training unit 140 calculates a special token for each of the k similar translations. <sep>Source language sentence s and similar translation sentence t' i The concatenation of and is used as input to the translation model, and the translation model is trained so that the target language sentence t is output from the translation model. More specifically, training the translation model means optimizing the parameters of the translation model.

[0047] If the training has converged (Yes in S104), the process ends. If the training has not converged, the process returns to S101.

[0048] The trained translation model obtained by training device 100 is stored in translation model DB 150. The translation model is read from translation model DB 150, input to inference device 200, and used for inference in inference device 200.

[0049] (Configuration and Operation Examples of the Inference Device 200) Next, we will explain an example configuration and operation of inference device 200. Fig. 4 shows an example configuration of inference device 200. As shown in Fig. 4, inference device 200 includes a translation memory DB 210, a search unit 220, a translation model DB 230, an inference unit 240, a reranking unit 250, an input unit 260, and an output unit 270.

[0050] The translation memory DB 210 stores a collection of target language sentences as a translation memory. The translation model DB 230 stores trained translation models (specifically, parameters). The inference unit 240 reads out the translation models from the translation model DB 230 and uses them for inference. Both the search unit 220 and the reranking unit 250 have multilingual sentence embedding generation models.

[0051] An example of the operation of the inference device 200 will be described with reference to the flowchart of FIG.

[0052] In S201, a source language input sentence s is input by the input unit 260. The source language input sentence s is passed to the search unit 220 and the inference unit 240, respectively.

[0053] In S202, the search unit 220 performs a cross-language search on a translation memory consisting of a set of target language sentences, using a multilingual sentence embedding generation model and a source language input sentence s as a query, and retrieves k target language similar sentences (similar translations) t'1, t'2, ..., t' k Specifically, in cross-language search, vector neighborhood search is used.

[0054] In S203, the inference unit 240 calculates a special token for each of the k similar translations. <sep>The source language input sentence s and the similar translation sentence t' are separated by a i The concatenation of these sentences is input to the translation model, and an output sentence is obtained from the translation model. This output sentence is a candidate for the final output sentence, so it may be called an output candidate sentence. Since there are k similar translations, the inference unit 240 performs k translations (which may also be called decoding) to obtain k output sentences. i get.

[0055] k output sentences o i , and the reranking score Q i The information such as probability values required for calculating the reranking score Q is passed to the reranking unit 250. i Calculate the reranking score Q i may be passed to the reranking unit 250. In this case, the reranking score Q i It may be interpreted that the function of calculating the above is included in the reranking unit 250.

[0056] In S204, the reranking unit 250 i The reranking score Q for each i Calculate the reranking score Q i Based on the output statement i Here, as shown in the formula below, the reranking score Q i Let i be the index that maximizes * Let's say.

[0057]

number

[0058] In S205, the output unit 270 outputs the reranking score Q i is the maximum output sentence o i* Alternatively, the k reranked output sentences may be output in descending order of score. Alternatively, the output may be performed in a manner other than the above. The reranking unit 250 may include the output unit 270.

[0059] The reranking process executed by the reranking unit 250 will be described in detail below.

[0060] "Source language input sentence s and similar translation sentence t' i ” output sentence based on the translation model i The output probability of p MT (o i |s,t´ i )

[0061] The reranking unit 250 MT (o i |s,t´ i ) can be calculated using the following formula:

[0062]

number

[0063] In the technique disclosed in Non-Patent Document 2, the following logarithmic likelihood Q is used as the reranking score. i (Hossain) is used.

[0064]

number

[0065]

number

[0066] In this embodiment, a reranking score may be used that takes into account the log-likelihood normalized by the sentence length and the similarity between the input sentence and the output candidate sentence based on multilingual sentence embedding, as follows:

[0067]

number

[0068] It is also possible to use a reranking score using the similarity between the input sentence and the output candidate sentence based on the multilingual sentence embedding, without normalizing by sentence length.

[0069] The flow of the above-mentioned inference process is summarized in Figure 6. Figure 6 shows that the translated sentence with the highest reranking score is output, but as mentioned above, multiple translated sentences may be output.

[0070] (About the experiment) An experiment was conducted to verify the effectiveness of the technology according to the present embodiment, and the experimental setup and results will be described below.

[0071] In this experiment, we used the ASPEC English-Japanese corpus of Asian academic papers and the English-French corpus of the EU Bookshop Corpus (hereafter referred to as EUBC), which was created based on publications from European institutions. The translation directions were English to Japanese and English to French, respectively.

[0072] For training the translation model, only 100,000 sentence pairs randomly sampled from each corpus were used, and the rest were used as a monolingual translation memory. Figure 7 shows the number of sentences for each of the training sentences, development sentences, test sentences, and translation memory (target language sentences).

[0073] For Japanese text, we used MeCab to tokenize it, then decomposed it into subwords using Byte Pair Encoding (BPE) with 32,000 operations. For English and French text, we used Moses tokenizer to tokenize it, then decomposed it into subwords using Byte Pair Encoding (BPE) with 32,000 operations.

[0074] <Experimental Setup> In this experiment, the search unit 220 performed a search for similar translations using three methods: edit distance (sss+ed), mSBERT, and LaBSE, and the results were compared.

[0075] Of these, the edit distance sss+ed can only search for similar translations within the same language, so the search target was limited to 100,000 source language sentences from the training sentences. A similar sentence search was performed based on the edit distance between the input sentence and the source language sentence in the translation memory, and the target language sentence corresponding to the searched source language sentence was determined to be a similar translation. On the other hand, mSBERT and LaBSE can use the input sentence as a query to perform a cross-language search directly on the target language sentence, so the search target was 2 million sentences for ASPEC and 8.42 million sentences for EUBC.

[0076] During training, we conducted a preliminary experiment comparing a normal method that does not perform similar sentence search (no search) with a method that uses up to four similar translations (top1 to top4). As a result, the translation accuracy was highest when the number of similar translations during training was k=2, so k=2 was used in subsequent experiments.

[0077] During inference, there are three methods: "a method that does not use similar translations (no similar translations)", "a method that uses only the top 1 similar translation (k=1) as in NFR", and "a method that uses the log likelihood Q corresponding to Retrieve-Edit-Rerank for the top 32 similar translations (k=32)". (Hossain) "The proposed method uses the normalized log-likelihood Q (proposed1) "A re-ranking method based on the normalized log-likelihood and the similarity between the input sentence and the output sentence, which is another proposed method, is used for the top 32 (k=32) similar translations. (proposed2) We compared "methods for reranking based on the above."

[0078] Among these, the combination of "no search" and "no similar translation" represents the translation accuracy of the normal vanilla Transformer. The combination of sss+ed and k=1 corresponds to the conventional Neural Fuzzy Repair. sss+ed and Q (Hossain) This combination corresponds to the conventional method of Retrieve-Edit-Rerank.

[0079] The automatic evaluation scale BLEU was used to evaluate translation accuracy. For the experiment, a Transformer model implemented in PyTorch (registered trademark) was used. The encoder and decoder each had six layers, the hidden dimension was 512, the FF layer dimension was 2048, and the number of multi-heads was 8. Training was performed for 30 epochs with a batch size of 96 sentences and 6,000 warm-up steps, and the BLEU of the test sentence at the epoch with the highest BLEU of the development sentence is reported as the experimental result.

[0080] <Experimental Results> Figure 8 shows a comparison of translation accuracy using different similar translation search methods and reranking methods. In the ASPEC English-Japanese translation and EUBC English-French translation in this experiment, the combination of "sss+ed similar sentence search in translation memory (parallel data) based on edit distance and k=1 using only the top similar translation," which corresponds to the conventional Neural Fuzzy Repair method, did not show any improvement in accuracy compared to "no search" and "no similar translation," which correspond to translation using the baseline Transformer.

[0081] On the other hand, the conventional method, Retrieve-Edit-Rerank, corresponds to the "sss+ed and log likelihood Q (Hossain) The "combination of reranking based on" showed a slight improvement in accuracy.

[0082] In contrast to this, the technology according to the present embodiment uses "cross-lingual search of translation memories (monolingual) using a multilingual sentence embedding generation model (mSBERT, LaBSE) and log likelihood Q normalized by sentence length" (proposed1) Furthermore, Q considers the similarity between input and output sentences based on multilingual sentence embedding. (proposed2) The combination of reranking based on "translation accuracy" shows a greater improvement.

[0083] (Example of hardware configuration) Any of the devices described in this embodiment (training device 100, inference device 200) can be realized, for example, by causing a computer to execute a program. This computer may be a physical computer or a virtual machine on the cloud.

[0084] That is, the device can be realized by executing a program corresponding to the processing performed by the device using hardware resources such as a CPU and memory built into a computer. The program can be recorded on a computer-readable recording medium (such as a portable memory) and stored or distributed. The program can also be provided via a network such as the Internet or email.

[0085] Fig. 9 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 9 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a CPU 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, and the like, all of which are interconnected via a bus B. The computer may further include a GPU.

[0086] A program for realizing processing on the computer is provided by a recording medium 1001 such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0087] The memory device 1003 reads and stores a program from the auxiliary storage device 1002 when an instruction to start the program is received. The CPU 1004 realizes the functions related to the device in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) or the like according to the program. The input device 1007 is composed of a keyboard, mouse, buttons, a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the results of calculations.

[0088] (Summary of implementation form, effects, etc.) As described above, the technology described in this embodiment makes it possible to perform translation using a source language input sentence and a target language sentence that is similar to the source language input sentence, without using bilingual data as a translation memory.

[0089] In other words, in conventional technology, bilingual data for the language pair and region to be translated was required as a translation memory, but in the technology of this embodiment, similar translations to an input sentence are extracted from a set of target language sentences by cross-language search using a multilingual sentence embedding generation model, so that monolingual data (rather than bilingual data) for the target region of the target language can be used as a translation memory.

[0090] Furthermore, in the Retrieve-Edit-Rerank method, which uses multiple similar translations, log-likelihood is used as the reranking score. In contrast, the technology according to this embodiment normalizes the log-likelihood by the length of the output sentence, thereby correcting the tendency to prioritize short output sentences. Furthermore, by taking into account the similarity based on the multilingual sentence embedding of the input sentence and the output sentence in the reranking score, translation accuracy can be further improved.

[0091] The following additional notes are provided regarding the above-described embodiments.

[0092] <Additional Notes> (Additional note 1) a search unit that extracts a plurality of similar translations to a source language input sentence from a set of target language sentences by performing a cross-language search using the source language input sentence as a query; an inference unit that generates a plurality of output candidate sentences by performing translation using the source language input sentence and each of the plurality of similar translations; a reranking unit that performs reranking on the plurality of output candidate sentences; An inference device comprising: (Additional note 2) The reranking unit performs reranking using a score obtained by normalizing the log likelihood of the output candidate sentence by the length of the output candidate sentence. 2. The inference device according to claim 1. (Additional note 3) The re-ranking unit performs re-ranking using a score having a logarithmic likelihood of an output candidate sentence and a similarity between the output candidate sentence and the source language input sentence. 3. An inference device according to claim 1 or 2. (Additional note 4) An inference method executed by an inference device, comprising: a search step of extracting a plurality of similar translations to the source language input sentence from a set of target language sentences by performing a cross-language search using the source language input sentence as a query; an inference step of generating a plurality of output candidate sentences by performing translation using the source language input sentence and each of the plurality of similar translations; a reranking step of reranking the plurality of output candidate sentences; An inference method comprising: (Additional note 5) A program for causing a computer to function as each part of the inference device described in any one of appendixes 1 to 3.

[0093] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0094] 100 training equipment 110 Training Data DB 120 Translation Memory DB 130 Search Department 140 Training Department 150 Translation Model DB 200 Reasoning device 210 Translation Memory DB 220 Search Department 230 Translation Model DB 240 Reasoning section 250 Reranking 260 Input section 270 Output section 1000 Drive Device 1001 Recording media 1002 Auxiliary storage 1003 Memory device 1004 CPU 1005 Interface device 1006 Display device 1007 Input Device 1008 Output Device< / sep> < / sep>

Claims

1. a search unit that extracts a plurality of similar translations to a source language input sentence from a set of target language sentences by performing a cross-language search using the source language input sentence as a query; an inference unit that generates a plurality of output candidate sentences by performing translation using the source language input sentence and each of the plurality of similar translations; a reranking unit that performs reranking on the plurality of output candidate sentences; An inference device comprising:

2. The reranking unit performs reranking using a score obtained by normalizing the log likelihood of the output candidate sentence by the length of the output candidate sentence. The inference device according to claim 1 .

3. The re-ranking unit performs re-ranking using a score having a logarithmic likelihood of an output candidate sentence and a similarity between the output candidate sentence and the source language input sentence. The inference device according to claim 1 .

4. An inference method executed by an inference device, comprising: a search step of extracting a plurality of similar translations to the source language input sentence from a set of target language sentences by performing a cross-language search using the source language input sentence as a query; an inference step of generating a plurality of output candidate sentences by performing translation using the source language input sentence and each of the plurality of similar translations; a reranking step of reranking the plurality of output candidate sentences; An inference method comprising:

5. A program for causing a computer to function as each unit of the inference device according to any one of claims 1 to 3.

Images