RU2779526C2 - Method and device for text translation at discourse level - Google Patents

Abstract

FIELD: information processing.

SUBSTANCE: invention relates to a method, a device and a machine-readable carrier for text translation at a discourse level. In the method, the text to be translated is obtained, while the text to be translated is a unit of text in the text at the discourse level to be translated, the relevant text is obtained for the text to be translated, while the relevant text includes at least one of the following: the preceding source text, the subsequent source text or the preceding target text, wherein the preceding source text is at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is at least one unit of text after the text to be translated in the text at the discourse level, and the preceding target text is the translation text of the mentioned preceding source text, and the text to be translated is translated in accordance with the relevant text, while the translation of the text to be translated in accordance with the relevant text includes: obtaining the initial encoding result by encoding the text to be translated, while the initial encoding result is semantic information of the text to be translated; obtaining the relevant encoding result by encoding the relevant text, while the relevant encoding result is semantic information of the relevant text; obtaining the target encoding result by processing the initial encoding result using all or part of encoding results in the relevant encoding result, and translating the text to be translated by decoding the target encoding result.

EFFECT: increase in the accuracy of text translation.

12 cl, 8 dwg

Description

[1] The present application claims priority according to the PRC Patent Application No. 201810463138.3 entitled "Method and Apparatus for Translating Text at the Discourse Level" and filed with the PRC Patent Office on May 15, 2018, and the content of said application is fully incorporated herein by links.

Technical field

[2] The present invention relates to the technical field of text translation and, in particular, to a method and apparatus for translating text at the discourse level.

Prerequisites for the creation of the invention

[3] The text at the level of discourse consists of a sequence of sentences. For example, the text at the discourse level can be the text of an oral report, a journal article, a literary work, etc. Since the main property of texts at the discourse level is the coherence and consistency of sentences, such a text is not just a set of consecutive sentences, but has a semantic unity with complete structures and clear functions.

[4] To translate such texts, methods of translating texts at the level of discourse should be used. Among the known methods of translating texts at the discourse level, the text at the discourse level to be translated is usually translated directly as a whole, considering it as a separate object of translation. However, the resulting translation is not perfect in the sense that the accuracy of the translation is not high, which negatively affects the user's experience.

The essence of the invention

[5] The purpose of the present invention is to provide a method and apparatus for translating text at the discourse level, which will improve the accuracy of the result of translating text at the discourse level.

[6] In accordance with embodiments of the present invention, a method for translating text at the discourse level is provided. The method includes:

[7] obtaining a text to be translated, wherein the text to be translated is a unit of text in the text at the discourse level to be translated;

[8] obtaining the relevant text for the text to be translated, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text,

[9] meanwhile, the preceding source text is understood to be at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is understood to be at least one unit of text after the text to be translated in the text at the discourse level , and by the previous target text is meant the text of the translation of the said previous source text, and

[10] translation of the text to be translated according to the relevant text.

[11] Optionally, the translation of the text to be translated according to the relevant text includes:

[12] translation of the text to be translated according to the relevant text using a pre-built encoding-decoding model.

[13] Optionally, the translation of the text to be translated according to the relevant text includes:

[14] obtaining the initial encoding result by encoding the text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated;

[15] obtaining a relevant encoding result by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text;

[16] obtaining a target encoding result by processing the initial encoding result using all or part of the encoding results in the relevant encoding result, wherein the target encoding result represents the semantic information of the text to be translated and carries all or part of the semantic information of the relevant text, and

[17] translating the text to be translated by decoding the target encoding result.

[18] Optionally, obtaining the initial encoding result by encoding the text to be translated includes:

[19] obtaining the initial encoding result by encoding the words of the text to be translated.

[20] Accordingly, obtaining a relevant encoding result by encoding the relevant text includes:

[21] encoding the words of each text unit in the relevant text to obtain encoding results corresponding to the corresponding text units, and determining said encoding results as the relevant encoding result.

[22] Optionally, after encoding the words of the text to be translated, the method also includes:

[23] obtaining a total result corresponding to the text to be translated by adding together the results of encoding the words of the text to be translated.

[24] Accordingly, after encoding the words of each unit of text in the relevant text, the method also includes:

[25] in the case where the relevant text includes the preceding source text, adding together the word encoding results of each text unit in the preceding source text to obtain a total result corresponding to each text unit;

[26] in the case where the relevant text includes subsequent source text, adding together the word encoding results of each unit of text in the subsequent source text to obtain a total result corresponding to each unit of text, and,

[27] in the case where the relevant text includes the preceding target text, adding together the word encoding results of each text unit in the preceding target text to obtain a total result corresponding to each text unit.

[28] Optionally, processing the initial encoding result using all or part of the encoding results in the relevant encoding result includes:

[29] determining, in the case where the relevant text includes the preceding source text, the encoding result of the preceding source text from the relevant encoding result;

[30] determining, in the case where the relevant text includes the subsequent source text, the encoding result of the subsequent source text from the relevant encoding result, and

[31] processing the initial encoding result according to the determined encoding result.

[32] Optionally, processing the initial encoding result according to the determined encoding result includes:

[33] determining, in the case where the relevant text includes the preceding source text, the degree of correlation between the text to be translated and the preceding source text according to the initial encoding result of the text to be translated and the encoding result of the preceding source text.

[34] determining, in the case where the relevant text includes a subsequent source text, the degree of correlation between the text to be translated and the subsequent source text according to the initial encoding result of the text to be translated and the encoding result of the subsequent source text, and

[35] processing the initial encoding result in accordance with the determined encoding result and the determined degrees of correlation.

[36] Optionally, translating the text to be translated by decoding the target encoding result includes:

[37] decoding, in the case where the relevant text includes the previous target text, the encoding target result using the encoding result of the previous target text in the relevant encoding result.

[38] Optionally, decoding the target encoding result using the encoding result of the preceding target text in the relevant encoding result includes:

[39] determining the degree of correlation between the text to be translated and the previous target text, and

[40] decoding the target encoding result in accordance with the determined degree of correlation and the encoding result of the previous target text.

[41] Optionally, determining the degree of correlation between the text to be translated and the previous target text includes:

[42] determining the degree of correlation between the text to be translated and the preceding source text as the degree of correlation between the text to be translated and the preceding target text.

[43] In accordance with embodiments of the present invention, a device for translating text at the discourse level is provided. The device contains:

[44] a text to be translated acquisition unit, which is configured to obtain a text to be translated, wherein the text to be translated is a unit of text in a discourse level text to be translated;

[45] a relevant text obtaining unit configured to obtain relevant text for the text to be translated, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text, wherein the previous source text is understood as at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is understood to be at least one unit of text after the text to be translated in the text at the discourse level, and the preceding target text is the text of the translation of the aforementioned preceding source text, and

[46] a translation unit of the text to be translated, which is configured to translate the text to be translated according to the relevant text.

[47] Optionally, the translation unit of the text to be translated is configured to translate the text to be translated according to the relevant text using a pre-built encoding-decoding model.

[48] Optionally, the translation block of the text to be translated contains:

[49] a first encoding subunit, configured to obtain an initial encoding result by encoding a text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated;

[50] a second encoding subunit, configured to obtain a relevant encoding result by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text;

[51] a result processing subunit configured to obtain a target encoding result by processing the initial encoding result using all or part of the encoding results in the relevant encoding result, wherein the target encoding result represents the semantic information of the text to be translated and carries all the semantic information of the relevant text or part of it, and

[52] a decoding execution sub-unit configured to translate text to be translated by decoding a target encoding result.

[53] Optionally, the decoding execution sub-unit is configured to decode, in the case where the relevant text includes the previous target text, the target encoding result using the encoding result of the previous target text in the relevant encoding result.

[54] In accordance with embodiments of the present invention, a device for translating text at the discourse level is provided. The device contains a processor, memory and a system bus,

[55] while the processor and memory are connected to each other via the system bus, and

[56] the memory is configured to store one or more programs containing instructions that, when executed by a processor, cause, by that processor, any of the discourse-level text translation methods described above.

[57] According to embodiments of the present invention, a computer-readable storage medium is provided. The computer-readable storage medium contains instructions which, when executed on a computer, cause that computer to perform any of the above-described methods of translating text at the discourse level.

[58] In accordance with embodiments of the present invention, there is provided a computer program product that, when executed on a terminal device, causes that terminal device to perform any of the discourse-level text translation methods described above.

[59] When using the method and apparatus for translating text at the discourse level according to the embodiments of the present invention, each unit of text in the text at the discourse level to be translated is taken as the text to be translated, then the relevant text is obtained from the text at the discourse level for the text to be translated, and translating the text to be translated based on said relevant text. It can be seen that when translating the text to be translated, not only the current context of the text to be translated is taken into account, but also the content of the relevant text for the text to be translated, which makes it possible to improve the accuracy of the result of the translation of the text to be translated and, accordingly, improve the user experience.

Brief description of the drawings

[60] The following will briefly describe the drawings used in describing embodiments of the present invention or known technologies to explain technical solutions in accordance with embodiments of the present invention or in accordance with the prior art. Obviously, the accompanying drawings in the following description illustrate only some of the embodiments of the present invention. Specialists in the art, without making creative efforts, can create other drawings based on them.

[61] FIG. 1 shows a flow diagram of a method for translating text at the discourse level in accordance with one embodiment of the present invention.

[62] FIG. 2 shows a text diagram at the discourse level in accordance with one embodiment of the present invention.

[63] FIG. 3 shows a topological block diagram of an encoding/decoding model in accordance with one embodiment of the present invention.

[64] FIG. 4 shows a flowchart for performing translation using a pre-built encoding/decoding model in accordance with one embodiment of the present invention.

[65] FIG. 5 is a flowchart showing the processing of initial encoding result processing according to one embodiment of the present invention.

[66] FIG. 6 is a network block diagram of a Gate Network in accordance with one embodiment of the present invention.

[67] FIG. 7 shows a block diagram of a device for translating text at the discourse level in accordance with one embodiment of the present invention.

[68] FIG. 8 shows a hardware block diagram of a device for translating text at the discourse level in accordance with one embodiment of the present invention.

Detailed description of embodiments of the invention

[69] In some methods of translating text at the discourse level, the entire text to be translated at the discourse level can be considered as the object of translation directly translated using the translation model. However, if each sentence of the text at the discourse level is translated using such methods, the contextual information for the translated sentence will not be taken into account, which will negatively affect the accuracy of the translation result and the user's perception. Also, when translating a sentence, the information that the translation model accepts may contain redundant information. More specifically, if the sentences preceding and following the translated sentence are not related to it in meaning, then in the case when the text at the discourse level is taken as the object of translation for this sentence, the irrelevant preceding and subsequent sentences will be redundant information.

[70] In order to solve the problems described above, a discourse-level text translation method according to embodiments of the present invention is provided. For each piece of text to be translated, where the text is, for example, a sentence, in a text at the discourse level, when translating the text to be translated, not only its current context is taken into account, but also the contextual information of the text to be translated, which improves accuracy. translation result of the text to be translated, and improve the user experience accordingly. In addition, in the embodiments of the present invention, when translating the text to be translated, the context information is dynamically selected according to the relevance of the text to be translated, the context information, which not only reduces redundant information in the information received in the translation model, but also improves the accuracy of the result. translation of the text to be translated. In addition, in the embodiments of the present invention, when translating the text to be translated, the result of the translation of the text preceding the text to be translated is also taken into account, which allows the result of the translation of the text to be translated to be associated with the result of the translation of the previous test. In this way, not only the accuracy of the translation of the text to be translated is improved, but also the smoothness of the translation text as a whole. In other words, the coherence and consistency of the sentences of the translated text at the discourse level are guaranteed.

[71] The technical solutions proposed in the embodiments of the present invention will be described clearly and completely in the following with reference to the attached drawings, to explain the objectives, technical solutions and advantages of the embodiments of the present invention. It is obvious that the considered embodiments of the present invention are only some, but not all possible options for its implementation. Other embodiments of the present invention, which can be obtained by those skilled in the art based on the embodiments of the present invention without creative effort, are within the spirit of the present invention.

[72] First Embodiment

[73] FIG. 1 shows a flow diagram of a method for translating text at the discourse level in accordance with one embodiment of the present invention. The method includes steps S101-S103 described below.

[74] In step S101, a text to be translated is obtained, wherein the text to be translated is a unit of text in a discourse-level text to be translated.

[75] The discourse-level text to be translated, that is, the discourse-level text prior to translation, will hereinafter be referred to simply as the discourse-level text for simplicity. This embodiment of the present invention imposes no restrictions on the type of text at the discourse level. For example, the text at the discourse level can be the text of an oral report, a journal article, a literary work, etc.

[76] The text at the level of discourse is a set of consecutive sentences. When translating text at the discourse level, the unit of translation can be a sentence or a sequential text of any other length in an embodiment of the present invention. If we assume that the translation is performed by sentences, then each sentence can be obtained in accordance with the sequence of sentences in the text at the discourse level. The sentence currently received is identified as the text to be translated and translated according to the steps below.

[77] It should be noted that this embodiment of the present invention does not impose restrictions on the language of the text at the discourse level, both before and after translation. For example, the original language may be Chinese and the target language may be English.

[78] In step S102, the relevant text for the text to be translated is obtained, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text.

[79] In this embodiment of the present invention, prior source text is understood to be at least one unit of text before the text to be translated in a discourse-level text, subsequent source text is understood to be at least one unit of text after the text to be translated in text at the discourse level, and the previous target text is understood as the text of the translation of the mentioned previous source text.

[80] More specifically, if the translation is performed by sentences, one or more sentences preceding the text to be translated can be derived from the text at the discourse level as the preceding source text. One or more sentences following the text to be translated can be derived from the text at the discourse level as a subsequent source text. In addition, since each sentence in the text is translated sequentially at the discourse level, at the time of translation of the text to be translated, the source text preceding it has already been translated. Thus, the translation text of the previous source text, that is, the previous target text, can be directly obtained.

[81] It should be noted that the number of sentences in the preceding source text and subsequent source text may be the same or different. If k sentences preceding the text to be translated are defined as the preceding source text, k or l sentences following the text to be translated can be defined as the subsequent source text, where k and l are integers, where k≠l, k ≥1, l≥1.

[82] For example, in the text at the discourse level shown in FIG. 2, the text to be translated is indicated by X, the sentence X _-1 preceding X may be identified as the preceding source text, or several sentences preceding X, such as sentence X _-1 and sentence X _-2 , may be designated as previous source text. Similarly, the sentence X ₁ following X may be designated as a subsequent source text, or multiple sentences following X, such as sentence X ₁ , sentence X ₂ and sentence X ₃ , may be designated as a subsequent source text.

[83] In step 103, translation of the text to be translated is performed according to the relevant text.

[84] In this embodiment of the present invention, when translating the text to be translated, the relevant text may also be taken into account. In other words, for the translation of the text to be translated, the semantic information of at least one of the following is taken into account: the previous source text, the subsequent source text, or the previous target text for the text to be translated, to obtain the translation result of the text to be translated. The degree of correlation between the text to be translated and the preceding source text, the degree of correlation between the text to be translated and the subsequent source text, and the degree of correlation between the text to be translated and the preceding target text may also be considered. Obviously, by taking into account the semantic content of the text to be translated and the relevant text, as well as the degree of correlation between the text to be translated and the relevant text, a more accurate result of the translation of the text to be translated can be obtained.

[85] It should be noted that when the relevant text includes the previous target text (translation result of the previous source text), that is, when the previous target text is used as an auxiliary in the translation of the text to be translated, not only the accuracy of the translation result of the text to be translated is improved, but and it is possible to link the result of the translation of the text to be translated with the result of the translation of the previous source text. In other words, the result of the translation of the text to be translated can be more related in meaning to the result of the translation of the previous source text, which guarantees the coherence and consistency of the sentences of the translated text at the discourse level.

[86] Note that for a description of the specific implementation of step S103, refer to the second embodiment of the present invention.

[87] Thus, according to this embodiment of the present invention, a discourse-level text translation method is provided in which each unit of text in the discourse-level text to be translated is taken as the text to be translated, then from the discourse-level text obtaining a relevant text for the text to be translated, and translating the text to be translated based on said relevant text. It can be seen that when translating the text to be translated, not only the current context of the text to be translated is taken into account, but also the content of the relevant text for the text to be translated, which makes it possible to improve the accuracy of the result of the translation of the text to be translated and, accordingly, improve the user experience.

[88] Second Embodiment

[89] This embodiment of the present invention includes a specific implementation of step S103 mentioned in the first embodiment of the present invention.

[90] In this embodiment of the present invention, the encoding-decoding model can be pre-built as a model for translating texts at the discourse level, that is, the encoding-decoding model performs text translation at the discourse level by encoding and then decoding. Accordingly, the above-mentioned step S103 may include specifically translating the text to be translated according to the relevant text using the pre-built encoding/decoding model.

[91] It should be noted that the application of the encoding-decoding model for translating text to be translated is discussed below with reference to FIG. 3, which shows a topological block diagram of the encoding-decoding model.

[92] To provide an introduction for Figure 3, consider an example in which the translation is performed by sentences, while taking k sentences before the text to be translated as the previous source text, and taking k sentences after the text to be translated as the subsequent source text. translation. Let's first define the following.

[93] 1. The text at the discourse level before translation is represented as {X _-k , …, X _-1 , X, X ₁ , … X _k }, where X represents the text to be translated, X _-1 ... X _-k in sequence represent the first sentence before X, ..., the kth sentence before X; X ₁ ... X _k sequentially represent the first sentence after X, ..., the kth sentence after X. Here, X _-1 ... X _-k are jointly defined as the preceding source text, and X ₁ ... X _k are jointly defined as the subsequent source text .

где

последовательно представляют первое слово, второе слово, …, n-е слово предложения Х_-k; остальные предложения предшествующего исходного текста представлены аналогичным образом. Все слова предложения Х_k в последующем исходном тексте представлены как

где

последовательно представляют первое слово, второе слово, …, n-е слово предложения Х_k; остальные предложения последующего исходного текста представлены аналогичным образом.[94] For simplicity, let's assume that each sentence contains n words (in practice, the actual number of words will be used). As shown in FIG. 3, all words of the text X to be translated are represented as {x ₁ , x ₂ , ..., x _n }, where x ₁ , x ₂ , ..., x _n sequentially represent the first word, the second word, ..., the nth word text X to be translated. Similarly, all the words of the sentence X _-k in the preceding source text are represented as

where

sequentially represent the first word, the second word, ..., the n-th word of the sentence X _-k ; the remaining sentences of the preceding source text are presented in a similar way. All words of the sentence Х _k in the following source text are presented as

where

sequentially represent the first word, the second word, ..., the n-th word of the sentence X _k ; the remaining sentences of the subsequent source text are presented in a similar way.

[95] 2. Text at the discourse level after translation is represented as {Y _-k , …, Y _-1 , Y, Y ₁ , … Y _k }, where Y represents the translation text of text X to be translated, Y _-1 …Y _-k sequentially represent the translation text X _-1 ...X _-k in the preceding source text, and Y ₁ ...Y _k sequentially represent the translation text X ₁ ...X _k in the subsequent source text. Here Y _-1 ...Y _-k are jointly defined as the previous target text.

, где

последовательно представляют первое слово, второе слово, …, n-е слово предложения Y_-k. Все слова предложения Y_-1 в предшествующем целевом тексте представлены как

, где

последовательно представляют первое слово, второе слово, …, n-е слово предложения Y_-1. Остальные предложения последующего целевого текста представлены аналогичным образом.[96] All words of the translation text Y for the text X to be translated are represented as {y ₁ , y ₂ , …, y _n }, where y ₁ , y ₂ , …, y _n successively represent the first word, the second word, … , the nth word of the text Y as shown in Fig. 3, all the words of the sentence Y _-k in the preceding target text are represented as

, where

sequentially represent the first word, the second word, ..., the n-th word of the sentence Y _-k . All words of sentence Y _-1 in the preceding target text are represented as

, where

sequentially represent the first word, the second word, ..., the n-th word of the sentence Y _-1 . The remaining sentences of the subsequent target text are presented in a similar way.

[97] Below, with reference to FIG. 3, the application of an encoding-decoding model for translating a text to be translated will be described.

[98] FIG. 4 is a flowchart for performing translation using a pre-built encoding/decoding model, which in one implementation of this embodiment includes steps S401-S404 described below in particular.

[99] In step S401, an initial encoding result is obtained by encoding the text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated.

[100] In one implementation, step S401 may include, in particular, obtaining the initial encoding result by encoding the text to be translated using a bidirectional recurrent neural network (Recurrent Neural Network, RNN).

[101] In step S401 or in specific implementations of step S401, when encoding the text to be translated, the initial encoding result can be obtained, in particular, by encoding all the words of the text to be translated.

[102] More specifically, as shown in FIG. 3, the words of the text X to be translated, that is, {x ₁ ,x ₂ ,...,x _n }, are entered into the bidirectional RNN as a vector of words.

To vectorize words in this embodiment of the present invention, any of the methods known or proposed in the future can be used.

[103] The formula for calculating each of the hidden states of a bidirectional RNN network is as follows:

[104]

представляет вектор слов, составленный из слов х_i, h_i-1 представляет (i-1)-e скрытое состояние, используемое для представления семантической информации слова х, а h_i. представляет i-e скрытое состояние, используемое для представления семантической информации слова[105] where i=1, 2, …, n, f is a non-linear function, for example, a sigmoid function or a network with long and short-term memory (Long Short-Term Memory, LSTM), W and U are the training parameters of a bidirectional RNN network ,

represents a word vector composed of words x _i , h _i-1 represents the (i-1)-e hidden state used to represent the semantic information of word x, and h _i . represents an ie hidden state used to represent the semantic information of a word

после чтения по порядку вектора слов, составленного из слов текста X, подлежащего переводу, где

представляет информацию текста X, подлежащего переводу, в прямом порядке; обратная RNN-сеть формирует обратную скрытую последовательность

после чтения в обратном порядке вектора слов, составленного из слов текста X, подлежащего переводу, где

представляет информацию текста, подлежащего переводу, в обратном порядке.[106] A bidirectional RNN network includes a forward RNN network and a reverse RNN network. Forward RNN Network Forms Direct Hidden Sequence

after reading in order a vector of words composed of the words of the text X to be translated, where

presents the information of the text X to be translated in direct order; reverse RNN-network forms reverse hidden sequence

after reading in reverse order a vector of words composed of the words of the text X to be translated, where

represents the information of the text to be translated in reverse order.

и обратной скрытой последовательности

соответствуют друг другу, соединяют вместе и получают последовательность состояний, соответствующую слову в этой позиции. Другими словами, последовательности состояний, соответствующие словам в {х₁,х₂,…,х_n}, имеют следующий вид: последовательностью состояний, соответствующей х₁, является

последовательностью состояний, соответствующей х₂, является

а последовательностью состояний, соответствующей х_n, является

Соответственно, семантическая информация текста X, подлежащего переводу, может быть представлена как Н={h₁, h₂, …, h_n}, где Н - начальный результат кодирования, полученный кодированием текста, подлежащего переводу.[107] Blocks of hidden states whose positions are in direct hidden sequence

and reverse hidden sequence

match each other, connect together and get the sequence of states corresponding to the word in this position. In other words, the state sequences corresponding to the words in {x ₁ ,x ₂ ,…,x _n } are as follows: the state sequence corresponding to x ₁ is

the sequence of states corresponding to x ₂ is

and the sequence of states corresponding to x _n is

Accordingly, the semantic information of the text X to be translated can be represented as H={h ₁ , h ₂ , …, h _n }, where H is the initial encoding result obtained by encoding the text to be translated.

, полученный путем сложения вместе последовательностей состояний в Н, может быть определен в качестве нового начального результата кодирования, для представления семантической информации текста X, подлежащего переводу,[108] In addition, in one implementation, after encoding the words of the text to be translated, the method may also include obtaining a total result corresponding to the text to be translated by adding together the results of encoding the words of the text to be translated. In such an implementation, after obtaining the result H of coding, the total result

, obtained by adding together the sequences of states in H, can be defined as a new initial encoding result, to represent the semantic information of the text X to be translated,

[109] where

[110] In step S402, a relevant encoding result is obtained by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text.

[111] In one implementation, step S402 may include, in particular, obtaining a relevant encoding result by encoding the relevant text using a bidirectional RNN.

[112] In step S402 or specific implementations of step S402, when encoding relevant text, in particular, the words of each text unit in the relevant text may be encoded to obtain encoding results corresponding to the corresponding text units, said encoding results being determined as the relevant encoding result.

…, семантическая информация предложения Х_-k представлена как

а H_-1, …, Н_-k представляет собой результат кодирования, полученный кодированием предшествующего исходного текста, где в

…,

а в

…,

[113] More specifically, as shown in FIG. 3, when the relevant text includes the preceding source text, that is, the relevant text contains K sentences X _-1 ... X _-k before the text X to be translated, X _-1 ... X _-k can be encoded similarly to the encoding of the text X to be translated in step S401, where it is only necessary to replace the text X to be translated with X _-1 ... X _-k separately, so that the semantic information of the sentence X _-1 is represented as

…, the semantic information of the sentence X _-k is presented as

and H _-1 , ..., H _-k is the encoding result obtained by encoding the previous source text, where in

…,

and in

…,

семантическая информация предложения Х_k представлена как

а Н₁, …, Н_k представляет собой результат кодирования, полученный кодированием последующего исходного текста, где в

…,

…,

[114] When the relevant text includes subsequent source text, that is, the relevant text contains K sentences X ₁ ...X _k after the text X to be translated, X ₁ ...X _k may be encoded similarly to the encoding of the text X to be translated in step S401, where it is only necessary to replace the text X to be translated with X ₁ ... X _k separately, so that the semantic information of the sentence X ₁ is represented as

the semantic information of the sentence Х _k is presented as

and H ₁ , ..., H _k is the encoding result obtained by encoding the subsequent source text, where in

…,

…,

семантическая информация предложения Y_-k представлена как

а

представляет собой результат кодирования, полученный кодированием предшествующего целевого текста, где в

…,

а в

…,

[115] When the relevant text includes the preceding target text, i.e. the relevant text contains the translation text Y ₁ ...Y _-k for K sentences X _-1 ...X _-k before the text X to be translated, Y ₁ ...Y _-k may be encoded similarly encoding the text X to be translated in step S401, where it is only necessary to replace the text X to be translated with Y _-1 ... Y _-k separately, so that the semantic information of the sentence Y _-1 is represented as

the semantic information of the sentence Y _-k is represented as

a

is the encoding result obtained by encoding the preceding target text, where in

…,

and in

…,

[116] In addition, in one implementation, after encoding the words of each unit of text in the relevant text, the method may also include steps A1-C1.

[117] In step A1, in the case where the relevant text includes the previous source text, the word encoding results of each text unit in the previous source text are added together to obtain a total result corresponding to each text unit.

для всех последовательностей состояний в Н_-1 может использоваться для представления семантической информации предложения Х_-1, суммарный результат

для всех последовательностей состояний в Н_-k может использоваться для представления семантической информации предложения Х_-k, а

…,

совместно определяют в качестве результата кодирования предшествующего исходного текста,[118] After obtaining the results of H _-1 , ..., H _-k encoding for the previous source text, the total result

for all sequences of states in H _-1 can be used to represent the semantic information of the sentence X _-1 , the total result

for all state sequences in H _-k can be used to represent the semantic information of the sentence X _-k , and

…,

jointly defined as the result of encoding the previous source text,

[119] where

[120] ...;

[121]

[122] In step B1, in the case where the relevant text includes subsequent source text, the word encoding results of each text unit in the subsequent source text are added together to obtain a total result corresponding to each text unit.

для всех последовательностей состояний в Н₁ может использоваться для представления семантической информации предложения Х₁, суммарный результат

для всех последовательностей состояний в Н_k может использоваться для представления семантической информации предложения Х_k, а

, …,

совместно определяют в качестве результата кодирования последующего исходного текста,[123] After obtaining the results H ₁ , ..., H _k encoding the subsequent source text, the total result

for all sequences of states in H ₁ can be used to represent the semantic information of the sentence X ₁ , the total result

for all sequences of states in H _k can be used to represent the semantic information of the sentence X _k , and

, …,

jointly determined as the result of encoding the subsequent source text,

[124] where

[125] ...;

[126]

[127] In step C1, in the case where the relevant text includes the previous target text, the word encoding results of each text unit in the previous target text are added together to obtain a total result corresponding to each text unit.

кодирования предшествующего исходного текста, суммарный результат

для всех последовательностей состояний в

может использоваться для представления семантической информации предложения Y_-1, суммарный результат

для всех последовательностей состояний в

может использоваться для представления семантической информации предложения

а

,…,

совместно определяют в качестве результата кодирования предшествующего целевого текста,[128] After receiving the results

encoding of the previous source text, the total result

for all state sequences in

can be used to represent the semantic information of the sentence Y _-1 , the total result

for all state sequences in

can be used to represent the semantic information of a sentence

a

,…,

jointly determined as the encoding result of the preceding target text,

[129] where

[130] ...;

[131]

[132] It should be noted that in this embodiment of the present invention, the order of performing the steps S401 and S402 described above is not limited. Steps S401 and S402 may be performed simultaneously, step S401 may be performed before the execution of step S402, or step S402 may be performed before the execution of step S401.

[133] In step S403, the target encoding result is obtained by processing the initial encoding result using all or part of the encoding results in the relevant encoding result, wherein the target encoding result represents the semantic information of the text to be translated, and carries all or part of the semantic information of the relevant text .

кодирования предшествующего целевого текста, могут использоваться для обработки начального результата Н кодирования текста, подлежащего переводу, при этом обработанный результат кодирования определяют в качестве целевого результата кодирования.[134] In this embodiment of the present invention, one or more of the following encoding results: the encoding result H _-1 , ..., H _-k of the preceding source text encoding, the encoding result H ₁ , ..., H _k of the subsequent source text, and the result

encoding of the previous target text may be used to process the initial encoding result H of the text to be translated, wherein the processed encoding result is determined as the target encoding result.

кодирования предшествующего исходного текста, результат

кодирования последующего исходного текста и результат

кодирования предшествующего целевого текста, могут использоваться для обработки начального результата Н кодирования текста, подлежащего переводу, причем обработанный результат кодирования определяют в качестве целевого результата Н_с кодирования.[135] Alternatively, one or more of the following encoding results: result

encoding of the previous source text, the result

encoding the subsequent source text and the result

encoding of the preceding target text may be used to process the initial encoding result H _of the text to be translated, wherein the processed encoding result is determined as the target encoding result Hc.

[136] It should be noted that, when processing the initial encoding result, the semantic information of text units in the relevant text may be selected by different degree values corresponding to the degree of correlation between the text to be translated and the relevant text, so that the target encoding result may carry information with different degrees. and present semantic information H _{from the} text to be translated.

[137] Note that for a description of the specific implementation of step S403, refer to the third embodiment of the present invention.

[138] In step S404, the text to be translated is translated by decoding the target encoding result.

[139] When decoding the target result H _with encoding, known decoding methods, for example, using the attention model (see Fig. 3) or future decoding methods proposed in the future, can be used to obtain the translation result of the text to be translated.

[140] Note also that for a description of the specific implementation of step S404, you can refer to the fourth embodiment of the present invention.

[141] Thus, in this embodiment of the present invention, an encoding-decoding model is used to encode and decode the text to be translated as well as the relevant text. When the text to be translated is translated using the encoding and decoding method, not only the current context of the text to be translated, but also the relevant text for the text to be translated is taken into account. Thus, the accuracy of the translation result of the text to be translated can be improved and the user experience improved accordingly.

[142] Third Embodiment

[143] This embodiment of the present invention includes a specific implementation of step S403 mentioned in the second embodiment of the present invention.

[144] FIG. 5 is a flowchart showing the processing of the initial encoding result. "Processing the initial encoding result using all or part of the encoding results in the relevant encoding result" in the above-mentioned step S403 also includes steps S501-S503 described below.

[145] In step S501, in the case where the relevant text includes the previous source text, the encoding result of the previous source text is determined from the relevant encoding result.

[146] In this embodiment of the present invention, the result

в соответствии со вторым вариантом осуществления изобретения.encoding of the preceding source text can be H _-1 , ..., H _-k or

according to the second embodiment of the invention.

[147] In step S502, in the case where the relevant text includes the subsequent source text, the encoding result of the subsequent source text is determined from the relevant encoding result.

результатом кодирования последующего исходного текста может являться

в соответствии со вторым вариантом осуществления изобретения.[148] In this embodiment of the present invention, if the encoding result of the preceding source text is H _-1 , ..., H _-k , the encoding result of the subsequent source text may be H ₁ , ..., H _k in accordance with the second embodiment of the invention; similarly, if the result of encoding the preceding source text is

the result of encoding the subsequent source text may be

according to the second embodiment of the invention.

[149] In step S503, processing of the initial encoding result is performed in accordance with the determined encoding result.

[150] In one implementation, step S503 may include, in particular, steps A2-C2, described below.

[151] In step A2, in the case where the relevant text includes the previous source text, the degree of correlation between the text to be translated and the previous source text is determined according to the initial encoding result of the text to be translated and the encoding result of the previous source text.

[152] The degree of relevance between the text to be translated and each sentence in the preceding source text (hereinafter referred to as the preceding sentence) is uncertain, and if the text to be translated is not at all related to the preceding sentence, the preceding sentence will be redundant information. If the text to be translated is translated with reference to the preceding sentence, this will adversely affect the accuracy of the translation result of the text to be translated, that is, the accuracy of the translation result will be low. Therefore, in order to make more rational use of the encoding result of the previous source text, in this embodiment of the present invention, the degree of correlation between the text to be translated and each of the previous sentences is determined separately, so that the semantic information corresponding to the previous sentence can be applied dynamically using the determination result degree of correlation in order to avoid redundancy of information caused by the preceding sentence.

а результатами кодирования предшествующих предложений предшествующего исходного текста могут быть, соответственно,

Может использоваться шлюзовая сеть (Gate Network) со структурой нейронной сети для формирования коэффициента корреляции, соответствующего каждому предшествующему предложению, причем коэффициент корреляции представляет степень корреляции между текстом, подлежащим переводу, и соответствующим предшествующим предложением.[153] At this step, the initial encoding result of the text to be translated may be

and the encoding results of previous sentences of the previous source text can be, respectively,

A Gate Network with a neural network structure may be used to generate a correlation coefficient corresponding to each preceding sentence, the correlation coefficient representing the degree of correlation between the text to be translated and the corresponding preceding sentence.

кодирования текста, подлежащего переводу, и результаты

кодирования, отвечающие соответствующим предшествующим предложениям, используют в качестве входных данных для шлюзовой сети. Затем шлюзовая сеть вычисляет коэффициент корреляции, соответствующий каждому из предшествующих предложений, согласно следующей формуле: [154] FIG. 6 shows a network block diagram of a gateway network. In specific implementations, the initial result

encoding of the text to be translated and the results

encodings corresponding to the respective previous proposals are used as input to the gateway network. The gateway network then calculates the correlation coefficient corresponding to each of the preceding sentences according to the following formula:

[155]

- начальный результат кодирования текста, подлежащего переводу;

- результат кодирования i-го предложения, предшествующего тексту, подлежащему переводу; λ_-i - коэффициент корреляции i-го предложения, предшествующего тексту, подлежащему переводу, который представляет степень корреляции между текстом, подлежащим переводу, и i-м предложением, предшествующим тексту, подлежащему переводу; U и V - параметры обучения шлюзовой сети, которые получают путем обучения на большом объеме заранее собранных данных, при этом каждый из них имеет размер d×1.[156] where i=1, 2 … k;

- the initial result of encoding the text to be translated;

- the result of encoding the i-th sentence preceding the text to be translated; λ _-i - correlation coefficient of the i-th sentence preceding the text to be translated, which represents the degree of correlation between the text to be translated and the i-th sentence preceding the text to be translated; U and V are the training parameters of the gateway network, which are obtained by training on a large amount of pre-collected data, each of which has a size of d×1.

и

представляют собой матрицы размера 1×d, где d - количество скрытых узлов в процедуре кодирования посредством двунаправленной RNN-сети, которое заранее сконфигурировано. Таким образом, скалярная величина размерностью 1 может быть получена умножением матриц, указанных в формуле (2), а затем преобразована с помощью нелинейной функции, то есть сигмоидной функции, с получением скалярного значения λ_-i, лежащего в диапазоне от 0 до 1. Чем больше значение λ_-1, тем выше степень корреляции между текстом, подлежащим переводу, и i-м предложением, предшествующим тексту, подлежащему переводу, и наоборот.[157] It should be noted that

and

are 1×d matrices, where d is the number of hidden nodes in the bidirectional RNN coding procedure, which is preconfigured. Thus, a scalar value of dimension 1 can be obtained by multiplying the matrices indicated in formula (2), and then transformed using a non-linear function, that is, a sigmoid function, to obtain a scalar value λ _-i ranging from 0 to 1. Than the greater the value of λ _-1 , the higher the degree of correlation between the text to be translated and the i-th sentence preceding the text to be translated, and vice versa.

[158] As can be seen, the correlation coefficients {X _-k ,X _-k-1 ,...,X _-1 ) corresponding to the corresponding preceding sentences in the preceding source text can be calculated using formula (2).

[159] In step B2, in the case where the relevant text includes the subsequent source text, the degree of correlation between the text to be translated and the subsequent source text is determined according to the initial encoding result of the text to be translated and the encoding result of the subsequent source text.

[160] The relevance between the text to be translated and each sentence in the subsequent source text (hereinafter referred to as the subsequent sentence) is uncertain, and if the text to be translated is not at all related to the subsequent sentence, the subsequent sentence will be redundant with respect to the text to be translated. translation. If the text to be translated is translated with reference to a subsequent sentence, this will negatively affect the accuracy of the translation result, that is, the accuracy of the translation result will be low. Therefore, in order to make more rational use of the encoding result of the subsequent source text, in this embodiment of the present invention, the degree of correlation between the text to be translated and each subsequent sentence individually is determined, so that the semantic information corresponding to the subsequent sentence can be applied dynamically using the result of determining the degree correlations to avoid information redundancy caused by the subsequent sentence.

, а результатами кодирования предшествующего предложения предшествующего исходного текста могут быть, соответственно,

Может использоваться шлюзовая сеть (Gate Network) со структурой нейронной сети для формирования коэффициента корреляции, соответствующего каждому предшествующему предложению, причем коэффициент корреляции представляет степень корреляции между текстом, подлежащим переводу, и соответствующим предшествующим предложением.[161] At this step, the initial encoding result of the text to be translated may be

, and the results of encoding the previous sentence of the previous source text can be, respectively,

A Gate Network with a neural network structure may be used to generate a correlation coefficient corresponding to each preceding sentence, the correlation coefficient representing the degree of correlation between the text to be translated and the corresponding preceding sentence.

кодирования текста, подлежащего переводу, и результаты

кодирования, отвечающие соответствующим последующим предложениям, используют в качестве входных данных для шлюзовой сети. Затем шлюзовая сеть вычисляет коэффициент корреляции, соответствующий каждому из последующих предложений, согласно следующей формуле:[162] FIG. 6 shows a network block diagram of a gateway network. In specific implementations, the initial result

encoding of the text to be translated and the results

the encodings corresponding to the respective subsequent sentences are used as input to the gateway network. The gateway network then calculates the correlation coefficient corresponding to each of the following sentences according to the following formula:

[163]

- начальный результат кодирования текста, подлежащего переводу;

- результат кодирования i-го предложения, следующего за текстом, подлежащим переводу; λ_i - коэффициент корреляции i-го предложения, следующего за текстом, подлежащим переводу, который представляет степень корреляции между текстом, подлежащим переводу и i-м предложением, следующим за текстом, подлежащим переводу; U и V - параметры обучения шлюзовой сети, которые получены путем обучения на большом объеме заранее собранных данных, при этом каждый из них имеет размер d×1.[164] where i=1,2…k;

- the initial result of encoding the text to be translated;

- the result of encoding the i-th sentence following the text to be translated; λ _i - correlation coefficient of the i-th sentence following the text to be translated, which represents the degree of correlation between the text to be translated and the i-th sentence following the text to be translated; U and V are the training parameters of the gateway network, which are obtained by training on a large amount of pre-collected data, each of which has a size of d×1.

и

представляют собой матрицы размера 1×d, где d - количество скрытых узлов в процедуре кодирования посредством двунаправленной RNN-сети, которое сконфигурировано заранее. Таким образом, скалярная величина размерностью 1 может быть получена умножением матриц, указанных в формуле (3), а затем преобразована с помощью нелинейной функции, то есть сигмоидной функции, для получения скалярного значения λ_i, лежащего в диапазоне от 0 до 1. Чем больше значение λ_i, тем выше степень корреляции между текстом, подлежащим переводу, и i-м предложением, следующим за текстом, подлежащим переводу, и наоборот.[165] It should be noted that

and

are 1×d matrices, where d is the number of hidden nodes in the bidirectional RNN coding procedure, which is preconfigured. Thus, a scalar value of dimension 1 can be obtained by multiplying the matrices indicated in formula (3), and then transformed using a non-linear function, that is, a sigmoid function, to obtain a scalar value λ _i ranging from 0 to 1. The more value λ _i , the higher the degree of correlation between the text to be translated and the i-th sentence following the text to be translated, and vice versa.

[166] As can be seen, the correlation coefficients {X ₁ ,X ₂ ,...,X _k } for the respective subsequent sentences in the following source text can be calculated using formula (3).

[167] In step C2, an initial encoding result is processed according to the determined encoding result and the determined degree of correlation.

[168] After calculating the coefficients {X _-k ,X _-k-1 ,...,X _-1 } of the correlation corresponding to the respective previous sentences of the previous source text, the previous sentences can be dynamically selected in accordance with their correlation coefficients; likewise, after computing the correlation coefficients {X ₁ ,X ₂ ,...,X _k } corresponding to the respective subsequent sentences of the subsequent source text, subsequent sentences may be dynamically selected based on these correlation coefficients. This procedure is described in more detail below.

[169] Based on the initial encoding result H={h ₁ ,h ₂ ,...,h _n } of the text to be translated, which is obtained in step S401 in the second embodiment of the present invention, for the corresponding h _i (i=1.2... n) for each word in the text to be translated, contextual information can be entered according to the following formula:

[170]

- результаты кодирования соответствующих предшествующих предложений, и

- результаты кодирования соответствующих последующих предложений.[171] where {X _-k ,X _-k-1 ,…,X _-1 } are the correlation coefficients corresponding to the respective previous proposals and obtained using the gateway network, {X ₁ ,X ₂ ,…,X _k } are the coefficients correlations corresponding to the corresponding subsequent sentences and obtained using the gateway network,

- the encoding results of the corresponding preceding sentences, and

- the coding results of the corresponding subsequent sentences.

[172] Obviously, using the formula (3) context information can be entered separately for each word in the text to be translated.

то есть в выходные данные Н_с подсистемы декодирования, как показано на фиг. 3.Through the encoding process described above, the semantic information of the text to be translated is converted into

that is, the output H _from the decoding subsystem as shown in FIG. 3.

[173] It should be noted that H _c not only includes semantic information

text to be translated, but also takes into account {λ _-k , λ _-k-1 , …, λ _-1 } and {λ ₁ , λ ₂ , …, λ _k } to dynamically select the contextual information used, which eliminates information redundancy caused by the overlay of information, and improve the accuracy of the translation result.

[174] Thus, the encoding method proposed in this embodiment of the present invention allows you to enter semantic information in the relevant text or part thereof, at the encoding end, as well as enter information about the degrees of correlation between the text to be translated and the entered relevant text. Thus, dynamic selection of contextual information for the text to be translated is provided, which not only reduces redundant information in the information received by the encoding-decoding model, but also improves the accuracy of the translation result of the text to be translated.

[175] Fourth Embodiment

[176] This embodiment of the present invention includes a specific implementation of step S404 mentioned in the second embodiment of the present invention.

или

Эти результаты кодирования могут использоваться как вспомогательные при декодировании целевого результата Н_с кодирования с получением перевода текста, подлежащего переводу.[177] In this embodiment, "translating the text to be translated by decoding the target encoding result" in the above step S404 includes specifically decoding, in the case where the relevant text includes the preceding target text, the target encoding result using the result encoding the preceding target text in the relevant encoding result. The encoding result of the preceding target text in step S402 in the second embodiment of the present invention may be

or

These encoding results can be used as an aid in decoding the target encoding result Hc _to obtain a translation of the text to be translated.

[178] In one implementation of this embodiment, decoding the target encoding result using the encoding result of the previous target text in the relevant encoding result may include steps A3 and B3 described below.

[179] In step A3, the degree of correlation between the text to be translated and the previous target text is determined.

[180] The previous target text includes the result of the translation of each previous sentence in the previous source text (hereinafter referred to as translated sentences). The degree of correlation between each translated sentence and the text to be translated can be determined separately.

[181] In one implementation, step A3 may include, in particular, determining the degrees of correlation between the text to be translated and the previous source text as the degrees of correlation between the text to be translated and the previous target text. In such an implementation, the degrees of correlation between the text to be translated and the corresponding translated sentences of the preceding target text can be respectively represented by correlation coefficients {λ _-k , λ _-k-1 , …, λ _-1 } corresponding to the corresponding previous sentences of the preceding source text and calculated in step A2 of the third embodiment of the present invention. For example, λ _-k represents the correlation between the text to be translated and the kth sentence before the text to be translated.

[182] In step B3, the target encoding result is decoded according to the determined degree of correlation and the encoding result of the previous target text.

кодирования (со ссылкой на шаг С1 второго варианта осуществления изобретения) для предшествующего целевого текста выполняют согласно следующей формуле:[183] In this step, the calculation using the coefficients {λ _-k , λ _-k-1 , ..., λ _-1 } of the correlation and results

encoding (with reference to step C1 of the second embodiment of the invention) for the previous target text is performed according to the following formula:

[184]

представляет результат кодирования предшествующего целевого текста, который используют при декодировании. Когда декодируют целевой результат Н_с кодирования, выполняют последовательное декодирование посредством RNN-сети. В данной реализации в любой момент в процессе декодирования может добавляться

согласно следующей формуле:[185] As shown in FIG. 3,

represents the encoding result of the preceding target text, which is used in decoding. When _the target encoding result Hc is decoded, serial decoding is performed by the RNN. In this implementation, at any time during the decoding process,

according to the following formula:

[186]

[187] where g is a non-linear function, such as a sigmoid function or an LSTM network; W, U, and V may be known decoding network learning parameters, and P is a newly introduced decoding network learning parameter in this embodiment of the present invention; s _j represents the hidden state at the current time j; s _j-1 represents the hidden state at the last j-1 time; y _j-1 is a word vector of the encoding result at the last time point, and c _j is the output of the attention model, which can be a network structure in a common encoding-decoding structure using the same encoding and calculation methods as in the prior art .

[188] After obtaining, by decoding, the hidden state s _j at the current time j, the probability distribution for the translation results of the word being translated at the current time j is estimated using the soft max function in the overall encoding-decoding structure. The translation result of the translated word is obtained in accordance with the probability distribution.

[189] Thus, the decoding method according to this embodiment of the present invention includes the input of the translation result of the previous source text, that is, the previous target text, at the decoding end, as well as the input of information about the degrees of correlation between the text to be translated and the previous target text. text. Thus, the result of the translation of the text to be translated can be related to the result of the translation of the previous text, which not only improves the accuracy of the translation of the text to be translated, but also improves the smoothness of the entire text of the translation as a whole, that is, the coherence and consistency of the sentences of the translated text are guaranteed. at the discourse level.

[190] Fifth Embodiment

[191] In this embodiment, the present invention provides a device for translating text at the level of discourse, while for a corresponding description, you should refer to the above embodiments of the method.

[192] FIG. 7 is a block diagram of a device for translating text at the discourse level in accordance with this embodiment of the present invention. Device 700 contains:

[193] a translatable text acquisition unit 701, which is configured to acquire a translatable text, wherein the translatable text is a unit of text in a discourse-level text to be translated;

[194] a relevant text obtaining block 702, which is configured to obtain relevant text for the text to be translated, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text, and under the previous source text means at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is understood to mean at least one unit of text after the text to be translated in the text at the discourse level, and the preceding target text is the text of the target said prior source text, and

[195] A text to be translated unit 703 that is configured to translate the text to be translated according to the relevant text.

[196] In one implementation of this embodiment of the present invention, the text-to-translate block 703 can be configured, in particular, to translate the text to be translated according to the relevant text using a pre-built encoding-decoding model.

[197] In one implementation of this embodiment of the present invention, the translation block 703 of the text to be translated may comprise:

[198] the first encoding subunit, configured to obtain the initial encoding result by encoding the text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated;

[199] a second encoding subunit, configured to obtain a relevant encoding result by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text;

[200] a result processing subunit configured to obtain a target encoding result by processing the initial encoding result using all or part of the encoding results in the relevant encoding result, wherein the target encoding result represents the semantic information of the text to be translated and carries all the semantic information of the relevant text or part of it, and

[201] A decoding execution sub-unit configured to translate text to be translated by decoding a target encoding result.

[202] In one implementation of this embodiment of the present invention, the first encoding subunit may be configured, in particular, to obtain the initial encoding result by encoding each word of the text to be translated.

[203] Accordingly, the second encoding sub-unit may be configured specifically to encode the words of each text unit in the relevant text to obtain encoding results corresponding to the respective text units and to determine said encoding results as the relevant encoding result.

[204] In one implementation of this embodiment of the present invention, the translation block 703 of the text to be translated may also comprise:

[205] the first addition sub-unit, configured to add together, after encoding the words of the text to be translated, the results of encoding the words of the text to be translated, obtaining a total result corresponding to the text to be translated, and

[206] A second adder configured to add together, after encoding the words of each text unit in the relevant text, the word encoding results of each text unit in the preceding source text in the case where the relevant text includes the preceding source text, to obtain a total result corresponding to each unit of text, adding together the word encoding results of each text unit in the subsequent source text, in the case where the relevant text includes the subsequent source text, to obtain a total result corresponding to each text unit, and adding together the word encoding results of each text unit in the previous target text in when the relevant text includes the previous target text, with a total result corresponding to each unit of text.

[207] In one implementation of this embodiment of the present invention, the result processing subunit comprises:

[208] the first determination sub-unit, configured to determine, in the case where the relevant text includes the previous source text, an encoding result of the previous source text from the relevant encoding result;

[209] a second determination sub-unit, configured to determine, in the case where the relevant text includes the subsequent source text, an encoding result of the subsequent source text from the relevant encoding result, and

[210] a result obtaining subunit configured to process the initial encoding result according to the determined encoding result to obtain a target encoding result.

[211] In one implementation of this embodiment of the present invention, the result obtaining sub-unit comprises:

[212] the first correlation degree determination sub-unit, configured to determine, in the case where the relevant text includes the previous source text, the degree of correlation between the text to be translated and the previous source text according to the initial encoding result of the text to be translated and the encoding result the previous source text,

[213] a second correlation degree determination sub-unit, configured to determine, in the case where the relevant text includes the subsequent source text, the degrees of correlation between the text to be translated and the subsequent source text according to the initial encoding result of the text to be translated and the encoding result subsequent source text, and

[214] a target result obtaining sub-unit, configured to process the initial encoding result in accordance with the determined encoding result and the determined degree of correlation, to obtain the target encoding result.

[215] In one implementation of this embodiment of the present invention, the decoding execution subunit is specifically configured to decode, in the case where the relevant text includes the previous target text, the target encoding result using the encoding result of the previous target text in the relevant encoding result.

[216] In one implementation of this embodiment of the present invention, the decoding execution subunit comprises:

[217] a third correlation degree determination sub-unit, configured to determine the degree of correlation between the text to be translated and the preceding target text, and

[218] An encoding result processing sub-unit, configured to decode the target encoding result according to the determined degree of correlation and the encoding result of the previous target text.

[219] In one implementation of this embodiment of the present invention, the third correlation degree determination subunit is configured to specifically determine the degree of correlation between the text to be translated and the previous source text as the degree of correlation between the text to be translated and the previous target text. .

[220] Sixth embodiment of the invention

[221] In this embodiment of the present invention, another device for translating text at the discourse level is proposed, and for a corresponding description, refer to the above embodiments of the method.

[222] FIG. 8 is a hardware block diagram of a device for translating text at the discourse level in accordance with this embodiment of the present invention. The voice interaction device 800 includes a memory 801, a receiver 802, and a processor 803 associated with the memory 801 and the receiver 802, wherein the memory 801 is configured to store a set of program instructions, and the processor 803 is configured to call the program instructions stored in the memory 801 to perform the following operations:

[223] obtaining the text to be translated, wherein the text to be translated is a unit of text in the text at the discourse level to be translated;

[224] obtaining relevant text for the text to be translated, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text,

[225] while the preceding source text is understood to be at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is understood to be at least one unit of text after the text to be translated in the text at the discourse level , and by the previous target text is meant the text of the translation of the said previous source text, and

[226] translation of the text to be translated according to the relevant text.

[227] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operation:

[228] translating the text to be translated according to the relevant text using a pre-built encoding-decoding model.

[229] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[230] obtaining an initial encoding result by encoding the text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated;

[231] obtaining a relevant encoding result by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text;

[232] obtaining a target encoding result by processing the initial encoding result using all or part of the encoding results in the relevant encoding result, wherein the target encoding result represents the semantic information of the text to be translated and carries all or part of the semantic information of the relevant text, and

[233] translating the text to be translated by decoding the target encoding result.

[234] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[235] obtaining the initial encoding result by encoding the words of the text to be translated, and

[236] encoding the words of each text unit in the relevant text to obtain encoding results corresponding to the corresponding text units, and determining said encoding results as the relevant encoding result.

[237] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[238] obtaining a total result corresponding to the text to be translated by adding together the results of encoding the words of the text to be translated;

[239] in the case where the relevant text includes the preceding source text, adding together the word encoding results of each text unit in the preceding source text to obtain a total result corresponding to each text unit;

[240] in the case where the relevant text includes subsequent source text, adding together the word encoding results of each text unit in the subsequent source text to obtain a total result corresponding to each text unit, and,

[241] in the case where the relevant text includes the preceding target text, adding together the word encoding results of each text unit in the preceding target text to obtain a total result corresponding to each text unit.

[242] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[243] determining, in the case where the relevant text includes the previous source text, the encoding result of the previous source text from the relevant encoding result;

[244] determining, in the case where the relevant text includes the subsequent source text, the encoding result of the subsequent source text from the relevant encoding result, and

[245] processing the initial encoding result according to the determined encoding result.

[246] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[247] determining, in the case where the relevant text includes the previous source text, the degree of correlation between the text to be translated and the previous source text according to the initial encoding result of the text to be translated and the encoding result of the previous source text;

[248] determining, in the case where the relevant text includes subsequent source text, the degree of correlation between the text to be translated and the subsequent source text according to the initial encoding result of the text to be translated and the encoding result of the subsequent source text, and

[249] processing the initial encoding result in accordance with the determined encoding result and the determined degree of correlation.

[250] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operation:

[251] decoding, in the case where the relevant text includes the previous target text, the encoding target result using the encoding result of the previous target text in the relevant encoding result.

[252] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operations:

[253] determining the degree of correlation between the text to be translated and the previous target text, and

[254] decoding the target encoding result in accordance with the determined degree of correlation and the encoding result of the previous target text.

[255] In one implementation of this embodiment of the present invention, the processor 803 is configured to call program instructions stored in memory 801 to perform the following operation:

[256] determining the degree of correlation between the text to be translated and the previous source text as the degree of correlation between the text to be translated and the previous target text.

[257] In some embodiments of the present invention, processor 803 may be a Central Processing Unit (CPU), memory 801 may be internal storage in the form of Random Access Memory (RAM), and receiver 802 may include A standard physical interface, which can be an Ethernet interface or an Asynchronous Transfer Mode (ATM) interface. Processor 803, receiver 802, and memory 801 may be integrated into one or more independent circuits or hardware units, such as custom integrated circuits (Application Specific Integrated Circuit, ASIC).

[258] In accordance with this embodiment of the present invention, a computer-readable storage medium is provided. A computer-readable storage medium contains instructions that, when executed on a computer, cause that computer to perform any of the above-described methods of translating text at the discourse level.

[259] According to this embodiment, the present invention also provides a computer program product that, when executed on a terminal device, causes that terminal device to perform any of the discourse-level text translation methods described above.

[260] From the above description of the embodiments of the present invention, it will be apparent to those skilled in the art that all or part of the steps in the discussed embodiments of the methods may be implemented by software on an appropriate standard hardware platform. Therefore, the essence or that part of the proposed technical solutions, which contributes to the prior art, can be implemented as a software product. The computer program product may be stored on a storage medium, such as ROM/RAM memory, a magnetic or optical disk, or similar medium containing a set of instructions to enable execution, by a computing device (which may be a personal computer, a server, or a network communication device such as as a media gateway, etc.), methods according to various embodiments of the present invention or parts thereof.

[261] It should be noted that the various embodiments of the present invention discussed in this application are listed sequentially, with the differences of each of the embodiments of the invention from the rest; therefore, for a description of identical or similar parts in various embodiments of the present invention, you can refer to other options. The device according to the embodiments of the present invention is described briefly as it corresponds to the method according to the embodiments of the present invention, and the description of the method can be referred to for a description of the relevant parts.

[262] It should also be understood that terms indicating relationships, such as "first" or "second", are used herein solely to distinguish one entity or operation from another entity or operation, and do not necessarily imply or require the existence of an actual order or relationship between these objects or operations. In addition, the terms "comprises", "comprises" and any variations thereof are intended to be non-exclusive, so that a process, method, object, or device incorporating a set of elements may include not only these elements, but any others not explicitly listed, or elements inherent in a process, method, object, or device. Without further limitation, the expression "includes" before an element does not exclude the presence of other such elements in a process, method, object, or device that includes said element.

[263] The above description of the embodiments of the invention enables those skilled in the art to make or practice the present invention. Various modifications of the embodiments of the present invention may be apparent to those skilled in the art, and the general principle of the invention disclosed herein may be implemented in other embodiments thereof, within the spirit of the present invention. Accordingly, the present invention is not limited to the embodiments described herein, and its essence is defined in the broadest scope in accordance with the principle and new features of the invention.

Claims (39)

1. A computer-implemented method of translating text at the discourse level, including: receiving the text to be translated, while the text to be translated is a unit of text in the text at the discourse level to be translated, obtaining the relevant text for the text to be translated, wherein the relevant text includes at least one of the following: the previous source text, the subsequent source text, or the previous target text, wherein the preceding source text is at least one unit of text before the text to be translated in the text at the discourse level, the subsequent source text is at least one unit of text after the text to be translated in the text at the discourse level, and the preceding target the text is the translation text of said prior source text, and translation of the text to be translated according to the relevant text, while the translation of the text to be translated, in accordance with the relevant text, includes: obtaining an initial encoding result by encoding the text to be translated, wherein the initial encoding result represents the semantic information of the text to be translated; obtaining a relevant encoding result by encoding the relevant text, wherein the relevant encoding result represents the semantic information of the relevant text; obtaining a target encoding result by processing the initial encoding result using all or part of the encoding results in the relevant encoding result; and translating the text to be translated by decoding the target encoding result. 2. The method of claim 1, wherein translating the text to be translated according to the relevant text includes: translating the text to be translated according to the relevant text using a prebuilt encoding-decoding model. 3. The method of claim 2, wherein the target encoding result represents the semantic information of the text to be translated and carries all or part of the semantic information of the relevant text. 4. The method according to claim 3, wherein obtaining the initial encoding result by encoding the text to be translated includes: obtaining the initial encoding result by encoding the words of the text to be translated, while accordingly obtaining the relevant encoding result by encoding the relevant text includes: encoding the words of each text unit in the relevant text to obtain encoding results corresponding to the respective text units, and determining said encoding results as the relevant encoding result. 5. The method according to claim 4, also including, after encoding the words of the text to be translated: obtaining an overall result corresponding to the text to be translated by adding together the results of encoding the words of the text to be translated, whereby, respectively, after encoding the words of each unit of text in the relevant text, the method also includes: in the case where the relevant text includes the preceding source text, adding together the word encoding results of each text unit in the preceding source text to obtain a total result corresponding to each text unit; in the case where the relevant text includes subsequent source text, adding together the word encoding results of each text unit in the subsequent source text to obtain a total result corresponding to each text unit, and in the case where the relevant text includes the preceding target text, adding together the word encoding results of each text unit in the preceding target text to obtain a total result corresponding to each text unit. 6. The method of claim 3, wherein processing the initial encoding result using all or part of the encoding results in the relevant encoding result comprises: determining, in the case where the relevant text includes the preceding source text, an encoding result of the preceding source text from the relevant encoding result; determining, in the case where the relevant text includes the subsequent source text, the encoding result of the subsequent source text from the relevant encoding result, and processing the initial encoding result in accordance with the determined encoding result. 7. The method of claim 6, wherein processing the initial encoding result according to the determined encoding result comprises: Determining, in the case where the relevant text includes the previous source text, the degree of correlation between the text to be translated and the previous source text according to the initial encoding result of the text to be translated and the encoding result of the previous source text; determining, in the case where the relevant text includes the subsequent source text, the degree of correlation between the text to be translated and the subsequent source text according to the initial encoding result of the text to be translated and the encoding result of the subsequent source text, and processing the initial encoding result in accordance with the determined encoding result and the determined degree of correlation. 8. The method according to any one of paragraphs. 3-7, in which the translation of the text to be translated by decoding the target encoding result includes: decoding, in the case where the relevant text includes the preceding target text, of the encoding target result using the encoding result of the previous target text in the relevant encoding result. 9. The method of claim 8, wherein decoding the target encoding result using the encoding result of the preceding target text in the relevant encoding result includes: determining the degree of correlation between the text to be translated and the previous target text, and decoding the target encoding result in accordance with the determined degree of correlation and the encoding result of the previous target text. 10. The method of claim 9, wherein determining the degree of correlation between the text to be translated and the previous target text includes: determining the degree of correlation between the text to be translated and the previous source text as the degree of correlation between the text to be translated and the previous target text. 11. A device for translating text at the discourse level, comprising: processor, memory and system bus, wherein the processor and memory are coupled to each other via a system bus, and the memory is configured to store one or more computer programs containing instructions that, when executed by the processor, cause the processor to execute the method of any one of claims. 1−10. 12. A computer-readable storage medium containing instructions that, when executed on a computer, cause that computer to execute the method according to any one of paragraphs. 1−10.

Images