Context sentence determination system, machine translation device, and learning device
The context sentence determination system accurately identifies relevant context sentences for machine translation by evaluating word and sentence similarities, addressing the issue of incorrect context selection in conventional methods and enhancing translation accuracy.
Patent Information
- Application Number
- PCT/JP2024/014310
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-05-19
- Filing Date
- 2024-04-08
- Publication Date
- 2024-11-28
AI Technical Summary
Conventional context sentence determination methods in machine translation often inaccurately identify context sentences, leading to incorrect translation results, even when a sentence with a high translation evaluation value is not actually relevant to the translation target.
A context sentence determination system that acquires translation target sentences and multiple context candidate sentences, calculates context evaluation values using word meaning similarity and sentence meaning similarity models, and determines appropriate context sentences based on these evaluations to ensure accurate context selection for machine translation.
This approach effectively eliminates the determination of inappropriate context sentences, providing a context that truly relates to the translation target, thereby improving the accuracy of machine translation by considering the appropriate context for words with ambiguous meanings.
Smart Images

Figure JP2024014310_28112024_PF_FP_ABST
Abstract
Description
Context sentence determination system, machine translation device and learning device
[0001] The present invention relates to a context sentence determination system for determining a context to be used in machine translation, a machine translation device, and a learning device.
[0002] In machine translation, taking context into account during execution can sometimes make it possible to appropriately translate words with ambiguous meanings. For example, when translating the English sentence "He is famous scholar in this area" into Japanese, if "He teaches politics" is taken into account as context, "area" will be translated as "field." Here, context refers to a sentence with a strong semantic connection that contains information from the sentences before and after it, such as a document or dialogue. The number of sentences before the sentence to be translated that the context exists varies depending on the sentence to be translated.
[0003] Non-Patent Document 1 listed below describes a context sentence determination model (neural network) that determines an appropriate context for a sentence to be translated. This context sentence determination model is executed as a preprocessing step for translation execution, and when the sentence to be translated and multiple context candidate sentences are given as input, an appropriate context for the sentence to be translated is output. This context sentence determination model is generated by learning using the sentence to be translated and the context candidate sentences as input and a predetermined context as output.
[0004] In Non-Patent Document 1, a method for determining a context sentence during model training uses a translation evaluation value for the translation result when each context candidate sentence is used as the context. The translation evaluation value is an index representing translation accuracy, and the sentence that was considered when outputting a translation result with a high translation evaluation value is determined as the context. The background to using this index for determining the context is the assumption that when an appropriate context for the sentence to be translated is considered, it is possible to output a translation result with high accuracy.
[0005] Xiaomian Kang, Yang Zhao, Jiajun Zhang, Chengqing Zong, "Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning," [online], November 2020, Association for Computational Linguistics, [Retrieved April 19, 2023], Internet <URL: https: / / aclanthology.org / 2020.emnlp-main.175 / >
[0006] In machine translation, even if a sentence that is not actually context for the sentence to be translated is used as context, it may still be considered as context because it can produce a correct translation result and have a high translation evaluation value. For example, when translating the English sentence "He is a famous scholar in this area." into Japanese, even if "Good morning everyone." is considered as context, the English word "area" may be translated as "area" in Japanese, resulting in a high translation evaluation value. As a result, a sentence that is inappropriate as context may be determined as context. Therefore, it is possible that the machine translation will not be correct.
[0007] Therefore, an object of the present invention is to provide a context sentence determination system that can determine appropriate context sentences in machine translation, and a machine translation device and a learning device that utilize this system.
[0008] The context sentence determination system of the present invention includes an acquisition unit that acquires a sentence to be translated and a plurality of context candidate sentences, and a determination unit that determines one or more context sentences from the plurality of context candidate sentences using the meanings of the sentence to be translated and each of the plurality of context candidate sentences.
[0009] According to the present invention, the appropriate context sentence can be determined.
[0010] 1 is a schematic diagram illustrating the processing concept of a context sentence determination method. FIG. 1 is a functional block diagram illustrating the functional configuration of a context sentence determination system 10 of the present disclosure. FIG. 2 is a flowchart illustrating the flow of processing performed in the context sentence determination system 10 of embodiment 1. FIG. 3 is a schematic diagram of a context sentence determination method using a word semantic similarity calculation model 12a as a context evaluation value calculation unit 12. FIG. 4 is a diagram illustrating how semantic vectors are extracted from each word vector by BERT in the word semantic similarity calculation model 12a. FIG. 5 is a diagram illustrating how the cosine similarity of each semantic vector is calculated in the word semantic similarity calculation model 12a. FIG. 6 is a diagram illustrating how the word semantic similarity is calculated based on the cosine similarity and IDF value of each semantic vector in the word semantic similarity calculation model 12a. FIG. 7 is a diagram illustrating the similarity between each semantic vector output from BERT in the word semantic similarity calculation model 12a. FIG. 8 is a diagram illustrating how the word semantic similarity is calculated based on the cosine similarity of each semantic vector in the word semantic similarity calculation model 12a. FIG. 9 is a schematic diagram illustrating the processing concept of a context sentence determination method using the semantic similarity of the entire sentence between a translation target sentence and a context candidate sentence. FIG. 1 is a schematic diagram showing the processing concept of a context sentence determination method using coreference analysis. FIG. 2 is a functional block diagram showing the functional configuration of a context sentence determination system 10c of embodiment 4. FIG. 3 is a diagram showing a coreference relationship. FIG. 4 is a diagram showing an overview of an entire machine translation system. FIG. 5 is a functional block diagram showing the functional configuration of a machine translation device 20 of embodiment 5. FIG. 6 is a flowchart showing the flow of processing performed by the machine translation device 20 and the context sentence determination system 10 according to embodiment 5. FIG. 7 is a diagram showing an overview of an entire context sentence determination model generation method. FIG. 8 is a schematic diagram showing translation execution processing. FIG. 9 is a functional block diagram showing the functional configuration of a context sentence determination model learning device 30. FIG. 10 is a flowchart showing the flow of learning processing of the context sentence determination model learning device 30 according to embodiment 6. FIG. 11 is a block diagram showing the functional configuration of a context sentence determination system 10d. FIG. 12 is a diagram showing an example of the hardware configuration of the context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 according to an embodiment of the present disclosure.
[0011] The present disclosure will be described with reference to the accompanying drawings. Whenever possible, the same parts are designated by the same reference numerals and redundant description will be omitted.
[0012] [Embodiment 1] In one embodiment of the present disclosure, a context sentence determination device will be described with reference to Figures 1 to 3. Figure 1 is a schematic diagram illustrating the processing concept of a context sentence determination method. Figure 1(a) is a diagram illustrating an outline of processing by a context evaluation value calculation unit 12. In the present disclosure (including embodiments 2 to 6), it is assumed that a sentence to be translated is in English and is to be translated into Japanese, but the present disclosure is not limited to these languages.
[0013] 1(a) shows that when the sentence to be translated is "He is famous scholar in this area." and the candidate context sentence is "He teaches politics in his school," each sentence is input to the context evaluation value calculation unit 12, which outputs a context evaluation value of 37.88. This context evaluation value is expressed as a percentage, and the same is true for word similarity, which will be described later.
[0014] 1B is a correspondence table that associates a translation target sentence, a context candidate sentence, a context evaluation value, and a determination result indicating whether or not the sentence is a context. As shown in the figure, the translation target sentence indicates a sentence to be translated. The context candidate sentence is a candidate sentence to be referenced when the translation target sentence is translated.
[0015] 1B is a value output from the context evaluation value calculation unit 12. A context evaluation value is associated with each context candidate sentence.
[0016] FIG. 1( c) shows that five context candidate sentences for a sentence to be translated and the sentence to be translated are input to the context sentence determination system 10 to determine one or several context sentences. The one or several context sentences are sentences that are referenced during machine translation. The context sentence determination system 10 (including the context evaluation value calculation unit 12) generates the above-mentioned correspondence table 12x and determines one context sentence based on the context evaluation value of each context candidate sentence. This correspondence table 12x is generated by the context evaluation value calculation unit 12. In the present disclosure, "He teaches politics in his school.", which has the highest context evaluation value, is determined as the context sentence when translating the sentence to be translated. This makes it possible to determine an appropriate context sentence for one sentence to be translated. Note that the present disclosure is not limited to determining one context sentence with the highest context evaluation value, and multiple context sentences with a context evaluation value equal to or greater than a threshold may be determined.
[0017] Fig. 2 is a functional block diagram showing the functional configuration of a context sentence determination system 10 according to the present disclosure. The context sentence determination system 10 is a system for executing the process shown in Fig. 1. The context sentence determination system 10 includes an acquisition unit 11, a context evaluation value calculation unit 12, a context sentence determination unit 13, and an output unit 14.
[0018] The acquisition unit 11 is a part that acquires the translation target sentence and the context candidate sentences. The translation target sentence and the context candidate sentences are stored in advance in a machine translation target sentence DB (not shown), and a predetermined number of sentences are acquired as context candidate sentences, going back from the translation target sentence. In the present disclosure, five sentences are used, but this is not limited to this, and the number may be six or more, or four or less.
[0019] The context evaluation value calculation unit 12 is a part that calculates the evaluation value between the translation target sentence and the context candidate sentence, calculates the strength of the semantic connection between the two sentences (hereinafter referred to as the context evaluation value), and outputs the context evaluation value to a correspondence table 12x between the translation target sentence and the context candidate sentence. In this embodiment 1, a conceptual process is shown. Specific processes of the context evaluation value calculation unit 12 will be described in embodiments 2 to 4.
[0020] The context sentence determination unit 13 is a part that determines a context sentence based on the context evaluation value described in the correspondence table 12x.
[0021] The output unit 14 outputs the determined context sentence as a determination result. For example, the output unit 14 associates the translation target sentence with the context sentence and outputs the result to a training database or to a machine translation unit.
[0022] As shown in the figure, the context sentence determination system 10 acquires a sentence to be translated in language A and context candidate sentences in language A, and outputs the context of the sentence to be translated. In the present disclosure, English is used as language A and Japanese is used as language B, but this is not limited to this and the system can be applied to other languages.
[0023] FIG. 3 is a flowchart showing the flow of processing performed by the context sentence determination system 10 of the first embodiment.
[0024] The acquisition unit 11 of the context sentence determination system 10 acquires a translation target sentence (language A) and context candidate sentences (language A) (S101). The translation target sentence is a sentence to be translated when translation is executed, and the context candidate sentences represent multiple sentences that can serve as context sentences for the translation target sentence. In FIG. 1, the translation target sentence is "He is a famous scholar in this area.", and the context candidate sentences are a total of five sentences, such as "Good evening everyone." and "Welcome to the seminar."
[0025] In the context sentence determination system 10, the context evaluation value calculation unit 12 receives the translation target sentence and each of the context candidate sentences as input, calculates the strength of the semantic connection between the two sentences (hereinafter referred to as the context evaluation value), and outputs the calculated value to the context sentence determination unit 13 (S102). For example, in FIG. 1, the context evaluation value calculation unit 12 calculates and outputs a context evaluation value of 37.88 for the context candidate sentence "He teaches politics in his school." and the translation target sentence "He is famous scholar in this area.". Specific methods for calculating the context evaluation value will be described later in embodiments 2, 3, and 4.
[0026] Then, the context evaluation value calculation unit 12 outputs the context evaluation values of the translation target sentences for each of the context candidate sentences, and generates a correspondence table 12x.
[0027] The context sentence determination unit 13 determines one sentence to be used as the context based on the context evaluation value (S103). For example, the context with the highest context evaluation value is determined. For example, in FIG. 1, the context sentence determination unit 13 determines "He teaches politics in his school," which has the highest context evaluation value, as the context sentence. Note that the number of context sentences to be determined does not need to be limited to one sentence, and may be the top two sentences with the highest context evaluation value, sentences with a context evaluation value equal to or greater than a set threshold, etc.
[0028] In the context sentence determination system 10, the output unit 14 outputs the determined context sentence as a determination result.
[0029] The operation and effect of the context sentence determination system 10 in the first embodiment configured as described above will be described below. According to the first embodiment, it is possible to eliminate the possibility that an inappropriate sentence is determined as a context sentence, which was a problem in conventional context sentence determination methods, and to provide a sentence that truly resembles a context sentence as a context sentence. Therefore, by applying the context sentence determination system 10 before executing translation of contextual data or during training of the context sentence determination model 30b described below, it is possible to present appropriate context sentences to the user.
[0030] [Embodiment 2] Next, embodiment 2 of the present disclosure will be described. Embodiment 2 of the present disclosure differs from embodiment 1 in that the context evaluation value calculation unit 12 in Fig. 2 and step S102 in Fig. 3 specifically use the semantic similarity between words contained in the sentence to be translated and the context candidate sentence as the context evaluation value.
[0031] The following mainly describes the differences in configuration and operation from the first embodiment described above, and the same components as those in the first embodiment described above are denoted by the same reference numerals, and the description thereof will be omitted as appropriate.
[0032] In step S103, the context sentence determination system 10 functions as the word semantic similarity calculation model 12a in the context evaluation value calculation unit 12, and calculates the semantic similarity of words between sentences as a context evaluation value. The similarity becomes high when there are many words with similar meanings between two sentences.
[0033] 4 is a schematic diagram of a context sentence determination method using a word semantic similarity calculation model 12a as the context evaluation value calculation unit 12. In FIG. 4(a), the translation target sentence "He is famous scholar in this area." and the context candidate sentence "He teaches politics in his school." are input to the word semantic similarity calculation model 12a, which outputs a context evaluation value of 37.88. This word semantic similarity calculation model 12a is a calculation model that calculates the similarity in meaning of words contained in a sentence.
[0034] For example, the translation target sentence "He is a famous scholar in this area." and the context candidate sentence "He teaches politics in his school." contain words that can be used in academic contexts, such as "scholar," "teaches," and "school," as well as the word "He," which is common to both sentences. Due to the presence of these words, the word semantic similarity calculation model 12a outputs a context evaluation value of 37.88. In other words, the word semantic similarity calculation model 12a outputs a high similarity between two sentences that contain many words used in the same field.
[0035] The word semantic similarity calculation model 12a uses BERTScore, a representative method, to output an evaluation value. The reason for using this index as a context evaluation value is that the meanings of words contained in sentences with strong contextual connections are similar.
[0036] 4B shows a correspondence table 12x that associates the translation target sentence, the context candidate sentence, the word similarity output by the word semantic similarity calculation model 12a, and the determination result of whether or not the sentence is a context sentence. The context sentence determination unit 13 determines whether or not each context candidate sentence is a context for the translation target sentence based on the word similarity (e.g., the similarity with the largest value).
[0037] 4(c) shows that a translation target sentence and a context syntax are input to the context sentence determination system 10, and one context sentence is output. Note that, as described above, multiple context sentences may be determined and output.
[0038] The following describes the operation and effect of the context sentence determination system 10 in this embodiment 2. According to embodiment 2, when a sentence that truly seems to be a context contains many words that have similar meanings to the sentence to be translated, it can be correctly determined as a context.
[0039] FIG. 5 is a diagram showing details of the word similarity calculation procedure. The word semantic similarity calculation model 12a has BERT (Bidirectional Encoder Representations from Transformers). BERT is a natural language processing technology that is primarily capable of interpreting the meaning of a sentence. The BERTSocre is an index obtained using the semantic vectors of words extracted by BERT, and is an index for measuring the similarity between two documents.
[0040] As shown in the figure, the target sentence "He is a famous scholar in this area." and the context candidate sentence "He teaches politics in his school." are input to BERT. BERT calculates the word vectors x 1 , x 2 ... etc. are extracted on a word-by-word basis, and the semantic vector x' of the word is 1 , x' 2 ... and output it. In FIG. 5, BERT converts "He" (feature x 1 ) to generate a context-sensitive semantic vector x' 1Similarly, BERT extracts features x and y from the translation target sentence and context candidate sentence, and outputs semantic vectors x' and y' for each word.
[0041] FIG. 6 shows how the word semantic similarity calculation model 12a calculates the cosine similarity between semantic vectors output from BERT. FIG. 6(a) shows how the cosine similarity is calculated for all combinations of feature vectors by cross-multiplying them together. FIG. 6(b) shows the cosine similarity of each word combination between the translation target sentence and the context candidate sentence in a matrix format. The word semantic similarity calculation model 12a extracts the combination with the highest cosine similarity between the semantic vectors of each word (horizontal direction) in the translation target sentence and each word in the context candidate sentence.
[0042] In Figure 6(b), the cosine similarity between He in the sentence to be translated and He in the context candidate sentence is 0.87, which is the highest value extracted. Similar combination extraction is performed for all words in the sentence to be translated.
[0043] 7A and 7B are explanatory diagrams for calculating evaluation values using idf. Fig. 7A shows the cosine similarity between words (semantic vectors) in a sentence to be translated and a context candidate sentence, and the idf values of words in the sentence to be translated. Fig. 7B shows the calculation of evaluation values using idf values.
[0044] Here, the word semantic similarity calculation model 12a calculates the BERTScore according to Equation 1 using the similarity with the word determined above and the idf value. i , y i indicate the words in the translation target sentence and the context candidate sentence, respectively, and x' i , y′ i The semantic vector of maxx' is shown below. i T y' i In this formula, the similarity is calculated by changing j (the index of the context candidate sentence) by one for i (the index of the sentence to be translated), and then finding the maximum value, thereby calculating the maximum similarity. The idf (Inverse Document Frequency) value represents the reciprocal of the proportion of documents that contain a certain word in a document set. The more frequently a word appears in other documents, the smaller the idf value, and the less frequently a word appears in other sentences, the larger the idf value. In the present disclosure, the translation target sentence and each context candidate sentence correspond to the document set, and the idf value is calculated based on the frequency of appearance therein. Because the idf value represents the rarity of a word, the weight of the similarity of rare words is increased, allowing for calculation of similarity that takes into account the unique characteristics of the sentence.
[0045] Instead of the evaluation value using the idf value shown in Equation 1, it may be based on the average value of the similarity.
[0046] FIG. 8 is a diagram showing the similarity between each semantic vector output from BERT in the word semantic similarity calculation model 12a. As shown in FIG. 8(a), the word semantic similarity calculation model 12a extracts one word vector from the sentence to be translated and calculates the similarity (cosine similarity) with each of all words in the context candidate sentence. For each word in the sentence to be translated, the word semantic similarity calculation model 12a determines the word in the context candidate sentence that has the greatest similarity. For example, for the sentence to be translated, "scholar," in FIG. 8(b), "school" in the context candidate sentence is determined to be the word with the greatest similarity.
[0047] As for other words included in the sentence to be translated, as shown in FIG. 5, the word semantic similarity calculation model 12a calculates the semantic vector x' of the word in the sentence to be translated. 2 etc. are extracted in order, and the semantic vector x' 2 and the semantic vector y' of the context candidate sentence 1 In the example of FIG. 8(b), the similarity between the semantic vector x′ 4 scholar and the semantic vector y' 6 The similarity with school is output as a high value of 0.82.
[0048] The word semantic similarity calculation model 12a calculates the average value of the maximum similarity values obtained from each word, and sets this average value as the similarity between the translation target sentence and the context candidate sentence. The context sentence determination unit 13 determines the context sentence based on this average value. Note that other statistical indicators may be used instead of the average value.
[0049] 9A and 9B are diagrams showing the details of the calculation procedure for determining an evaluation value using the average value described in FIG. 8 above. FIG. 9A shows the cosine similarity between the words (semantic vectors) of the translation target sentence and the context candidate sentence. FIG. 9B is an explanatory diagram showing the calculation of an evaluation value using the average value. Here, the idf value as in FIG. 7 is not used, and only the average value of the maximum similarity is calculated as shown in Equation 2. In this way, the word semantic similarity calculation model 12a calculates the similarity between the translation target sentence and the context candidate sentence using the maximum similarity value obtained from each word and the inverse document frequency (IDF) value. The calculation formula is shown in Figure 7. Because the IDF value indicates the rarity of a word, the cosine similarity of rare words is weighted more heavily, allowing for the calculation of similarity that takes into account the unique characteristics of the sentence. The context sentence determination unit 13 determines the context sentence based on the calculated similarity between the sentences.
[0050] The word semantic similarity calculation model 12a calculates the average value of the maximum similarity values obtained from each word using the calculation formula shown in FIG.
[0051] Using the idf value allows the calculation of similarity that captures the characteristics of the text more effectively. For example, the importance of "scholar" and "school" can be increased, while the importance of "he," "in," and "is" can be decreased.
[0052] The context sentence determination unit 13 determines the context for the sentence to be translated based on the similarity calculated by the above method.
[0053] Next, a description will be given of the operation and effect of embodiment 2. According to embodiment 2, when a sentence that is truly likely to be a context contains many words having a similar meaning to the sentence to be translated, the sentence can be correctly determined as a context.
[0054] Next, a third embodiment of the present disclosure will be described. The third embodiment of the present disclosure differs from the first and second embodiments in that the context evaluation value calculation unit 12 and step S102 specifically use the semantic similarity of the entire sentence between the translation target sentence and the context candidate sentence as the context evaluation value.
[0055] The following mainly describes the differences in configuration and operation from the above-described first and second embodiments, and the same configurations as those in the above-described first and second embodiments are denoted by the same reference numerals, and the description thereof will be omitted as appropriate.
[0056] FIG. 10 is a schematic diagram illustrating the processing concept of a context sentence determination method using the semantic similarity of the entire sentence between a translation target sentence and a context candidate sentence. In this diagram, the context evaluation value calculation unit 12 of the context sentence determination system 10 functions as a sentence semantic similarity calculation model 12b, calculating the semantic similarity of the sentence as a context evaluation value. The sentence semantic similarity calculation model 12b includes SentenseBERT, which vectorizes sentences. Using SentenseBERT, it generates sentence vectors for the input sentences and outputs the similarity between the sentence vectors. This similarity is output so that it is higher when the meanings of the two sentences are similar. For example, as shown in FIG. 10(a), the sentence semantic similarity calculation model 12b inputs the translation target sentence "He is a famous scholar in this area." and the context candidate sentence "He teaches politics in this area.", outputs a sentence semantic similarity of 39.34, and generates a correspondence table 12x.
[0057] As shown in FIG. 10B, the sentence semantic similarity calculation model 12b associates the sentence semantic similarity with the translation target sentence and the context candidate sentence, and describes the sentence semantic similarity in the correspondence table 12x.
[0058] The context sentence determination unit 13 refers to this correspondence table 12x and determines the context candidate sentence corresponding to the highest sentence semantic similarity as the context sentence. In Figure 10(b), the context sentence determination unit 13 calculates the highest value of 39.34 for the translation target sentence "He is famous scholar in this area." and the context candidate sentence "He teaches politics in his school." as both having similar meanings that can be used in an academic context, and determines that this context candidate sentence is the context.
[0059] 10(c) shows that five context candidate sentences for the sentence to be translated are input to the context sentence determination system 10 to determine one or several context sentences. This context sentence determination system 10 determines one or several context sentences based on the sentence semantic similarity in the correspondence table. In the present disclosure, "He teaches politics in his school.", which has the highest sentence semantic similarity, is determined as one context sentence.
[0060] A representative method for calculating this evaluation value is SentenceBERT. SentenceBERT is a learning model trained to make the sentence vectors generated from a set of similar sentences closer to each other. In the present disclosure, this model is used to determine the similarity between two sentences. The reason for using this index as a context evaluation value is that sentences that have a strong semantic connection in the context are similar in meaning as sentences.
[0061] Next, the operation and effect of the context sentence determination system 10b according to the third embodiment of the present disclosure will be described. According to the third embodiment, an effect is obtained in which a truly context-like sentence can be correctly determined as a context when it has a similar meaning to the sentence to be translated. Therefore, it is possible to present an appropriate context using a more specific calculation method than the first embodiment. Furthermore, in the second embodiment, a sentence could be considered as a context if it contained similar words regardless of its meaning, but in the third embodiment, it is possible to take into account the meaning of the entire sentence.
[0062] [Fourth Embodiment] Next, a fourth embodiment of the present disclosure will be described. The fourth embodiment of the present disclosure differs from the first, second, and third embodiments in that, in step S102, the presence or absence of words in a coreference relationship with words in the sentence to be translated is used as the context evaluation value. A coreference relationship refers to the relationship between two noun expressions when the two expressions refer to the same discourse target. Coreference analysis is a process in natural language processing in which, when a document is input, words that represent the same entity are grouped into the same cluster. The input is a document, and the output is cluster information. The target is not limited to noun phrases. In this fourth embodiment, the input is a document that connects the sentence to be translated and context candidate sentences, and the output is cluster information of words in a coreference relationship within the document.
[0063] The following mainly describes the differences in configuration and operation from the above-described first, second, and third embodiments, and the same configurations as those in the above-described first, second, and third embodiments are designated by the same reference numerals, and the description thereof will be omitted as appropriate.
[0064] 11 is a schematic diagram showing the processing concept of a context sentence determination method using coreference analysis. In FIG. 11( a), the coreference analysis model 12c performs coreference analysis on a series (plurality) of immediately preceding context candidate sentences, including the sentence to be translated, and outputs whether or not the context candidate sentences have a coreference relationship with words in the sentence to be translated. In coreference analysis, by inputting data on multiple consecutive sentences, it is possible to extract a group of words that have a coreference relationship.
[0065] In Figure 11(a), the following six sentences are treated as one document for coreference analysis: "Good morning, everyone. Welcome to the seminar. Today's speaker is Tom. He is my friend. He teaches politics in his school. He is a famous scholar in this area."
[0066] The coreference analysis model 12c then extracts a group of words that refer to the same discourse phenomenon from these five sentences. Here, the coreference analysis model 12c determines that "politics" and "Tom" are in a coreference relationship with "this area" and "He" in the sentence to be translated, respectively, and that sentences containing them have a coreference relationship.
[0067] 11(b) shows a correspondence table 12x that lists the translation target sentence, the context candidate sentence, the presence or absence of co-reference information, and the determination result of whether or not it is a context sentence. The context sentence determination unit 13 determines whether or not it is a context sentence based on the presence or absence of co-reference information in this correspondence table 12x.
[0068] The reason behind using this index as a context evaluation value is the assumption that sentences that are strongly connected in terms of context contain words that are in a coreference relationship.
[0069] 11(c) shows that five context candidate sentences including the sentence to be translated are input to a context sentence determination system 10c to determine one context sentence. This context sentence determination system 10c determines one context sentence based on the presence or absence of coreference information in the correspondence table 12x. In the present disclosure, of the context candidate sentences determined to have coreference information, "He teaches politics in his school." is determined as the context sentence.
[0070] The following can be cited as criteria for determining one context sentence based on coreference information. For example, if the sentence to be translated contains a demonstrative pronoun such as "it" or "they" or a determiner such as "the" or "this" through morphological analysis of the coreference result, a context candidate sentence containing a noun phrase corresponding to the demonstrative pronoun or determiner may be determined as one context sentence. In FIG. 11( a), the coreference analysis model 12c actually outputs a result indicating that "He" in the sentence to be translated also has a coreference relationship with "Tom," as well as with "He" in "He is my friend," and with "He" in "He teaches politics in his school." However, by applying this method, it becomes possible to determine context sentences only for sentences that include details about demonstratives.
[0071] Furthermore, in combination with embodiments 2 and 3, word similarity (or sentence-level similarity) may also be calculated separately from the co-reference result, and the context candidate sentence with the highest context evaluation value with the translation target sentence among the context candidate sentences for which there is a co-reference result may be determined as one context sentence.
[0072] Alternatively, instead of determining a single context sentence, all of the context candidate sentences that have been output as having a coreference relationship with the sentence to be translated may be concatenated into a single sentence. For example, in the example shown in Figure 11, "Today's speaker is Tom. He teaches politics in his school." can be determined as a single context sentence.
[0073] 12 is a functional block diagram showing the functional configuration of a context sentence determination system 10c according to embodiment 4. The context sentence determination system 10c is a system for executing the processing shown in the figure. The context sentence determination system 10c includes an acquisition unit 11, a coreference analysis model 12c, a context sentence determination unit 13, and an output unit 14.
[0074] The acquisition unit 11c is a part that acquires a plurality of context candidate sentences and translation target sentences. In this fourth embodiment, the acquisition unit 11c acquires a series of sentences. Although not shown in FIG. 11( a), an identification flag is attached to each of the translation target sentence and the context candidate sentence so that they can be distinguished. The acquisition unit 11c identifies the identification flag and passes the translation target sentence and the context candidate sentence to the coreference analysis model 12c. This identification flag includes information indicating that the sentence is a translation target sentence and, for the context candidate sentence, indicating how many sentences before (n sentences before) it is from the translation target sentence. For example, since "Good evening, everyone." is a sentence that is five sentences before the translation target sentence, an identification flag indicating n=5 is assigned. Since "He teaches politics in his school." is one sentence before the translation target sentence, an identification flag indicating n=1 is assigned.
[0075] The coreference analysis model 12c is a part that outputs the coreference relationship between the translation target sentence and each context candidate sentence in a series of sentences. FIG. 13 is a diagram showing the coreference relationship between the context candidate sentence and the translation target sentence. As shown in FIG. 13, the coreference analysis model 12c performs coreference analysis on the sentence from "Good morning, everyone." to "He is a famous scholar in this area." as a single document, and outputs that "politics" and "Tom" are in a coreference relationship with "this area" and "He" in the translation target sentence, respectively. The coreference analysis model 12c outputs whether or not this relationship exists to the correspondence table 12x (see FIG. 11(b)).
[0076] In the present disclosure, the coreference analysis model 12c analyzes the coreference relationship between the translation target sentence and the context candidate sentence, and outputs cluster information of words in the coreference relationship.
[0077] The context sentence determination unit 13c determines one context sentence based on the presence or absence of coreference information in the correspondence table 12x obtained by the coreference analysis model 12c. In the present disclosure, the context sentence determination unit 13c determines "He teaches politics in his school." as the context sentence. Note that the presence or absence of coreference information may be "present" for multiple sentences. In this case, as described above, the multiple context candidate sentences may be treated as one sentence (for example, "Today's speaker is Tom. He teaches politics in his school."), or may be combined with word similarity or sentence similarity in other embodiments.
[0078] Next, the operation and effect of the context sentence determination system 10 according to the fourth embodiment of the present disclosure will be described. According to the fourth embodiment, when a sentence that truly seems to be a context contains a word that has a coreference relationship with a word in the sentence to be translated, the effect of being able to correctly determine the context can be obtained. Therefore, it is possible to present an appropriate context using a more specific calculation method than in the first embodiment. Furthermore, when differentiating between translations of demonstrative pronouns such as "it" and "they," the fourth embodiment can determine a sentence containing the words they refer to as the context, thereby enabling more accurate context determination than the second and third embodiments.
[0079] [Embodiment 5] Next, embodiment 5 of the present disclosure will be described. Embodiment 5 of the present disclosure relates to a machine translation device 20 that applies any one of the context sentence determination systems 10 (including 10a, 10b, and 10c) of embodiments 1 to 4, and determines a context sentence each time translation is executed, thereby performing translation in consideration of context.
[0080] The following mainly describes the differences in configuration and operation from the above-described first embodiment, and the same configurations as those in the above-described first embodiment are denoted by the same reference numerals, and the description thereof will be omitted as appropriate.
[0081] In this embodiment, a machine translation system to which the present invention is applied will be described with reference to Figures 14 to 17. Figure 14 is a diagram showing an overview of the entire machine translation system.
[0082] As shown in the figure, the context sentence determination system 10 inputs five context candidate sentences and determines one possible context for the sentence to be translated. Here, the context candidate sentence "He teaches politics in his school" is output as the context sentence for the sentence to be translated, "He is famous scholar in this area." Note that while the context sentence determination system 10 is described here as a representative example, it is also possible to use a context sentence determination system 10a equipped with a word semantic similarity calculation model 12a, or another context sentence determination system 10b, etc.
[0083] 15 is a functional block diagram showing the functional configuration of a machine translation device 20 according to the fifth embodiment. The machine translation device 20 includes an acquisition unit 21, a machine translation unit 22, and an output unit 23.
[0084] The acquisition unit 21 is a part that acquires a translation target sentence (language A) and a context sentence (language A) determined by the context sentence determination system 10 b. The translation target sentence is stored in a translation target sentence storage unit (not shown) and is acquired according to one context sentence determined by the context sentence determination system 10 b.
[0085] The machine translation unit 22 is a part that translates the sentence to be translated by referring to the determined one context sentence.
[0086] The output unit 23 is a part that outputs the translated sentence (language B).
[0087] FIG. 16 is a flowchart showing the flow of processing performed by the machine translation device 20 and the context sentence determination system 10 according to the fifth embodiment.
[0088] The acquisition unit 11 of the context sentence determination system 10 acquires a translation target sentence (language A) and a context candidate sentence (language A) (S201). The context sentence determination unit 13 of the context sentence determination system 10 determines one context sentence based on the translation target sentence (language A) and the context candidate sentence (language A) (S202). Then, the output unit 14 outputs the determined context sentence (language A) to the machine translation device 20.
[0089] The acquisition unit 21 of the machine translation device 20 acquires the translation target sentence and one context sentence (in language A) for the translation target sentence from the context sentence determination system 10, and the machine translation unit 22 performs machine translation of the translation target sentence taking the context sentence into consideration (S203).The output unit 23 then outputs the translation result (in language B) (S204).
[0090] Next, the operation and effect of the fifth embodiment will be described. According to the fifth embodiment, it is possible to obtain an effect that machine translation can be performed taking into account appropriate context. Therefore, it is possible to present to the user a correct translation result that is compared with the result of a normal machine translation, takes into account appropriate context, and translates words with ambiguous meanings.
[0091] [Sixth Embodiment] Next, a sixth embodiment of the present disclosure will be described. The sixth embodiment of the present disclosure differs from the fifth embodiment in that it uses a context sentence determination model (neural network) trained using data for determining context sentences in step S202 of Fig. 16 or the like. The context sentence determination model is a classification model trained using a translation target sentence and a context candidate sentence as input and a context determined from a context evaluation value as output.
[0092] In the sixth embodiment, this context sentence determination model is used to determine one context sentence from the sentence to be translated and multiple context candidate sentences, which differs from the fifth embodiment in that a context evaluation value is calculated and a context sentence is determined each time translation is performed.
[0093] The following mainly describes the differences in configuration and operation from the above-described first and fifth embodiments, and the same configurations as those in the above-described first and fifth embodiments are denoted by the same reference numerals, and their description will be omitted as appropriate.
[0094] In this embodiment, a method for generating a context sentence determination model will be described with reference to FIGS.
[0095] Fig. 17 is a diagram showing an overview of the entire method for generating a context sentence determination model. Fig. 17(a) shows the collection of training data for generating a context sentence determination model, and Fig. 17(b) shows a training method using a training DB.
[0096] As shown in FIG. 17(a), a context sentence is determined using the context sentence determination system 10 shown in any one of the first to fourth embodiments, and the current translation target sentence, multiple context candidate sentences, and the context sentence determined therefrom are stored in the learning DB 30a.
[0097] As shown in FIG. 17(b), the context sentence determination model learning device 30 uses the translation target sentences and context candidate sentences stored in the learning DB 30a as explanatory variables and the determined context sentence as the objective variable, and learns using known machine learning to generate a context sentence determination model 30b.
[0098] 18 is a schematic diagram showing the translation execution process. As shown in the figure, when a plurality of context candidate sentences and a translation target sentence are input, the context sentence determination model 30b outputs one context sentence determined from the context candidate sentences. Note that in this disclosure, for the sake of convenience, the sentences are assumed to be the same as those used in the learning process, but the present disclosure can of course also be applied to different sentences.
[0099] The machine translation device 20 acquires the context sentence output from the context sentence determination model 30b and translates the translation target sentence by referring to it. In the present disclosure, the machine translation device 20 acquires the context sentence "He teaches politics in his school." and the translation target sentence "He is famous scholar in this area." to be referenced during translation, and performs machine translation. In the present disclosure, the Japanese translation result can be "He is a famous scholar in this field." In other words, "in this area" in English can be translated as "in this field" in Japanese, rather than as "in this region." When translating the English word "area" into Japanese, it can mean either "region" or "field," and it is necessary to determine which meaning is appropriate based on the context.
[0100] 19 is a functional block diagram showing the functional configuration of the context sentence determination model learning device 30. The context sentence determination model learning device 30 includes an acquisition unit 31, a learning unit 32, and a context sentence determination model 30b.
[0101] The acquisition unit 31 retrieves training data from the training DB 30a. In the present disclosure, the training translation target sentences and multiple training context candidate sentences are used as explanatory variables, and a determined training context sentence is acquired as a target variable. These training translation target sentences, training context candidate sentences, and training context sentences (collectively referred to as training data) are based on the translation target sentences, context candidate sentences, and context sentences obtained according to, for example, embodiments 1 to 4.
[0102] The learning unit 32 is a part that performs machine learning based on the acquired learning data to generate the context sentence determination model 30b. That is, the learning unit 33 performs machine learning using the translation target sentence (language A) and the context candidate sentence (language A) as explanatory variables and a corresponding context sentence as a target variable to generate the context sentence determination model 30b.
[0103] The context sentence determination model 30b is a machine learning model that outputs one context sentence when a translation target sentence and multiple context candidate sentences are input.
[0104] FIG. 20 is a flowchart showing the flow of the learning process of the context sentence determination model learning device 30 according to the sixth embodiment.
[0105] The acquisition unit 31 acquires a translation target sentence (language A), a context candidate sentence (language A), and one corresponding context sentence from the learning DB 30a (S301). The learning unit 33 performs machine learning using the translation target sentence (language A) and the context candidate sentence (language A) as explanatory variables and the corresponding one context sentence as a target variable to generate a context sentence determination model 30b (S302).
[0106] By using this context sentence determination model 30b in the context sentence determination unit 13 of S202 in Figure 2, it is possible to obtain machine translation results that take appropriate context into consideration. Figure 21 is a block diagram showing the functional configuration of a context sentence determination system 10d. This context sentence determination system 10d is configured to include an acquisition unit 11d, a context sentence determination model 30b, and an output unit 14. The acquisition unit 11d is a part that acquires a translation target sentence and multiple context candidate sentences. When the translation target sentence and multiple context candidate sentences are input, the context sentence determination model 30b outputs one or more appropriate context sentences that take meaning into consideration. The output unit 14 outputs one or more context sentences.
[0107] Next, the operation and effect of embodiment 6 will be described. According to embodiment 6, by using a model trained in advance, it is possible to acquire context more quickly than in embodiment 5, where context is acquired during translation execution. Therefore, in a method that requires high processing speed, such as a real-time translation method, it is possible to reduce delays during translation execution and present the user with a correct translation result that takes appropriate context into consideration.
[0108] [Operational Effects of the Context Sentence Determination System, Machine Translation Device, and Learning Device According to the Present Disclosure] In the context sentence determination system 10 according to the present disclosure shown in the first embodiment, the acquisition unit 11 acquires a translation target sentence and a plurality of context candidate sentences. Then, the context sentence determination unit 13 determines one or more context sentences from the plurality of context candidate sentences using the meanings of the translation target sentence and each of the plurality of context candidate sentences.
[0109] For example, in the first embodiment, the context sentence determination system 10 includes a context evaluation value calculation unit 12 as a similarity deriving unit. The context evaluation value calculation unit 12 derives a context evaluation value taking into account the meanings of the translation target sentence and multiple context candidate sentences. The context evaluation value indicates the similarity between the sentences. The context sentence determination unit 13 determines one or more context sentences based on the context evaluation value.
[0110] In the second embodiment, the context evaluation value calculation unit 12 functions as a word semantic similarity calculation model 12 a, which outputs a plurality of inter-word similarities between words included in the sentence to be translated and words included in each of a plurality of context candidate sentences.
[0111] More specifically, the word semantic similarity calculation model 12a derives the inter-word similarity for each combination of a word contained in the sentence to be translated and a word contained in each of the multiple context candidate sentences, and outputs the average value as the similarity.
[0112] Alternatively, the similarity of the entire sentence may be determined without considering the similarity on a word-by-word basis. For example, as in the third embodiment, the similarity deriving unit may function as the sentence semantic similarity calculation model 12b. This sentence semantic similarity calculation model 12b outputs the inter-sentence similarity of the meaning of each sentence unit between the translation target sentence and each of the multiple context candidate sentences.
[0113] These similarities are calculated using vectors. That is, the word semantic similarity calculation model 12 a or the sentence semantic similarity calculation model 12 b derives semantic vectors of the sentences to be translated and the context candidate sentences or of the words contained in the sentences, and derives similarities based on the semantic vectors.
[0114] In this way, it is possible to determine the context sentence to be referenced when machine translating a sentence to be translated based on the similarity between sentences or words. A context sentence determined based on the similarity between sentences or words is a context sentence that takes into account the meaning of the sentence. This is a natural way of determining context, and it is possible to determine an appropriate context sentence.
[0115] Another method is processing based on coreference relationships, as described in embodiment 4. In embodiment 4, a context sentence determination system 10c includes a coreference analysis model 12c unit that acquires coreference relationships between a translation target sentence and each of multiple context candidate sentences. A context sentence determination unit 13c determines one or more context sentences based on these coreference relationships.
[0116] In this way, if a sentence that truly resembles a context contains a word that has a coreference relationship with a word in the sentence to be translated, it can be correctly determined as the context. Therefore, it is possible to present an appropriate context using a more specific calculation method than in the first embodiment.
[0117] Furthermore, by using a context determination model that has been trained in advance on the context sentences determined as described above, appropriate context sentences can be determined in the same manner as described above.
[0118] That is, the context sentence determination system 10d in embodiment 6 includes a context sentence determination model 30b, and when a translation target sentence and multiple context candidate sentences are input, it outputs one or several context sentences. This context sentence determination model 30b is trained using one or several training context sentences determined using the meanings of the training translation target sentence and each of the multiple training context candidate sentences. The training context sentences determined in this case are the context sentences obtained by the context sentence determination system 10 described in embodiments 1 to 4.
[0119] This context sentence determination model 30b is trained using the translation target sentence and each of the plurality of context candidate sentences as explanatory variables and the training context sentence as a target variable.
[0120] That is, the context sentence determination model learning device 30 includes a learning DB 30a (learning data storage unit) that stores the learning context sentences, learning context candidate sentences, and learning translation target sentences acquired in embodiments 1 to 4. The learning unit 32 then generates a context sentence determination model 30b through machine learning, using the learning translation target sentences and learning context candidate sentences in the learning data as explanatory variables and the learning context as a target variable.
[0121] In this way, the context sentence determination model 30b is generated, and thus it is possible to appropriately determine a context sentence taking semantics into consideration.
[0122] In the fifth embodiment, a machine translation device 20 is shown which performs machine translation using the context and the translation target sentence determined in the context sentence determination systems 10, 10a, and 10d in the first to fourth embodiments.
[0123] In this machine translation device 20, an acquisition unit 21 acquires a translation target sentence and a determined context sentence, and a machine translation unit 22 translates the translation sentence by machine translation based on the context sentence.
[0124] This allows translation to be based on appropriate contextual sentences, and avoids the use of meaningless and inappropriate contextual sentences, resulting in more accurate translation.
[0125] The context sentence determination system, machine translation device, and learning device of the present disclosure have the following configurations.
[0126] [1] A context sentence determination system comprising: an acquisition unit that acquires a sentence to be translated and a plurality of context candidate sentences; and a determination unit that determines one or more context sentences from the plurality of context candidate sentences using the meanings of the sentence to be translated and each of the plurality of context candidate sentences.
[0127] [2] The context sentence determination system according to [1], further comprising a similarity derivation unit that derives a similarity between the translation target sentence and the plurality of context candidate sentences, taking into account their respective meanings, and the determination unit determines one or more context sentences based on the similarity.
[0128] [3] The context sentence determination system according to [2], wherein the similarity derivation unit derives a plurality of inter-word similarities between words included in the sentence to be translated and words included in each of the plurality of context candidate sentences.
[0129] [4] The context sentence determination system according to [3], wherein the similarity derivation unit derives the inter-word similarity for each combination of a word included in the sentence to be translated and a word included in each of the plurality of context candidate sentences, and derives an average value thereof as the similarity.
[0130] [5] The context sentence determination system according to any one of [2] to [4], wherein the similarity derivation unit derives semantic vectors of the sentences to be translated and the context candidate sentences or words included in the sentences, and derives similarity based on the semantic vectors.
[0131] [6] The context sentence determination system according to [1], further comprising a coreference analysis unit that acquires coreference relationships between the translation target sentence and each of the plurality of context candidate sentences, and the determination unit determines one or more context sentences based on the coreference relationships.
[0132] [7] The context sentence determination system according to [1], wherein the determination unit is a context sentence determination model that inputs the translation target sentence and the plurality of context candidate sentences and outputs one or several context sentences, and the context sentence determination model is trained using one or several training context sentences determined using the meanings of a training translation target sentence and each of the plurality of training context candidate sentences.
[0133] [8] The context sentence determination system according to [7], wherein the context sentence determination model is trained using the translation target sentence and each of the plurality of context candidate sentences as explanatory variables and the training context sentence as a target variable.
[0134] [9] A machine translation device comprising: a translation acquisition unit that acquires a context sentence and a translation target sentence determined in the context sentence determination system according to any one of [1] to [8]; and a machine translation unit that translates the translation sentence by machine translation based on the context.
[0135]
[10] A learning device comprising: a training data storage unit that stores training context sentences, training context candidate sentences, and training translation target sentences determined by the context sentence determination system according to any one of [1] to [6]; and a learning unit that generates a context sentence determination model by machine learning using the training translation target sentences and the training context candidate sentences in the training data as explanatory variables and the training context sentences as a target variable.
[0136] The block diagrams used to explain the above embodiments show functional blocks. These functional blocks (components) are realized by any combination of hardware and / or software. Furthermore, the method for realizing each functional block is not particularly limited. That is, each functional block may be realized using a single device that is physically or logically coupled, or may be realized using two or more physically or logically separated devices that are connected directly or indirectly (e.g., via wire, wirelessly, etc.) and these multiple devices. The functional block may also be realized by combining the single device or multiple devices with software.
[0137] Functions include, but are not limited to, judgment, determination, assessment, calculation, computation, processing, derivation, investigation, search, confirmation, reception, transmission, output, access, resolution, selection, selection, establishment, comparison, assumption, expectation, consideration, broadcasting, notifying, communicating, forwarding, configuring, reconfiguring, allocating, mapping, and assignment. For example, a functional block (component) that performs transmission is called a transmitting unit or transmitter. As mentioned above, there are no particular limitations on how these functions are implemented.
[0138] For example, the context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 according to an embodiment of the present disclosure may function as a computer that performs processing of the context sentence determination method, the machine translation method, or the learning method of the present disclosure. Figure 22 is a diagram showing an example of the hardware configuration of the context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 according to an embodiment of the present disclosure. The above-described context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 may be physically configured as a computer device including a processor 1001, a memory 1002, a storage 1003, a communication device 1004, an input device 1005, an output device 1006, a bus 1007, etc.
[0139] In the following description, the term "apparatus" can be interpreted as a circuit, a device, a unit, etc. The hardware configurations of the context sentence determination system 10, the machine translation apparatus 20, and the context sentence determination model learning apparatus 30 may be configured to include one or more of the apparatuses shown in the figures, or may be configured to exclude some of the apparatuses.
[0140] Each function of the context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 is realized by loading predetermined software (programs) onto hardware such as a processor 1001 and a memory 1002, causing the processor 1001 to perform calculations, control communications via the communication device 1004, and control at least one of reading and writing of data in the memory 1002 and the storage 1003.
[0141] The processor 1001 controls the entire computer by running, for example, an operating system. The processor 1001 may be configured as a central processing unit (CPU) including an interface with peripheral devices, a control unit, an arithmetic unit, a register, etc. For example, the above-mentioned context evaluation value calculation unit 12 and context sentence determination unit 13 may be realized by the processor 1001.
[0142] The processor 1001 also loads programs (program codes), software modules, data, etc. from at least one of the storage 1003 and the communication device 1004 into the memory 1002 and executes various processes in accordance with these programs. The programs used are those that cause a computer to execute at least some of the operations described in the above-described embodiments. For example, the context evaluation value calculation unit 12 and the context sentence determination unit 13 may be implemented by a control program stored in the memory 1002 and running on the processor 1001, and similar implementations may be used for other functional blocks. While the above-described various processes have been described as being executed by a single processor 1001, they may also be executed simultaneously or sequentially by two or more processors 1001. The processor 1001 may be implemented by one or more chips. The programs may also be transmitted from a network via a telecommunications line.
[0143] The memory 1002 is a computer-readable recording medium and may be configured by, for example, at least one of a read-only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a random access memory (RAM), etc. The memory 1002 may also be referred to as a register, a cache, a main memory (primary storage device), etc. The memory 1002 may store executable programs (program codes), software modules, etc. for implementing the context determination method, machine translation method, or learning method according to an embodiment of the present disclosure.
[0144] Storage 1003 is a computer-readable recording medium, and may be composed of at least one of, for example, an optical disk such as a CD-ROM (Compact Disc ROM), a hard disk drive, a flexible disk, a magneto-optical disk (e.g., a compact disk, a digital versatile disk, a Blu-ray (registered trademark) disk), a smart card, a flash memory (e.g., a card, a stick, a key drive), a floppy (registered trademark) disk, a magnetic strip, etc. Storage 1003 may also be referred to as an auxiliary storage device. The above-mentioned storage medium may be, for example, a database, a server, or other appropriate medium including at least one of memory 1002 and storage 1003.
[0145] The communication device 1004 is hardware (transmission / reception device) for communicating between computers via at least one of a wired network and a wireless network, and is also referred to as, for example, a network device, a network controller, a network card, a communication module, etc. The communication device 1004 may be configured to include a high-frequency switch, a duplexer, a filter, a frequency synthesizer, etc. to realize, for example, at least one of frequency division duplex (FDD) and time division duplex (TDD).
[0146] The input device 1005 is an input device (e.g., a keyboard, a mouse, a microphone, a switch, a button, a sensor, etc.) that receives input from the outside. The output device 1006 is an output device (e.g., a display, a speaker, an LED lamp, etc.) that outputs to the outside. Note that the input device 1005 and the output device 1006 may be integrated into one device (e.g., a touch panel).
[0147] Furthermore, each device, such as the processor 1001 and the memory 1002, is connected by a bus 1007 for communicating information. The bus 1007 may be configured using a single bus, or may be configured using different buses between each device.
[0148] Furthermore, the context sentence determination system 10, the machine translation device 20, and the context sentence determination model learning device 30 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a field programmable gate array (FPGA), and some or all of the functional blocks may be realized by the hardware. For example, the processor 1001 may be implemented using at least one of these pieces of hardware.
[0149] The notification of information is not limited to the aspects / embodiments described in the present disclosure and may be performed using other methods. For example, the notification of information may be performed by physical layer signaling (e.g., Downlink Control Information (DCI) and Uplink Control Information (UCI)), higher layer signaling (e.g., Radio Resource Control (RRC) signaling, Medium Access Control (MAC) signaling, broadcast information (Master Information Block (MIB) and System Information Block (SIB))), other signals, or a combination thereof. Furthermore, the RRC signaling may be referred to as an RRC message, and may be, for example, an RRC Connection Setup message, an RRC Connection Reconfiguration message, or the like.
[0150] The order of the procedures, sequences, flowcharts, etc. of each aspect / embodiment described in this disclosure may be changed unless it is consistent. For example, the methods described in this disclosure present elements of various steps using an example order, and are not limited to the particular order presented.
[0151] Input and output information may be stored in a specific location (for example, memory) or may be managed using a management table. Input and output information may be overwritten, updated, or added to. Output information may be deleted. Input information may be sent to another device.
[0152] The determination may be made based on a value represented by one bit (0 or 1), a Boolean value (true or false), or a numerical comparison (e.g., comparison with a predetermined value).
[0153] The aspects / embodiments described in this disclosure may be used alone, in combination, or switched depending on the implementation. Notification of predetermined information (e.g., notification that "X is true") is not limited to explicit notification, but may be implicit (e.g., not notifying the predetermined information).
[0154] Although the present disclosure has been described in detail above, it is clear to those skilled in the art that the present disclosure is not limited to the embodiments described herein. The present disclosure can be implemented in modified and altered forms without departing from the spirit and scope of the present disclosure as defined by the claims. Therefore, the description of the present disclosure is intended to be illustrative and does not have any limiting meaning on the present disclosure.
[0155] Software shall be construed broadly to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, threads of execution, procedures, functions, etc., whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise.
[0156] Software, instructions, information, etc. may also be transmitted or received over a transmission medium. For example, if software is transmitted from a website, server, or other remote source using wired technologies (such as coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL)), and / or wireless technologies (such as infrared, microwave), then these wired and / or wireless technologies are included within the definition of transmission media.
[0157] The information, signals, etc. described in this disclosure may be represented using any of a variety of different technologies. For example, data, instructions, commands, information, signals, bits, symbols, chips, etc. that may be referred to throughout the above description may be represented by voltages, currents, electromagnetic waves, magnetic fields or magnetic particles, optical fields or photons, or any combination thereof.
[0158] Note that terms described in this disclosure and terms necessary for understanding this disclosure may be replaced with terms having the same or similar meanings. For example, at least one of a channel and a symbol may be a signal (signaling). Furthermore, a signal may be a message. Furthermore, a component carrier (CC) may be called a carrier frequency, a cell, a frequency carrier, etc.
[0159] Furthermore, the information, parameters, etc. described in the present disclosure may be expressed using absolute values, may be expressed using relative values from a predetermined value, or may be expressed using other corresponding information. For example, a radio resource may be indicated by an index.
[0160] The names used for the above-described parameters are not intended to be limiting in any way. Furthermore, the mathematical expressions using these parameters may differ from those explicitly disclosed in this disclosure. The various channels (e.g., PUCCH, PDCCH, etc.) and information elements may be identified by any suitable names, and therefore the various names assigned to these various channels and information elements are not intended to be limiting in any way.
[0161] In this disclosure, the terms "Mobile Station (MS)," "user terminal," "User Equipment (UE)," "terminal," and the like may be used interchangeably.
[0162] A mobile station may also be referred to by those skilled in the art as a subscriber station, mobile unit, subscriber unit, wireless unit, remote unit, mobile device, wireless device, wireless communication device, remote device, mobile subscriber station, access terminal, mobile terminal, wireless terminal, remote terminal, handset, user agent, mobile client, client, or some other suitable terminology.
[0163] As used in this disclosure, the terms "determining" and "determining" may encompass a wide variety of actions. "Determining" and "determining" may include, for example, judging, calculating, computing, processing, deriving, investigating, looking up, searching, inquiring (e.g., searching in a table, database, or other data structure), ascertaining, and the like. "Determining" and "determining" may also include receiving (e.g., receiving information), transmitting (e.g., sending information), input, output, accessing (e.g., accessing data in memory), and the like. Furthermore, "judgment" and "decision" can include regarding resolving, selecting, choosing, establishing, comparing, etc. as having been "judged" or "decided." In other words, "judgment" and "decision" can include regarding some action as having been "judged" or "decided." Furthermore, "judgment (decision)" can be interpreted as "assuming," "expecting," "considering," etc.
[0164] The terms "connected," "coupled," or any variation thereof, refer to any direct or indirect connection or coupling between two or more elements, and may include the presence of one or more intermediate elements between two elements that are "connected" or "coupled" to each other. The coupling or connection between elements may be physical, logical, or a combination thereof. For example, "connected" may be read as "access." As used in this disclosure, two elements may be considered to be "connected" or "coupled" to each other using one or more wires, cables, and / or printed electrical connections, as well as electromagnetic energy having wavelengths in the radio frequency range, microwave range, and optical (both visible and invisible) range, as some non-limiting and non-exhaustive examples.
[0165] As used in this disclosure, the phrase "based on" does not mean "based only on," unless expressly stated otherwise. In other words, the phrase "based on" means both "based only on" and "based at least on."
[0166] As used in this disclosure, any reference to an element using a designation such as "first," "second," etc. does not generally limit the quantity or order of those elements. These designations may be used in this disclosure as a convenient method of distinguishing between two or more elements. Thus, a reference to a first and a second element does not imply that only two elements may be employed or that the first element must in some way precede the second element.
[0167] When the terms "include," "including," and variations thereof are used in this disclosure, these terms are intended to be inclusive, similar to the term "comprising." Furthermore, when the term "or" is used in this disclosure, it is not intended to be an exclusive or.
[0168] In this disclosure, where articles are added by translation, such as a, an, and the in English, the disclosure may include that the nouns following these articles are in the plural form.
[0169] In the present disclosure, the term "A and B are different" may mean "A and B are different from each other." The term may also mean "A and B are each different from C." Terms such as "separate" and "coupled" may also be interpreted in the same way as "different."
[0170] 10, 10a, 10b, 10c, 10d...context sentence determination system, 11, 11c, 11d...acquisition unit, 12...context evaluation value calculation unit, 12a...word semantic similarity calculation model, 12b...sentence semantic similarity calculation model, 12c...coreference analysis model, 13, 13c...context sentence determination unit, 14...output unit, 20...machine translation device, 21...acquisition unit, 22...machine translation unit, 23...output unit, 30...context sentence determination model learning device, 31...acquisition unit, 32...learning unit, 33...learning unit, 30b...context sentence determination model
Claims
1. A context sentence determination system comprising: an acquisition unit that acquires a translation target sentence and a plurality of context candidate sentences; and a determination unit that determines one or more context sentences from the plurality of context candidate sentences using the meanings of the translation target sentence and each of the plurality of context candidate sentences.
2. The context sentence determination system according to claim 1, further comprising a similarity derivation unit that derives a similarity between the sentence to be translated and the plurality of context candidate sentences, taking into account their respective meanings, wherein the determination unit determines one or more context sentences based on the similarity.
3. The context sentence determination system according to claim 2, wherein the similarity derivation unit derives a plurality of inter-word similarities between words included in the sentence to be translated and words included in each of the plurality of context candidate sentences.
4. The context sentence determination system according to claim 3, wherein the similarity derivation unit derives the inter-word similarity for each combination of a word included in the sentence to be translated and a word included in each of the plurality of context candidate sentences, and derives the average value as the similarity.
5. The context sentence determination system according to claim 2, wherein the similarity derivation unit derives semantic vectors of the sentences to be translated and the context candidate sentences or words contained in the sentences, and derives the similarity based on the semantic vectors.
6. The context sentence determination system according to claim 1, further comprising a coreference analysis unit that acquires coreference relationships between the translation target sentence and each of the plurality of context candidate sentences, and the determination unit determines one or more context sentences based on the coreference relationships.
7. The context sentence determination system according to claim 1, wherein the determination unit is a context sentence determination model that inputs the translation target sentence and the multiple context candidate sentences and outputs one or several context sentences, and the context sentence determination model is trained using one or several training context sentences determined using the meanings of a training translation target sentence and each of the multiple training context candidate sentences.
8. The context sentence determination system according to claim 7, wherein the context sentence determination model is trained using the translation target sentence and each of the plurality of context candidate sentences as explanatory variables and the learning context sentence as a target variable.
9. A machine translation device comprising: a translation acquisition unit that acquires a context sentence and a translation target sentence determined in the context sentence determination system described in claim 1; and a machine translation unit that translates the translation sentence by machine translation based on the context.
10. A learning device comprising: a learning data storage unit that stores learning context sentences, learning context candidate sentences, and learning translation target sentences determined by the context sentence determination system described in claim 1; and a learning unit that generates a context sentence determination model by machine learning using the learning translation target sentences and the learning context candidate sentences in the learning data as explanatory variables and the learning context sentences as a target variable.
Citation Information
Patent Citations
Tapter machine translation system and method capable of selecting context information
CN113095091A
Statement translation method and device, computer equipment and storage medium
CN113392658A
Machine translation method, machine translation model training method and device, and storage medium
WO2021082518A1
