JP2022002090A - Information processing method, apparatus, device, and computer readable storage media - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an information processing method, an apparatus, an apparatus and a computer-readable storage medium for maintaining the consistency of word translation in a sentence. SOLUTION: In an information processing method, when a first information to be processed is acquired and S101, and a first substitute term is included in the first information to be processed, a first from one or a plurality of candidate word sets. Select the first candidate word set containing the substitute term of S102 (each of one or more candidate word sets includes a candidate word having at least two meanings matching and the expressions do not match), according to the constraint model. The first constraint result translated corresponding to the first alternative word in the candidate word set of 1 is acquired, and the translation result of the information to be processed is modified according to S103 and the first constraint result. By doing so, S104 to generate the second information to be processed. [Selection diagram] Fig. 1

Description

The present application relates to the field of information processing, specifically to information processing methods, devices, devices and computer-readable storage media.

When linguistic translation is done, it is very important to ensure the consistency of all translated words in a sentence, as words with the same semantics may have different expressions in the document. Maintaining consistent word translations in sentences is especially important when translating long documents (especially legal documents).

In view of the above problems, the present disclosure provides information processing methods, devices, devices and computer readable storage media.

According to one aspect of the present disclosure, when the first information to be processed is acquired and the first substitute term is included in the first information to be processed, one or a plurality of candidate word sets are used. A first candidate word set containing the first alternative term is selected, and each of the one or more candidate word sets is restricted to include candidate words having at least two meanings matching and expressions not matching. By the model, the first constraint result translated corresponding to the first alternative word in the first candidate word set is acquired, and the first process is performed according to the first constraint result. By modifying the translation result of the information to be processed, the information processing method including the generation of the second information to be processed is provided.

According to one aspect of the present disclosure, obtaining the first constraint result translated corresponding to the first alternative word in the first candidate word set by the constraint model is the first constraint. Based on pairing each candidate word in the candidate word set with one or more translation results of each candidate word to form a translation pair, the characteristics of each translation pair, and / or the representation of the semantic space of each translation pair. And to acquire the first constraint result.

According to one aspect of the present disclosure, the constraint model is a supervised first constraint model, the first constraint based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. Acquiring the result includes acquiring the first constraint result based on the characteristics of each translation pair by the first constraint model, and the first constraint model with the teacher is based on training data. Obtained by learning.

According to one aspect of the present disclosure, the constraint model is an unsupervised second constraint model, the first constraint based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. Obtaining the result includes obtaining the first constraint result based on the representation of the semantic space of each translation pair by the second constraint model, and the second constraint model without the teacher is No need to learn in advance.

According to one aspect of the present disclosure, it is the center of the semantic space of all translation pairs to obtain the first constraint result based on the representation of the semantic space of each translation pair by the second constraint model. It involves acquiring the first constraint result based on the distance between the representation and the representation of the semantic space of each translation pair.

According to one aspect of the present disclosure, the constraint model is a third constraint model including a first constraint model with supervision and a second constraint model with no supervision, and features and / or each of the translation pairs. Acquiring the first constraint result based on the representation of the semantic space of the translation pair is N (N is 2 or more) based on the representation of the semantic space of each translation pair by the second constraint model. Based on the characteristics of each translation pair in the second candidate word set by the acquisition of the second candidate word set containing the translation pair (which is an integer of) and the first constraint model, the first The first constraint model with the teacher is obtained by learning from the training data, and the second constraint model without the teacher does not need to be learned in advance.

According to one aspect of the present disclosure, the distance is an Euclidean distance.

According to one aspect of the present disclosure, the characteristics of the translation pair are the frequency with which one or more translations were selected, whether or not they were recently selected, the length of the translation, and whether or not they were included in the formal word. Includes semantic relevance.

According to one aspect of the present disclosure, the first information to be processed is part of the first processed text, and the one or more candidate word sets are generated based on the first processed text. Or generated based on one or more pieces of information to be processed prior to the first piece of information to be processed.

According to one aspect of the present disclosure, the neural network is trained and based on the first processed text or based on one or more pieces of information to be processed prior to the first information to be processed. Generate the one or more candidate word sets.

According to one aspect of the present disclosure, a second process is performed by modifying the translation result of the information to be processed in the first process by using the constrained decoding process according to the first constraint result. Generate information to be done.

According to one aspect of the present disclosure, words that do not exist in the user personalized information are deleted from the first candidate word set before the first constraint result is acquired by the constraint model.

According to one aspect of the present disclosure, the user personalized information is constructed based on one or more of the user's translation history, the user's style trends, and the field of translation.

According to one aspect of the present disclosure, the third information to be processed is translated to generate the fourth information to be processed, and the fourth information to be processed includes a third alternative term. Then, a third candidate word set containing the third alternative word is selected from one or more candidate word sets, and each of the one or more candidate word sets has at least two meanings and is expressed. The information to be processed is obtained by acquiring the third constraint result in the third candidate word set by the constraint model and including the candidate words that do not match, and according to the third constraint result. By modifying it, a fifth information processing method including the generation of information to be processed is provided.

According to one aspect of the present disclosure, obtaining the third constraint result in the third candidate word set by the constraint model is a feature of each candidate word in the third candidate word set and / or each. It includes acquiring the third constraint result based on the representation of the semantic space of the candidate word.

According to one aspect of the present disclosure, the constraint model is a supervised first constraint model, based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. Therefore, to acquire the third constraint result is to acquire the third constraint result based on the characteristics of each candidate word in the third candidate word set by the first constraint model. Including, the first constraint model with the teacher is obtained by learning from the training data.

According to one aspect of the present disclosure, the constraint model is a second constraint model without a teacher, based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. The acquisition of the third constraint result includes the acquisition of the third constraint result based on the representation of the semantic space of each candidate word by the second constraint model, without the teacher. The second constraint model does not need to be trained in advance.

According to one aspect of the present disclosure, obtaining the third constraint result based on the semantic space representation of each candidate word by the second constraint model is a central representation of the semantic space of all candidate words. Includes obtaining the third constraint result based on the distance between and the semantic space representation of each candidate word.

According to one aspect of the present disclosure, the constraint model is a third constraint model including a first constraint model with supervision and a second constraint model with no supervision, and features and / or each of the candidate words. Acquiring the third constraint result based on the representation of the semantic space of the candidate word is M (M is 2 or more) based on the representation of the semantic space of each candidate word by the second constraint model. Based on the characteristics of each candidate word in the fourth candidate word set by the acquisition of the fourth candidate word set containing the candidate words (which is an integer) and the first constraint model, the third candidate word set is used. The first constraint model with the teacher is obtained by learning from the training data, including the acquisition of the constraint result, and the second constraint model without the teacher does not need to be learned in advance.

According to one aspect of the present disclosure, the third processed information is part of the second processed text, and the one or more candidate word sets are in the translated text of the second processed text. Generated based on, or based on the translated text of one or more pieces of information to be processed prior to the third piece of information to be processed.

According to one aspect of the present disclosure, the first information acquisition unit to be processed for acquiring the first information to be processed and the first information to be processed include the first alternative term. The purpose is to select a first candidate word set containing the first alternative term from one or more candidate word sets, and each of the one or more candidate word sets has at least two meanings. A first candidate word set selection unit containing candidate words that match and the expressions do not match, and a first constraint result translated corresponding to the first alternative term in the first candidate word set by a constraint model. The second information to be processed is generated by modifying the translation result of the first information to be processed according to the first constraint result acquisition unit for acquiring the first constraint result and the first constraint result. Provided is an information processing apparatus including a second information generation unit to be processed.

According to one aspect of the present disclosure, a fourth information generation unit to be processed for translating the third information to be processed and generating the fourth information to be processed, and the fourth processing. When the information to be included includes a third substitute word, the third candidate word set containing the third substitute word is selected from one or more candidate word sets, and the one or more candidate words are selected. Each of the candidate word sets of the third candidate word set contains at least two candidate words whose meanings match and the expressions do not match, and the third constraint result in the third candidate word set is obtained by the constraint model. The third constraint result acquisition unit for acquisition and the fifth information to be processed are generated by modifying the fourth information to be processed according to the third constraint result. Provided is an information processing apparatus including an information generation unit to be processed in 5.

According to one aspect of the present disclosure, the information processing method is executed when the computer-readable instruction is executed by the processor, including a processor and a memory for storing computer-readable instructions. Acquires the first information to be processed, and when the first information to be processed includes the first substitute term, the first substitute word set including the first substitute term from one or a plurality of candidate word sets. One candidate word set is selected, and each of the one or more candidate word sets contains words whose meanings match and expressions do not match, and the constraint model allows the first candidate word set in the first candidate word set. By acquiring the first constraint result translated corresponding to the substitute term of 1 and modifying the translation result of the information to be processed according to the first constraint result, the second constraint result is obtained. Provide information processing equipment including generating information to be processed.

According to one aspect of the present disclosure, a computer-readable storage medium for storing a computer-readable program is provided, and the program causes a computer to execute the information processing method according to any one of the above. ..

In the above aspects of the present disclosure, by constraining the alternative terms with a constraint model and modifying the translation of the alternative terms in the information, the translations of all specific alternative terms in the document are matched, and the accuracy and expertise of the translation. To improve.

By describing the embodiments of the present disclosure in more detail in conjunction with the accompanying drawings, the above and other purposes, features, and advantages of the present disclosure will become more apparent. The drawings provide a further understanding of the embodiments of the present disclosure, form part of the specification, and are used in conjunction with the embodiments of the present disclosure to illustrate the present disclosure and are not intended to limit the present disclosure. In the drawings, the same reference symbol generally represents the same component or step.
It is a flowchart of the information processing method based on the source end by the Example of this disclosure. FIG. 2a is an exemplary schematic diagram of acquiring one or more candidate word sets based on a dictionary according to an embodiment of the present disclosure. FIG. 2b is an exemplary schematic diagram of acquiring one or more candidate word sets based on a dictionary according to an embodiment of the present disclosure. It is an exemplary schematic diagram showing one or more candidate word sets based on the neural network according to the embodiment of the present disclosure. It is a flowchart of the method of acquiring the first constraint result by the constraint model by the embodiment of this disclosure. It is an exemplary schematic diagram which acquires the 1st constraint result by the 1st constraint model by the Example of this disclosure. It is a flowchart of the method of acquiring the 1st constraint result by the 2nd constraint model by the Example of this disclosure. It is an exemplary schematic diagram which acquires the 1st constraint result by the 2nd constraint model by the Example of this disclosure. It is a flowchart of the method of acquiring the 1st constraint result by the 3rd constraint model by the Example of this disclosure. It is an exemplary schematic diagram of an information processing method based on a source end according to an embodiment of the present disclosure. It is another exemplary schematic diagram of the information processing method based on the source end according to the embodiment of this disclosure. It is a flowchart of another information processing method based on the target end by the Example of this disclosure. It is a flowchart of the method of acquiring the 3rd constraint result by the 2nd constraint model by the Example of this disclosure. It is a flowchart of the method of acquiring the 3rd constraint result by the 3rd constraint model by the Example of this disclosure. It is an exemplary schematic diagram of the information processing method based on the target end by the Example of this disclosure. It is another exemplary schematic diagram of the information processing method based on the target edge by the Example of this disclosure. It is a schematic diagram of the translation result by the Example of this disclosure. It is a functional block diagram of the information processing apparatus according to the Example of this disclosure. It is a functional block diagram of another information processing apparatus according to the Example of this disclosure. It is a schematic diagram of the information processing apparatus according to the Example of this disclosure. It is a figure of the example of the hardware composition of the electronic device which concerns on embodiment of this disclosure.

Hereinafter, the technical plan in the embodiments of the present disclosure will be described more clearly and completely together with the accompanying drawings in the embodiments of the present disclosure. Obviously, the examples described are only a part of the examples of the present disclosure and not all of them. All other examples obtained by one of ordinary skill in the art without creative effort based on the examples of the present disclosure shall be within the scope of protection of this application.

In this application, flowcharts are used to describe the steps of the method according to the embodiments of the present application. It should be understood that the previous step or the subsequent step is not always performed accurately in order. Alternatively, the various steps can be processed in reverse order or simultaneously. At the same time, other operations can be added to these processes and one or more steps can be removed from these processes.

First, the information processing method 100 for realizing the embodiment of the present disclosure will be described with reference to FIG. The present disclosure aligns the translations of all first terminology in a sentence by constraining the first terminology with a constraint model and modifying the translation of the first terminology in the information, and the translation accuracy. Improve sex and expertise.

Examples of the present disclosure and examples thereof will be described in detail together with the drawings.

At least one embodiment of the present disclosure provides information processing methods, devices, equipment and computer readable storage media. Hereinafter, the information processing provided by at least one embodiment of the present disclosure will be described, but is not limited to, by some exemplary examples and examples, and as will be described below, these will be described when there is no mutual contradiction. Different features of the embodiments and examples can be combined with each other to give new examples and examples, which also fall within the scope protected by the present disclosure.

Hereinafter, the information processing method based on the source end according to the embodiment of the present disclosure will be described with reference to FIGS. 1 to 10.

First, with reference to FIG. 1, an information processing method based on the source end according to the embodiment of the present disclosure will be described. The method can be completed automatically by a computer or the like. For example, the method can be applied to translate information or text. For example, the information processing acquisition method can be realized by software, hardware, firmware or any combination thereof, and is loaded by the processor of a device such as a mobile phone, a tablet, a laptop, a desktop computer, or a web server. Will be executed.

As shown in FIG. 1, the information processing method includes the following steps S101 to S104.

In step S101, the first information to be processed is acquired.

In step S102, when the first substitute word is included in the first information to be processed, the first candidate word set containing the first substitute word is selected from one or more candidate word sets. Each of the one or more candidate word sets includes candidate words whose meanings match at least two and whose expressions do not match.

In step S103, the constraint model acquires the first constraint result translated corresponding to the first alternative term in the first candidate word set.

In step S104, the second information to be processed is generated by modifying the translation result of the first information to be processed according to the first constraint result.

Regarding step S101, for example, the first information to be processed may be one paragraph or one text to be translated, or one phrase may be the text to be translated. It may be in any language form (eg, Chinese, English, Japanese, etc.), but is not limited herein. According to the information processing method of the present disclosure, the information to be processed first can be translated into a desired text (for example, Chinese is translated into desired English, Japanese, etc., but is not limited thereto), and , Match translations of alternative terms that have the same or similar meanings.

Regarding step S102, for example, the first pronoun may have a noun or pronoun whose meaning and expression do not match. For example, the first alternative term may be included in the first candidate word set, and the first candidate word set may include candidate words having at least two meanings matching and expressions not matching. Selecting a first candidate word set containing the first alternative term from one or more candidate word sets is called a co-reference query.

以上、第１の処理すべき情報、第１の代用語及び第１の候補単語集合に関する説明は、一例であり、第１の処理すべき情報には複数の候補単語集合を含む可能性があると理解されたく、ここで限定しない。

As described above, the description regarding the first information to be processed, the first alternative term, and the first candidate word set is an example, and the first information to be processed may include a plurality of candidate word sets. I want to understand that, so I don't limit it here.

For example, the information to be processed first is a part of the first processing text, and the one or more candidate word sets are based on the first processing text or the first processing. It may be generated based on one or more information to be processed prior to the information to be processed.

As an example, the first processed text may be one text to be translated, and when the first information to be processed is a sentence in the one text to be translated, the translation of the one text. One or more candidate word sets can be generated based on the text to be written.

Alternatively, as another example, the first processed text may be one text to be translated, and the first information to be processed is a sentence in the one text to be translated. , One or more candidate word sets can be generated based on one or more sentences before the sentence.

For example, one or a plurality of candidate word sets can be obtained based on a word translation in a dictionary or the like. As shown in FIGS. 2A and 2B, it is a schematic diagram of an exemplary example of acquiring one or more candidate word sets based on word translation in a dictionary according to an embodiment of the present disclosure.

As shown in FIGS. 2 (a) and 2 (b), when performing a Chinese-English translation, one English corresponds to a plurality of Chinese, and one Chinese also corresponds to a plurality of English. I have something to do. For example, when translating English into Chinese, one English (for example, "Docomo Beijing Labs" in FIG. 2 (a)) is translated into multiple Chinese (for example, "Toshima" in FIG. 2 (a)). "Beijing Research Institute", "DOCOMO Beijing Research Institute", "DOCOMO Beijing Research Institute") may be supported. Further, when translating Chinese into English, one Chinese (for example, "Toshima Beijing Research Institute" in FIG. 2 (b)) is in a plurality of Chinese (for example, in FIG. 2 (b)). "Docomo Beijing Labs", "Docomo Beijing communications laboratory co. Ltd.", "DBL"). This is because inconsistencies in translation results occur, and information processing based on a dictionary or the like can process only entity words (for example, nouns) in the dictionary, and cannot process pronouns contained in the information.

The present disclosure provides a method of obtaining a candidate word set based on a neural network. In this method, a candidate word set of all words (including nouns and pronouns) contained in the information can be obtained by an existing neural network. It should be understood that it is also possible to obtain a candidate word set based on other methods (eg, statistical methods, etc.), and is not limited here.

As an example, the trained neural network may be based on the one or more information to be processed, based on the first processed text, or based on one or more pieces of information to be processed prior to the first information to be processed. Multiple candidate word sets can be generated.

For example, the neural network may construct a candidate word set by acquiring terms and / or pronouns whose semantic meanings match and expressions do not match from the input information based on semantic relationships, implications, and the like. As an example, the neural network pairs and scores the word, the word vector (or the representation of the semantic space), and the word vector of another word, and the word with the highest score is the candidate word set of the word. It may be taken as a candidate word in.

As shown in FIG. 3, it is a schematic diagram of an exemplary example of acquiring one or more candidate word sets based on a neural network according to an embodiment of the present disclosure. A plurality of candidate words are called co-reference words whose meanings match and expressions do not match, and a method of acquiring a candidate word set can be called a "co-reference word discovery (or co-reference analysis)" method. In general, the method of finding a co-reference word is to pair and score all the words (nouns or noun collocations) before the phrase.

It is easily understood that neural networks can employ different neural network structures and include, but are not limited to, convolutional neural networks, recurrent neural networks (RNNs), and the like. The convolutional neural network includes, but is not limited to, a U-Net neural network, ResNet, DenseNet, and the like.

The method of acquiring one or more candidate word sets based on the neural network shown in FIG. 3 above is an example, and one or more candidate word sets are acquired by another method (for example, a statistical method). It can be done, but it is not limited here.

After acquiring one or more candidate word sets, the first candidate word set containing the first alternative term can be selected from the one or more candidate word sets. Next, in step S103, the constraint model acquires the first constraint result translated corresponding to the first substitute term in the first candidate word set.

Hereinafter, a method of obtaining the first constraint result according to the embodiment of the present disclosure will be described with reference to FIGS. 4 to 8.

As shown in FIG. 4, it is a flowchart of the method 200 for acquiring the first constraint result by the constraint model according to the embodiment of the present disclosure. For example, as shown in FIG. 4, it is the first candidate to acquire the first constraint result translated corresponding to the first alternative word in the first candidate word set by the constraint model. Each candidate word in the word set is paired with one or more translation results of each candidate word to form a translation pair (S201), based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. , Acquiring the first constraint result (S202) can be included.

As shown in FIG. 5, it is a schematic diagram of an exemplary example in which the first constraint result is acquired by the first constraint model according to the embodiment of the present disclosure.

For example, the constraint model is a supervised first constraint model, and acquiring the first constraint result based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair is possible. The first constraint model can include acquiring the first constraint result based on the characteristics of each translation pair, and the supervised first constraint model can be obtained by learning from the training data. Be done.

For example, the characteristics of a translation pair are the frequency with which the translation was selected, whether it was recently selected, the length of the translation, whether it was included in the formal word, the semantic relevance, and the embedded expression of the word in the phrase. It may contain one or more of the above. The feature may be obtained from a user dictionary, big data statistics, etc., but is not limited thereto. It should be understood that the features of the translation pair are not limited to one or more of the above features, and other features may be added as needed, but are not limited here.

As shown in FIG. 5, when translating Chinese into English, if the first alternative term is "Toshima Beijing Research Institute", the corresponding first candidate word set is {Toshima If it is Beijing Research Institute, Beijing Research Institute, DOCOMO Beijing Research Institute}, the translation result corresponding to the candidate word set is {Docomo Beijing Labs, Docomo Beijing communications laboratory co. Ltd. , DBL}.

First, {Toshina Beijing Research Institute, Beijing Research Institute, DOCOMO Beijing Research Institute} and {Docomo Beijing Labs, Docomo Beijing Communications laboratory co. Ltd. , DBL} are paired to form a translation pair 71. Next, for each of all the translation pairs 71, feature extraction 72 is performed using an existing appropriate method. For example, for each translation pair, the frequency at which the translation was selected, whether it was recently selected, the length of the translation, whether it was included in the formal word, and (1) the characteristics of the semantic relevance may be extracted. .. For example, as shown in FIG. 5, the characteristics of the extracted translation pair "Toshima Beijing Institute-Docomo Beijing Labs" are: frequency of translation selection: 28, (2) recently selected or not: 0, (3) Translation length: 3, (4) Whether or not it is included in the formal word: 0, (5) Semantic relevance: 3.12. Next, the score 74 of each translation pair is calculated based on each feature. Further, the acquired scores are passed through, for example, a neural network 75 including softmax, and the probability 76 corresponding to each score is acquired. Finally, the translation corresponding to the translation pair with the highest selection probability 76 is set as the first constraint result. For example, when the probability of the translation pair "Docomo Beijing Communications laboratory co. Ltd." is the highest, "Docomo beijing communications laboratory co. Ltd." is used as the first alternative term "Toshima". This is the first constraint result of "Laboratory".

For example, the supervised first constraint model described above may be obtained by pre-training with labeled training data. The above supervised first constraint model may be a neural network model, a statistical method model, or the like, but is not limited thereto.

For example, in FIG. 5, k1, k2, k3, k4, and k5 may be weights corresponding to each feature in the supervised first constraint model, or may be obtained by learning a neural network by a loss function. good. It is easily understood that the supervised first constraint model according to the embodiments of the present disclosure may employ different neural network configurations, including, but not limited to, convolutional neural networks, recurrent neural networks (RNNs), and the like. .. The convolutional neural network includes, but is not limited to, a U-Net neural network, ResNet, DenseNet, and the like. The present disclosure is not limited to translating Chinese into English, and the present disclosure makes any desired translation, for example, Chinese-Japanese translation, English-Japanese translation, German-Chinese translation. It should be understood that it can be applied in some cases.

As shown in FIG. 6, it is a flowchart of the method 300 which acquires the 1st constraint result by the 2nd constraint model by the embodiment of this disclosure. As shown in FIG. 7, it is a schematic diagram of an exemplary example in which the first constraint result is acquired by the second constraint model according to the embodiment of the present disclosure.

For example, the constraint model may be a second constraint model without a teacher, and the first constraint result is acquired based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. What is done can include obtaining the first constraint result based on the representation of the semantic space of each translation pair by the second constraint model, the second constraint model without the teacher. No need to learn in advance.

For example, the above unsupervised second constraint model may be, but is not limited to, a text rank, an embedded center, and the like.

As an example, according to the second constraint model, obtaining the first constraint result based on the representation of the semantic space of each translation pair is the central representation of the semantic space of all translation pairs and each translation pair. Acquiring the first constraint result can be included based on the distance from the representation of the semantic space.

As shown in FIG. 6, obtaining the first constraint result based on the distance between the central representation of the semantic space of all translation pairs and the representation of the semantic space of each translation pair is all translations. Get the central representation of the semantic space of a pair (S301), get the distance between the central representation of the semantic space of all translation pairs and the representation of the semantic space of each translation pair (S302), and get the translation with the shortest distance. It can include selecting the translation result corresponding to the pair as the first constraint result (S303).

For example, as shown in Fig. 7, when translating Chinese into English, the first alternative term is "Toshima Beijing Research Institute", and the corresponding first candidate word set is {Toshima Beijing Research. However, if it is Beijing Research Institute, DOCOMO Beijing Research Institute}, the translation result corresponding to the candidate word set is {Docomo Beijing Labs, Docomo Beijing communications laboratory co. Ltd. , DBL}.

First, {Toshina Beijing Research Institute, Beijing Research Institute, DOCOMO Beijing Research Institute} and {Docomo Beijing Labs, Docomo Beijing Communications laboratory co. Ltd. , DBL} are paired with each other to obtain a translation pair 81. Next, for each of the translation pairs 81, the representation 82 of the semantic space is extracted using an existing appropriate method (eg, an existing neural network, etc., not limited thereto), and each translation pair is extracted. Gets the representation 83 of the semantic space of. As shown in FIG. 7, the representation of the semantic space of the extracted translation pair "Toshima Beijing Institute-Docomo Beijing Labs" is X1 = [4.511, 0.768, 3.886, 4.08, 1 .604]. Next, the central representation 84 of the semantic space of all translation pairs is acquired. For example, the semantic space representation of all translation pairs may be averaged to obtain the central representation of the semantic space of all translation pairs. The distance between the semantic space representation of each translation pair and the central representation of the semantic space is then obtained (eg, the distance may be Euclidean distance, Manhattan distance, Mahalanobis distance, etc.). Finally, the translation result corresponding to the translation pair with the shortest distance is selected as the first constraint result. As shown in FIG. 7, since the distance corresponding to X3 is the smallest, "Docomo beijing communications laboratory co. Ltd." is used as the first constraint result 85 of the first alternative term "Toshina Beijing Research Institute". ..

It is understood that the method of obtaining the central representation and distance of the semantic space is not limited to that with reference to FIG. 7 above, and other appropriate methods can be adopted to obtain the central representation and distance of the semantic space. sea bream.

As shown in FIG. 8, it is a flowchart of the method 400 for acquiring the first constraint result by the third constraint model according to the embodiment of the present disclosure. The third constraint model includes a first constraint model with supervised learning and a second constraint model with no supervised learning, and the first constraint model with supervised learning is obtained by learning from training data and is unsupervised. The second constraint model does not need to be trained in advance. The method comprises the following steps S401-S402.

In step S401, the second constraint model acquires a second candidate word set containing N translation pairs based on the representation of the semantic space of each translation pair, where N is an integer of 2 or more.

In step S402, the first constraint model acquires the first constraint result based on the characteristics of each translation pair in the second candidate word set.

For example, for step S401, the distance between the central representation of the semantic space of all translation pairs and the representation of the semantic space of each translation pair is obtained based on the method described in FIG. 6 above, and then the distance. Are sorted in ascending order, and the N translation pairs may be selected as the second candidate word set. Next, for the second candidate word set, for example, the first constraint result is acquired by the first constraint model shown in FIG.

For example, after sorting in ascending order of distance, the translation pair corresponding to the previous two second distances is referred to as a second candidate word set (eg, translation pair "Dokomo Beijing Institutes laboratory co. "Ltd." And "Dokomo Beijing Labs") are selected. Next, with the two selected translation pairs as inputs, the first constraint model described with reference to FIG. 5 is based on the characteristics of each translation pair in the second candidate word set. Get the constraint result.

Although the method of acquiring the first constraint result according to the embodiment of the present disclosure has been described above with reference to FIGS. 4 to 8, the first constraint result is obtained before the first constraint result is acquired by the constraint model. Words that do not exist in the user personalized information may be deleted from the candidate word set of.

For example, the user personalized information may be constructed based on one or more of the user's translation history, the user's style trends, and the field of translation. For example, prior to information processing (eg, translation) using such methods, users may use their style trends (eg, based on written / verbal, gender, age, identity, etc.), areas of translation (eg, translation). For example, the desired rule can be automatically selected based on the law, news, medical care, intellectual property, etc., and then the constraint result can be optimized based on the rule. Instead, the user's favorite translation style is learned from the user's translation history, user information, and then the constraint result is optimized based on the style before information processing (eg, translation) is performed using the method. It may be converted. It is an example to optimize the constraint result based on the above user personalization information, and other valid methods can be adopted to optimize the constraint result, but it should be understood that it will not be explained again here.

Next, returning to FIG. 1, in step S104, the translation result of the first information to be processed is modified according to the first constraint result to generate the second information to be processed. ..

For example, information to be processed second by modifying the translation result of the information to be processed first by using constrained decoding processing according to the first constraint result. Can be generated. The constrained decryption process can modify the information to be processed first by using the first constraint result on the premise that the sentence sentence is kept fluent. The first method of modifying the information to be processed is not limited to that, and other known techniques can be used to modify the translation result, but it should be understood that it will not be explained again here.

Hereinafter, an example of the information processing method based on the source end according to the embodiment of the present disclosure will be described with reference to FIGS. 9 to 10.

As described above, the first processed text may be one text to be translated, and the first processed information may be a sentence in the one text to be translated. One or more candidate word sets can be generated based on the text to be translated. Alternatively, one or more candidate word sets may be generated based on the one or more sentences preceding the sentence.

As shown in FIG. 9, it is an exemplary schematic diagram that generates one or more candidate word sets based on the text to be translated, and FIG. 10 is one or more before the sentence. It is an exemplary schematic diagram which generates one or more candidate word sets based on a sentence.

As shown in FIGS. 9 to 10, when the first substitute term is included in the first information to be processed for the first information to be processed (that is, the current sentence 21), a co-reference query is first made. Selects the first candidate word set containing the first alternative term, (ie, the related candidate word set 23), and then by the constraint model, the first candidate word set in the first candidate word set. By acquiring the first constraint result 25 translated corresponding to the substitute term of 1, and modifying the translation result 27 of the first information to be processed 26 according to the first constraint result 25. , Generates second information to be processed (ie, modified translation result 28).

In FIG. 9, one or a plurality of candidate word sets 31 are obtained by the method of referring to FIG. 3 above (for example, co-reference word discovery 30) based on the text to be translated (that is, the full sentence 29). Generate. Instead, in FIG. 10, one or more candidate word sets 31 are presented in FIG. 10 by the method of reference to FIG. 3 above (eg, co-reference word discovery 33) based on one or more sentences 32 before the sentence. Generate.

As described above, the information processing method based on the source end according to the embodiment of the present disclosure has been described with reference to FIGS. 1 to 10. Hereinafter, the information processing method based on the target end according to the embodiment of the present disclosure will be described with reference to FIGS. 11 to 15.

First, the information processing method 500 based on the target end according to the embodiment of the present disclosure will be described with reference to FIG. The method can be completed automatically by a computer or the like. For example, the method can be applied to translate information or text. For example, the information processing acquisition method can be realized by software, hardware, firmware or any combination thereof, and is loaded by the processor of a device such as a mobile phone, a tablet, a laptop, a desktop computer, or a web server. Will be executed.

As shown in FIG. 11, the information processing method includes the following steps S501 to S504.

In step S501, the third information to be processed is translated to generate the fourth information to be processed.

In step S502, when the third substitute word is included in the fourth information to be processed, the third candidate word set containing the third substitute word is selected from one or more candidate word sets. Each of the one or more candidate word sets includes candidate words whose meanings match at least two and whose expressions do not match.

In step S503, the constraint model acquires the third constraint result in the third candidate word set.

In step S504, the fifth information to be processed is generated by modifying the fourth information to be processed according to the third constraint result.

Regarding step S501, for example, the third information to be processed may be one paragraph or one text to be translated, or one phrase may be the text to be translated. It may be in any language form (eg, Chinese, English, Japanese, etc.), but is not limited herein. The information processing method of the present disclosure allows a third piece of information to be processed to be translated into the desired text (eg, Chinese is translated into the desired English, Japanese, etc., but is not limited herein), and , Match translations of alternative terms that have the same or similar meanings.

For step S502, for example, the third pronoun may have a noun or pronoun that matches its meaning and does not match the expression. For example, the third alternative term may be included in the third candidate word set, and the third candidate word set may include candidate words whose meanings match at least two and whose expressions do not match. The meanings of the third substitute term and the third candidate word set are the same as the meanings of the first substitute term and the first candidate word set in FIG. 1 above, and are not explained again here.

For example, the third processed information is part of the second processed text, and the one or more candidate word sets are generated or based on the translated text of the second processed text. It is generated based on the translated text of one or more pieces of information to be processed prior to the third piece of information to be processed.

As an example, the second processed text may be one text to be translated, and the third information to be processed is a sentence in the one text to be translated, and the third processed text is described. It is possible to generate one or more candidate word sets based on the translated text of one or more information to be processed before the information to be processed.

Alternatively, as another example, the second processed text may be a text to be translated, and the third information to be processed is a sentence in the text to be translated, said. It is possible to generate one or more candidate word sets based on the translated text of the second processed text.

The present disclosure provides a method of obtaining a candidate word set based on a neural network. In this method, a candidate word set for all words (nouns, pronouns) included in the information can be obtained by a conventional neural network.

As an example, the trained neural network can generate the one or more candidate word sets.

For example, the neural network can construct a candidate word set by acquiring terms and / or pronouns whose semantic meanings match and expressions do not match from the input information based on semantic relationships, implications, and the like. As an example, the neural network may pair and score the word vector of a word with the word vector of another word, and take the word with the highest score as a candidate word in the candidate word set of the word.

It is easily understood that the neural networks according to the embodiments of the present disclosure can adopt different neural network configurations, including, but not limited to, convolutional neural networks, recurrent neural networks (RNNs), and the like. The convolutional neural network includes, but is not limited to, a U-Net neural network, ResNet, DenseNet, and the like.

The method of acquiring one or more candidate word sets based on the neural network is an example, and one or more candidate word sets can be acquired by other methods, but the present invention is not limited thereto.

After acquiring one or more candidate word sets, a third candidate word set containing the third alternative term can be selected from the one or more candidate word sets. Next, in step S503, the third constraint result in the third candidate word set is acquired by the constraint model.

Hereinafter, a method of obtaining the third constraint result according to the embodiment of the present disclosure will be described with reference to FIGS. 12 to 15.

For example, to obtain the third constraint result in the third candidate word set by the constraint model is a feature of each candidate word in the third candidate word set and / or a representation of the semantic space of each candidate word. Based on this, it can include acquiring the third constraint result.

For example, the constraint model is a first constrained model with a teacher, and the third is based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. Acquiring the constraint result of the above can include acquiring the third constraint result based on the characteristics of each candidate word in the third candidate word set by the first constraint model. The first constraint model with supervision is obtained by training with training data.

For example, the characteristics of a candidate word are the frequency with which the translation was selected, whether it was recently selected, the length of the translation, whether it was included in the formal word, the semantic relevance, and the embedded expression of the word in the phrase (embedding). It may contain one or more of the above. The feature can be obtained from a user dictionary, big data statistics, etc., but is not limited here. It should be understood that the features of the translation pair are not limited to one or more of the above features and other features may be added as needed, but are not limited here.

For example, when translating Chinese into English, if the third alternative term contained in the fourth information to be processed generated by translation is "Docomo Beijing Labs", the corresponding third Candidate word sets are {Docomo Beijing Labs, Docomo Beijing communications laboratory co. Ltd. , DBL}.

First, the translation set {Docomo Beijing Labs, Docomo Beijing communications laboratory co. Ltd. , DBL}, feature extraction is performed using an existing appropriate method. For example, for each translation, the frequency of selection of the translation, whether it was recently selected, the length of the translation, whether it is included in the formal word, and the characteristics of the semantic relevance may be extracted. Next, the score of each translation is calculated based on each feature. In addition, the acquired scores are passed through a neural network including softmax to acquire the probability corresponding to each score. Finally, the translation with the highest probability is selected as the third constraint result. For example, when the probability of "DOCOMO Beijing communications laboratory co. Ltd." is the highest, "DOCOMO Beijing communications laboratory co. Ltd." is set as the third constraint result.

The difference between the method of acquiring the third constraint result by the first model and the method of acquiring the first constraint result by the above first model is that the processing target is different, and the third model is the third. The processing target of the method of acquiring the constraint result of is the third candidate word set including the third alternative term, and the processing target of the method of acquiring the first constraint result by the first model is all translation pairs. Please understand that there is.

For example, the supervised first constraint model described above may be obtained by pre-training with labeled training data. The above supervised first constraint model may be a neural network model, a statistical method model, or the like, but is not limited thereto. It is easily understood that the supervised first constraint model according to the embodiments of the present disclosure can employ different neural network configurations, including, but not limited to, convolutional neural networks, recurrent neural networks (RNNs), and the like. The convolutional neural network includes, but is not limited to, a U-Net neural network, ResNet, DenseNet, and the like. The present disclosure is not limited to translating Chinese into English, and the present disclosure may include any desired translation, such as Chinese-Japanese translation, English-Japanese translation, German-Chinese translation. It should be understood that it can be applied when doing so.

As shown in FIG. 12, it is a flowchart of the method 600 for acquiring the third constraint result by the second constraint model according to the embodiment of the present disclosure. The constraint model may be a second constraint model without a teacher, based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. Obtaining the constraint result can include obtaining the first constraint result based on the representation of the semantic space of each candidate word by the second constraint model, and the second without the teacher. The constraint model does not need to be trained in advance.

For example, the above unsupervised second constraint model may be, but is not limited to, a text rank, an embedded center, and the like.

As an example, to obtain the first constraint result based on the representation of the semantic space of each candidate word by the second constraint model is the central representation of the semantic space of all candidate words and the semantic of each candidate word. It involves acquiring the first constraint result based on the distance to and from the representation of space.

As shown in FIG. 12, it is all candidates to obtain the first constraint result based on the distance between the central representation of the semantic space of all candidate words and the representation of the semantic space of each candidate word. Obtain the central representation of the semantic space of a word (S601), obtain the distance between the central representation of the semantic space of all candidate words and the representation of the semantic space of each candidate word (S602), and obtain the candidate with the shortest distance. It can include selecting a word as a third constraint result (S603).

The difference between the method of acquiring the third constraint result by the second model and the method of acquiring the first constraint result by the above second model is that the processing target is different, and the third model is the third. The processing target of the method of acquiring the constraint result of is the third candidate word set including the third alternative term, and the processing target of the method of acquiring the first constraint result by the second model is all translation pairs. Please understand that there is. Therefore, it will not be explained again here.

FIG. 13 is a flowchart of the method 700 for acquiring the third constraint result by the third constraint model according to the embodiment of the present disclosure. The third constraint model includes a first constraint model with supervised learning and a second constraint model with no supervised learning, and the first constraint model with supervised learning is obtained by learning from training data and is unsupervised. The second constraint model does not need to be trained in advance. The method includes the following steps S701-S702.

In step S701, the second constraint model acquires a fourth candidate word set containing M candidate words based on the semantic space representation of each candidate word, and M is an integer of 2 or more.

In step S702, the third constraint result is acquired based on the characteristics of each candidate word in the fourth candidate word set by the first constraint model.

For example, for step S701, the distance between the central representation of the semantic space of all candidate words and the representation of the semantic space of each candidate word is obtained based on the method described in FIG. 6 above, and then the distance is obtained. Sorted in ascending order of distance, the previous M candidate words may be selected as the fourth candidate word set. Next, for the fourth candidate word set, for example, a third constraint result is acquired by the first constraint model shown in FIG.

Steps S701 to S702 are similar to the method described with reference to FIG. 8, and include a method of acquiring a third constraint result by the third model and a method of acquiring the first constraint result by the above third model. The difference is that the processing target is different, and the processing target of the method of acquiring the third constraint result by the third model is the third candidate word set including the third alternative term, and the processing target is the third model. It should be understood that the processing target of the method of acquiring the first constraint result is all translation pairs, so that it will not be explained again here.

As described above, a method of acquiring the third constraint result according to the embodiment of the present disclosure will be described with reference to FIGS. 12 to 13, and then, returning to FIG. 11, in step S504, the third constraint result will be obtained. Correspondingly, by modifying the fourth information to be processed, the fifth information to be processed is generated.

For example, the fifth information to be processed is generated by modifying the information to be processed in the fourth by using the constrained decoding process according to the third constraint result. You may. The constrained decryption process can modify the fourth information to be processed by utilizing the third constraint result on the premise that the sentence sentence is kept fluent. The fourth method of modifying the information to be processed is not limited to that, and other known techniques can be used to modify the translation result, but it should be understood that it will not be explained again here.

Hereinafter, an example of the information processing method based on the target end according to the embodiment of the present disclosure will be described with reference to FIGS. 14 to 15.

As described above, as an example, the second processed text may be one text to be translated, and the third information to be processed is a sentence in the one text to be translated. One or more candidate word sets may be generated based on the translated text of one or more pieces of information to be processed before the third piece of information to be processed. Alternatively, as another example, one or more candidate word sets may be generated based on the translated text of the second processed text.

As shown in FIG. 14, it is a schematic diagram of an exemplary example of generating one or more candidate word sets based on the translated text of the second processed text, and FIG. 15 is the third information to be processed. It is a schematic diagram of an exemplary example of generating one or more candidate word sets based on the translated text of one or more pieces of information to be processed before.

As shown in FIG. 14, the second processed text may be one text to be translated, and the third information to be processed is a sentence in the one text to be translated. For the third information to be processed (current sentence 41), the translation 42 is first performed to generate the fourth information 43 to be processed, and then the co-reference query is used to generate the third alternative term. The candidate word set of 3 is obtained (that is, the related candidate word set 45), then the constraint model 46 obtains the third constraint result 47 in the third candidate word set, and finally, the third candidate word set. By modifying the fourth information 48 to be processed according to the constraint result 47 of 3, the fifth information 49 to be processed is generated.

In FIG. 14, one or more candidates by the above method (eg, co-reference word discovery 51) based on the initial translation 50 (ie, the fourth information to be processed) of the text to be translated. Generate a word set 52.

Alternatively, as shown in FIG. 15, the second processed text may be one text to be translated, and the third information to be processed is the text in the one text to be translated. .. One or more candidate word sets may be generated based on the translated text of one or more pieces of information to be processed before the third piece of information to be processed. In this case, one or more candidate word sets 56 are generated by the above method (eg, co-reference word discovery 55) based on one or more sentences 54 before the current sentence 53 being processed. ..

As described above, the information processing method based on the source end according to the embodiment of the present disclosure will be described with reference to FIGS. 1 to 10, and the information processing based on the target end according to the embodiment of the present disclosure will be described with reference to FIGS. 11 to 15. I explained the method. By the above method, the translation can be modified, the translations of all the alternative terms in the sentence are matched, and the accuracy and specialty of the translation are improved.

FIG. 16 is a schematic diagram of translation results according to the examples of the present disclosure. From FIG. 16, for the pronouns (underlined) contained in the information 60 to be processed, the translation results obtained using the method of the present disclosure can be obtained by other methods (eg, conventional translation method, dictionary replacement). It is more consistent than the translation results obtained in, can handle pronoun problems, and improves translation accuracy and expertise.

The information processing method according to the embodiment of the present invention has been described above with reference to the drawings. Hereinafter, the information processing apparatus according to the embodiment of the present disclosure will be described.

FIG. 17 illustrates a functional block diagram of the information processing apparatus according to the embodiment of the present disclosure. As shown in FIG. 17, the information processing apparatus 1300 according to the embodiment of the present disclosure includes a first information acquisition unit 1301 to be processed, a first candidate word set selection unit 1302, a first constraint result acquisition unit 1303, and a first constraint result acquisition unit 1303. It includes a second information generation unit 1304 to be processed. Each of the above modules performs each step of the information processing method according to the embodiment of the present disclosure described with reference to FIGS. 1 to 11 above, respectively. Those skilled in the art can realize these unit modules in various ways by hardware only, software only, or a combination thereof, and the present disclosure is not limited to any of them. For example, a central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), field programmable gate array (FPGA), or other form of data processing and / or instruction execution capability. It should be understood that these units may be implemented by processing units and corresponding computer instructions.

For example, the first information acquisition unit 1301 to be processed is used to acquire the first information to be processed.

For example, when the first information to be processed contains the first substitute word, the first candidate word set selection unit 1302 includes the first substitute word from one or more candidate word sets. One candidate word set is selected, and each of the one or more candidate word sets includes candidate words whose meanings match at least two and whose expressions do not match.

For example, the first constraint result acquisition unit 1303 is used to acquire the first constraint result translated corresponding to the first alternative term in the first candidate word set by the constraint model.

For example, the second information generation unit 1304 to be processed generates the second information to be processed by modifying the translation result of the first information to be processed according to the first constraint result. Used to do.

For example, the first constraint result acquisition unit 1303 pairs each candidate word in the first candidate word set with one or more translation results of each candidate word to form a translation pair, and features of each translation pair. And / or the first constraint result may be obtained based on the representation of the semantic space of each translation pair.

For example, the constraint model is a supervised first constraint model, and the first constraint result acquisition unit 1303 is based on the characteristics of each translation pair by the first constraint model. May be obtained, and the first supervised constraint model is obtained by learning from the training data.

For example, the constraint model is an unsupervised second constraint model, and the first constraint result acquisition unit 1303 is based on the representation of the semantic space of each translation pair by the second constraint model. The constraint result of the above may be acquired, and the second unsupervised constraint model does not need to be learned in advance.

For example, the first constraint result acquisition unit 1303 acquires the first constraint result based on the distance between the central representation of the semantic space of all translation pairs and the semantic space representation of each translation pair. May be good.

For example, the constraint model is a third constraint model including a first constraint model with supervised learning and a second constraint model without supervised learning, and the first constraint result acquisition unit 1303 is based on the second constraint model. , Obtaining a second candidate word set containing N translation pairs (N is an integer greater than or equal to 2) based on the semantic space representation of each translation pair, and by the first constraint model. The first supervised constraint model is obtained by training with training data, including acquiring the first constraint result based on the characteristics of each translation pair in the second candidate word set. The unsupervised second constraint model does not need to be trained in advance.

For example, the distance is an Euclidean distance.

For example, the characteristics of the translation pair include one or more of the frequency of selection of the translation, whether it was recently selected, the length of the translation, whether it is included in the formal word, and the semantic relevance. ..

For example, the first information to be processed is a part of the first processed text, and the one or more candidate word sets are generated based on the first processed text, or the first. Generated based on one or more information to be processed before the information to be processed in.

For example, the one or more candidates that learn a neural network and are based on the first processed text or based on one or more pieces of information to be processed prior to the first piece of information to be processed. Generate a word set.

For example, the second information to be processed is generated by modifying the translation result of the first information to be processed by using the constrained decoding process according to the first constraint result.

For example, before acquiring the first constraint result by the constraint model, words that do not exist in the user personalized information are deleted from the first candidate word set.

For example, the user personalized information is constructed based on one or more of the user's translation history, the user's style trends, and the field of translation.

FIG. 18 illustrates a functional block diagram of another information processing apparatus according to an embodiment of the present disclosure. As shown in FIG. 18, the information processing apparatus 1200 according to the embodiment of the present disclosure includes a fourth information generation unit 1201, a third candidate word set selection unit 1202, a third constraint result acquisition unit 1203, and a third constraint result acquisition unit 1203. It includes a fifth information generation unit 1204 to be processed. Each of the above modules performs each step of the information processing method according to the embodiment of the present disclosure described with reference to FIGS. 12 to 16, respectively. Those skilled in the art can realize these unit modules in various ways by hardware only, software only, or a combination thereof, and the present disclosure is not limited to any of them. For example, a central processing unit (CPU), graphics processing unit (GPU), tensor processing unit (TPU), field programmable gate array (FPGA), or other form of data processing and / or instruction execution capability. It should be understood that these units may be implemented by processing units and corresponding computer instructions.

For example, the fourth information generation unit 1201 to be processed is used to translate the third information to be processed and generate the fourth information to be processed.

For example, if the fourth information to be processed contains a third alternative term, the third candidate word set selection unit 1202 includes the third alternative term from one or more candidate word sets. 3 candidate word sets are selected, and each of the one or more candidate word sets includes candidate words whose meanings match at least two and whose expressions do not match.

For example, the third constraint result acquisition unit 1203 is used to acquire the third constraint result in the third candidate word set by the constraint model.

For example, the fifth information generation unit 1204 to be processed is to generate the fifth information to be processed by modifying the fourth information to be processed according to the third constraint result. Used.

Hereinafter, the information processing apparatus according to the embodiment of the present disclosure will be described with reference to FIG. FIG. 19 is a schematic diagram of the information processing apparatus 2000 according to the embodiment of the present disclosure. Since the information processing apparatus according to the present embodiment is the same as the details of the method described with reference to FIG. 1, here, for the sake of simplicity, detailed description of the same contents will be omitted.

Hereinafter, the information processing apparatus 1100 according to the embodiment of the present disclosure will be described with reference to FIG. FIG. 19 is a schematic diagram of an information processing apparatus according to an embodiment of the present disclosure. Since the function of the information processing apparatus of this embodiment is the same as the details of the method described with reference to FIG. 1, here, for the sake of simplicity, detailed description of the same contents will be omitted.

As shown in FIG. 19, the information processing apparatus 1100 includes a memory 1103 and a processor 1101. Although FIG. 19 shows that the information processing device 1100 includes only two devices, this is merely an example, and the information processing device 1100 may include one or more other devices. It should be noted that these devices are omitted here as they are not related to the idea of the invention.

The information processing apparatus 1100 based on the neural network according to the present disclosure includes a processor 1101 and a memory 1103, stores computer-readable instructions, and is an information processing method when the computer-readable instructions are executed by the processor 1001. The method is to acquire the first information to be processed, and when the first information to be processed includes the first alternative term, the method is described from one or more candidate word sets. Depending on the constraint model, a first candidate word set containing the first substitute term is selected, and each of the one or more candidate word sets contains words whose meanings match at least two and whose expressions do not match. , The first processing should be performed according to the acquisition of the first constraint result translated corresponding to the first alternative term in the first candidate word set and the first constraint result. It includes generating a second information to be processed by modifying the translation result of the information.

The technical effects of the information processing apparatus 1200, 1300 and the information processing apparatus 1100 in different embodiments may be referred to with reference to the technical effects of the information processing methods provided in the embodiments of the present disclosure, which are not described again here.

The information processing devices 1200 and 1300 and the information processing device 1100 can be used for various suitable electronic devices.

The present disclosure further includes a computer-readable storage medium, stores a computer-readable instruction, and causes the computer to execute an information processing method when the computer-readable instruction is executed by the computer. When the information to be obtained is acquired and the first information to be processed includes the first substitute word, the first candidate word containing the first substitute word from one or more candidate word sets. A set is selected, and each of the one or more candidate word sets contains words whose meanings match at least two and whose expressions do not match, and the constraint model allows the first candidate word set in the first candidate word set. By acquiring the first constraint result translated corresponding to the alternative term of, and modifying the translation result of the information to be processed according to the first constraint result, the second constraint result is obtained. Includes generating information to be processed in.

<Hardware configuration>
The block diagram used in the description of the above embodiment shows a block of functional units. These functional blocks (components) are realized by any combination of hardware and / or software. Further, the means for realizing each functional block is not particularly limited. That is, each functional block may be realized by one physically and / or logically coupled device, or directly and / or indirectly by two or more physically and / or logically separated devices. (For example, wired and / or wireless) may be connected and realized by these plurality of devices.

For example, the electronic device in one embodiment of the present invention may function as a computer that processes the attribute identification method of the present invention. FIG. 20 is a diagram showing an example of a hardware configuration of an electronic device according to an embodiment of the present invention. The electronic device 10 described above 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, and the like.

In the following description, the word "device" can be read as a circuit, a device, a unit, or the like. The hardware configuration of the electronic device 10 may be configured to include one or more of the devices shown in the figure, or may be configured not to include some of the devices.

For example, although only one processor 1001 is shown, it may be a plurality of processors. The process may be executed by one processor, simultaneously or sequentially by one or more processors, or by another method. The processor 1001 may be mounted on one or more chips.

For each function in the electronic device 10, the processor 1001 performs calculations by loading predetermined software (programs) on hardware such as the processor 1001 and the memory 1002, and communication by the communication device 1004, memory 1002, and storage 1003 are performed. It is realized by controlling the reading and / or writing of the data in.

Processor 1001 operates, for example, an operating system to control the entire computer. The processor 1001 may be configured by a central processing unit (CPU: Central Processing Unit) including an interface with peripheral devices, a control device, an arithmetic unit, a register, and the like.

Further, the processor 1001 reads a program (program code), a software module and data from the storage 1003 and / or the communication device 1004 into the memory 1002, and executes various processes according to these. As the program, a program that causes a computer to execute at least a part of the operations described in the above-described embodiment is used. For example, the control unit 401 of the electronic device 10 may be realized by a control program stored in the memory 1002 and operated by the processor 1001, and may be similarly realized for other functional blocks.

The memory 1002 is a computer-readable recording medium, and is, for example, a ROM (Read Only Memory), an EPROM (Erasable Program ROM), an EEPROM (Electrically Erasable Program ROM), a RAM (Random Access), or an appropriate recording medium. It may be composed of at least one of. The memory 1002 may be referred to as a register, a cache, a main memory (main storage device), or the like. The memory 1002 can store a program (program code), a software module, or the like that can be executed to implement the wireless communication method according to the embodiment of the present invention.

The storage 1003 is a computer-readable recording medium, for example, a flexible disk, a floppy (registered trademark) disk, a photomagnetic disk (for example, a read-only optical disk such as a CD-ROM (Compact Disc Rome), a digital versatile disk). , Blu-ray® discs), removable discs, hard disk drives, smart cards, flash memories (eg cards, sticks, key drives), magnetic strips, databases, servers, and at least one of other suitable recording media. It may be composed of. The storage 1003 may be referred to as an auxiliary storage device.

The communication device 1004 is hardware (transmission / reception device) for communicating between computers via a wired and / or wireless network, and is also referred to as, for example, a network device, a network controller, a network card, a communication module, or the like.

The input device 1005 is an input device (for example, a keyboard, a mouse, a microphone, a switch, a button, a sensor, etc.) that receives an input from the outside. The output device 1006 is an output device (for example, a display, a speaker, an LED lamp, etc.) that performs output to the outside. The input device 1005 and the output device 1006 may have an integrated configuration (for example, a touch panel).

Further, 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 composed of a single bus or may be composed of different buses between the devices.

In addition, the electronic device 10 includes a microprocessor, a digital signal processor (DSP: Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logical Processor), a PLD (Progrumable Digital Device), a hardware such as FPGA, and an FPGA (File). Often, the hardware may implement some or all of each functional block. For example, the processor 1001 may be implemented on at least one of these hardware.

Software, whether called software, firmware, middleware, microcode, hardware description language, or other names, is an instruction, instruction set, code, code segment, program code, program, subprogram, software module. , Applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, features, etc. should be broadly interpreted.

Further, software, instructions, information and the like may be transmitted and received via a transmission medium. For example, software from a website, server, or other remote source using wired technology (coaxial cable, fiber optic cable, twisted pair and digital subscriber line (DSL)) and / or wireless technology (infrared, microwave, etc.). When transmitted, these wired and / or wireless technologies are included within the definition of transmission medium.

Each aspect / embodiment described herein may be used alone, in combination, or switched during execution. Further, the order of the processing procedures, sequences, flowcharts, etc. of each aspect / embodiment described in the present specification may be changed as long as there is no contradiction. For example, the methods described herein present elements of various steps in an exemplary order and are not limited to the particular order presented.

As used herein, the phrase "according to" does not mean "according to" unless otherwise stated. In other words, the statement "according to" means both "according only" and "at least according".

Any reference to a unit using designations such as "first", "second" as used herein does not generally limit the quantity or order of those elements. These designations can be used herein as a convenient way to distinguish between two or more units. Therefore, references to the first and second units do not mean that only two units can be employed there, or that the first element must somehow precede the second unit.

When "included", "comprising", and variations thereof are used herein or in the claims, these terms are used in the same manner as the term "comprising". Intended to be inclusive. Furthermore, the term "or" as used herein or in the claims is intended to be non-exclusive.

Those skilled in the art will appreciate some patentable categories or situations in which the various aspects of this application include any new or useful process, combination with a machine, product, or substance, or any new or useful improvement to them. Can be understood to be exemplified and explained by. In contrast, various aspects of this application may be performed entirely by hardware or entirely by software (including firmware, resident software, microcode, etc.), a combination of hardware and software. May be executed by. The above hardware or software may be referred to as a "data block", "module", "engine", "unit", "element" or "system". Also, various aspects of this application may be expressed as computer products placed on one or more computer-readable media. The product contains computer readable program code.

This application uses specific terms to describe examples of this application. For example, "one example," "one example," and / or "several examples" means a feature, configuration, or characteristic according to at least one embodiment of the present application. Therefore, "one example," "one example," or "one alternative example" referred to more than once in different positions herein does not necessarily refer to the same embodiment. It should be emphasized and noted. Also, certain features, configurations, or properties in one or more embodiments of the present application can be combined appropriately.

Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. For example, terms defined in a dictionary should usually be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and unless explicitly defined so, an ideal or overly formal interpretation. It will be further understood that it will not be done.

Although the present invention has been described in detail above, it is clear to those skilled in the art that the present invention is not limited to the embodiments described in the present specification. The present invention can be implemented as modifications and modifications without departing from the spirit and scope of the present invention as determined by the description of the scope of claims. Therefore, the description of the present specification is for the purpose of exemplary explanation and does not have any limiting meaning to the present invention.

Claims (24)

Obtaining the first information to be processed and
When the first substitute word is included in the first information to be processed, the first candidate word set containing the first substitute word is selected from one or more candidate word sets, and the one or more candidate word sets are selected. Each of the plurality of candidate word sets contains at least two candidate words whose meanings match and whose expressions do not match.
To obtain the first constraint result translated corresponding to the first alternative term in the first candidate word set by the constraint model.
An information processing method including generating a second information to be processed by modifying a translation result of the first information to be processed according to the first constraint result. Using the constraint model, it is possible to obtain the first constraint result translated corresponding to the first alternative term in the first candidate word set.
Pairing each candidate word in the first candidate word set with one or more translation results of each candidate word to form a translation pair.
The information processing method according to claim 1, further comprising acquiring the first constraint result based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. The constraint model is a supervised first constraint model, and obtaining the first constraint result based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair is possible.
The first constraint model includes acquiring the first constraint result based on the characteristics of each translation pair, and the first constraint model with supervision is a claim obtained by learning from training data. The information processing method according to 2. The constraint model is an unsupervised second constraint model, and obtaining the first constraint result based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair is possible.
The second constraint model includes obtaining the first constraint result based on the representation of the semantic space of each translation pair, and the unsupervised second constraint model does not need to be trained in advance. The information processing method according to claim 2. With the second constraint model, it is possible to obtain the first constraint result based on the representation of the semantic space of each translation pair.
The information processing method according to claim 4, further comprising acquiring the first constraint result based on the distance between the central representation of the semantic space of all translation pairs and the representation of the semantic space of each translation pair. The constraint model is a third constraint model including a first constraint model with supervision and a second constraint model without supervising, and is based on the characteristics of each translation pair and / or the representation of the semantic space of each translation pair. To obtain the first constraint result,
By the second constraint model, a second candidate word set containing N translation pairs (N is an integer of 2 or more) is obtained based on the representation of the semantic space of each translation pair.
The first constraint model includes obtaining the first constraint result based on the characteristics of each translation pair in the second candidate word set.
The information processing method according to claim 2, wherein the first constraint model with supervised learning is obtained by learning from training data, and the second constraint model without supervised learning does not need to be learned in advance. The information processing method according to claim 5, wherein the distance is an Euclidean distance. The characteristics of the translation pair are claims including one or more of the frequency of selection, whether the translation was recently selected, the length of the translation, whether it is included in the formal word, and the semantic relevance. The information processing method according to 3 or 6. The first processing information is a part of the first processing text, and the one or more candidate word sets are generated based on the first processing text, or the first processing. The information processing method according to any one of claims 1 to 8, which is generated based on one or a plurality of information to be processed before the information to be processed. By learning a neural network, the one or more candidates may be based on the first processed text or based on one or more pieces of information to be processed prior to the first piece of information to be processed. The information processing method according to claim 9, which generates a word set. Claim 1 to generate the second information to be processed by modifying the translation result of the first information to be processed by using the constrained decoding process according to the first constraint result. The information processing method according to any one of 8 to 8. The information processing method according to any one of claims 1 to 8, wherein words that do not exist in the user personalized information are deleted from the first candidate word set before the first constraint result is acquired by the constraint model. The information processing method according to claim 12, wherein the user personalized information is constructed based on one or more of a user's translation history, a user's style tendency, and a field of translation. To translate the third information to be processed to generate the fourth information to be processed,
When the information to be processed includes the third substitute word, the third candidate word set containing the third substitute word is selected from one or more candidate word sets, and the one or more candidate word sets are selected. Each of the plurality of candidate word sets contains at least two candidate words whose meanings match and whose expressions do not match.
By the constraint model, the third constraint result in the third candidate word set is acquired.
An information processing method including generating a fifth information to be processed by modifying the fourth information to be processed according to the third constraint result. Obtaining the third constraint result in the third candidate word set by the constraint model is not possible.
The information processing method according to claim 14, further comprising acquiring the third constraint result based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. The constraint model is a supervised first constraint model, and the third constraint result is obtained based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. To get is
The first constraint model with the teacher includes learning data, including acquiring the third constraint result based on the characteristics of each candidate word in the third candidate word set by the first constraint model. The information processing method according to claim 15, which is obtained by learning from the above. The constraint model is an unsupervised second constraint model, and the third constraint result is obtained based on the characteristics of each candidate word in the third candidate word set and / or the representation of the semantic space of each candidate word. To get is
The second constraint model includes obtaining the third constraint result based on the representation of the semantic space of each candidate word, and the unsupervised second constraint model does not need to be learned in advance. Item 5. The information processing method according to Item 15. By the second constraint model, it is possible to obtain the third constraint result based on the representation of the semantic space of each candidate word.
The information processing method according to claim 17, wherein the third constraint result is acquired based on the distance between the central representation of the semantic space of all the candidate words and the semantic space representation of each candidate word. The constraint model is a third constraint model including a first constraint model with supervised learning and a second constraint model without supervised learning, and is based on the characteristics of each candidate word and / or the representation of the semantic space of each candidate word. To obtain the third constraint result,
Using the second constraint model, a fourth candidate word set containing M candidate words (M is an integer of 2 or more) is obtained based on the representation of the semantic space of each candidate word.
The first constraint model includes obtaining the third constraint result based on the characteristics of each candidate word in the fourth candidate word set.
The information processing method according to claim 15, wherein the first constraint model with supervised learning is obtained by learning from training data, and the second constraint model without supervised learning does not need to be learned in advance. The third information to be processed is a part of the second processed text, and the one or more candidate word sets are generated based on the translated text of the second processed text, or the first. The information processing method according to any one of claims 14 to 19, which is generated based on the translated text of one or a plurality of information to be processed before the information to be processed in 3. The first information acquisition unit to be processed for acquiring the first information to be processed, and
When the first substitute word is included in the first information to be processed, the purpose is to select the first candidate word set containing the first substitute word from one or more candidate word sets. , Each of the one or more candidate word sets has at least two meanings matching, and a first candidate word set selection unit containing candidate words whose expressions do not match.
A first constraint result acquisition unit for acquiring a first constraint result translated corresponding to the first alternative term in the first candidate word set by a constraint model, and a first constraint result acquisition unit.
A second information generation unit to be processed for generating a second information to be processed by modifying the translation result of the first information to be processed according to the first constraint result. Information processing equipment including. A fourth information generation unit to be processed for translating the third information to be processed and generating a fourth information to be processed,
When the third substitute word is included in the fourth information to be processed, it is for selecting a third candidate word set containing the third substitute word from one or more candidate word sets. , Each of the one or more candidate word sets includes a third candidate word set selection unit containing at least two matching meanings and mismatched expressions.
With the constraint model, a third constraint result acquisition unit for acquiring the third constraint result in the third candidate word set, and
Information including a fifth information generation unit to be processed for generating a fifth information to be processed by modifying the fourth information to be processed according to the third constraint result. Processing equipment. With the processor
Includes memory for storing computer-readable instructions,
When the computer-readable instruction is executed by the processor, the information processing method is executed.
The method obtains the first information to be processed and obtains the information to be processed.
When the first substitute word is included in the first information to be processed, the first candidate word set containing the first substitute word is selected from one or more candidate word sets, and the one or more candidate word sets are selected. Each of the candidate word sets contains words that have at least two matching meanings and mismatched expressions.
By the constraint model, the first constraint result translated corresponding to the first alternative term in the first candidate word set is acquired.
An information processing apparatus including generating a second information to be processed by modifying a translation result of the first information to be processed according to the first constraint result. A computer-readable storage medium for storing a computer-readable program, wherein the program is a computer-readable storage that causes a computer to execute the information processing method according to any one of claims 1 to 20. Medium.

Images