WO2019228016A1 - Intelligent writing method and apparatus - Google Patents

Abstract

Embodiments of the present description provide a method and apparatus for training a writing model, and a writing method and apparatus. The writing model comprises a word segmentation unit, a conversion unit, and a cyclic neural network. The training method comprises: obtaining a sample for training the writing model, the sample comprising a text and a calibration value for the text, and the calibration value indicating whether the text is smooth; and training the writing model by using the sample so that, after the text is inputted into the writing model, the absolute value of a difference between an output value and the calibration value of the writing model after the training is decreased with respect to that before the training.

Description

Intelligent writing method and device Technical field

The embodiments of the present specification relate to the field of machine learning technology, and more particularly, to a method and device for training a writing model, and a writing method and device.

Background technique

With the explosive development of today's new media, the means of publicity and publicity have become increasingly abundant, which has brought great challenges to writing and publicity. For each emergency and various social hotspot issues, not only need to respond in a timely manner, but also write articles of corresponding styles according to the characteristics of various media and readers, such as Weibo, WeChat, Zhihu, Headlines, and so on. This greatly increases the workload of writing. Therefore, automatic writing through intelligent writing tools can greatly improve the efficiency of writing, and seize the opportunity in brand publicity and public relations. Intelligent writing tools currently available include Giiso's intelligent writing robots and automatic writing tools in academia. In the above-mentioned existing intelligent writing tools, there is very little manual intervention in the process of writing a manuscript, and after the writing of the manuscript, a large amount of manpower is required for proofreading; in addition, the completed article is biased toward formatting, for example, the completed article It is tabular and cannot be used in multiple writing scenarios. Therefore, a more effective method of intelligent writing is needed.

Summary of the Invention

The embodiments of the present specification aim to provide a more effective intelligent writing method to solve the deficiencies in the prior art.

To achieve the above object, an aspect of the present specification provides a method for training a writing model, the writing model includes a word segmentation unit, a transformation unit, and a recurrent neural network, and the method includes: obtaining a sample for training the writing model, so The sample includes a text and a calibration value for the text, the calibration value indicating whether the text is smooth; and using the sample to train the writing model so that after the text is input to the writing model, Compared to before training, the absolute value of the difference between the output value of the writing model after training and the calibration value is reduced, and after the text is input to the writing model, the word segmentation unit The text obtains a plurality of participles arranged in order, wherein the plurality of participles are arranged in order according to a position before and after in the text, and the conversion unit converts the plurality of participles arranged in order into a corresponding order. Multiple word vectors, and the recurrent neural network outputs an output value of the writing model based on the sequentially arranged multiple word vectors.

In one embodiment, in the method for training a writing model, the text includes two non-adjacent texts in an existing article, and the calibration value is 0.

In one embodiment, in the method for training a writing model, the text includes two texts respectively belonging to two existing articles, and the calibration value is 0.

In one embodiment, in the method for training a writing model, the text sequence includes two texts arranged next to each other in an existing article, and the calibration value is 1.

In one embodiment, in the method for training a writing model, the existing article is an existing article belonging to a selected field, and the writing model is used for writing belonging to the selected area after the training. Articles in the field.

In one embodiment, in the method for training a writing model, training the writing model using the samples includes using the samples to train the writing model by a back propagation algorithm.

In one embodiment, in the method for training a writing model, the recurrent neural network includes one of the following networks: RNN, LSTM, and GRU.

Another aspect of the present specification provides a writing method, including: obtaining n first candidate texts of a first to-be-written text of an article to be written and m second candidate texts of a second to-be-written text, where n and m are predetermined The number, and in the article to be written, the first to-be-written text is located in front of the second to-be-written text and adjacent to the second to-be-written text; the first candidate text is Each text in is combined with each text in the second candidate text to obtain n × m input texts, where in each of the input texts, the first candidate text is located in the In front of the second candidate text; input the n × m input texts respectively into a writing model trained according to the above training method to obtain n × m model output values corresponding to the n × m input texts, Predicting the smoothness of the corresponding input text by the model output value; and determining the selected text of the first to-be-written text from the n first candidate texts based on the n × m model output values, and From the m second candidate texts Written text of the selected text to be the second.

Another aspect of the present specification provides a writing method, including: obtaining q candidate texts of to-be-written text in a to-be-completed article, where q is a predetermined number; and obtaining from the to-be-completed articles, the texts to be related to the to-be-written text Existing text of a predetermined length adjacent to and in front of the text to be written; combining the existing text and the q candidate texts separately to obtain q input texts, where in each of the input In the text, the existing text is located in front of the candidate text; the q input texts are respectively input into a writing model trained according to the above training method to obtain q models corresponding to the q input texts respectively An output value, the model output value predicts the smoothness of the corresponding input text; and a selected text of the text to be written is determined from the q candidate texts based on the q model output values.

In one embodiment, the writing method further includes, after determining the selected text, performing synonymous replacement of words and / or sentences in the selected text.

In one embodiment, in the writing method, obtaining the q candidate texts of the text to be written in the article to be completed includes acquiring at least one keyword of the text to be written; and based on the keywords To obtain q candidate texts of the text to be written.

In one embodiment, the writing method further includes, before acquiring at least one keyword of the text to be written, acquiring keywords of a topic of the article to be completed.

In one embodiment, in the writing method, obtaining the q candidate texts of the text to be written based on the keywords includes, according to a predetermined search and sorting method, selecting A search is performed in the corpus based on the keywords to obtain the q candidate texts.

In one embodiment, in the writing method, a corpus corresponding to the text to be written includes at least one of the following: a corpus corresponding to a field of the text to be written, and a form corresponding to the form of the text to be written Corpus.

Another aspect of the present specification provides a device for training a writing model. The writing model includes a word segmentation unit, a conversion unit, and a recurrent neural network. The device includes: an obtaining unit configured to obtain a sample for training the writing model. The sample includes a text and a calibration value for the text, the calibration value indicating whether the text is smooth; and a training unit configured to use the sample to train the writing model such that: After the model inputs the text, compared to before training, the absolute value of the difference between the output value of the writing model after training and the calibration value decreases, wherein the text is input to the writing model After that, the word segmentation unit obtains a plurality of sequentially arranged word segments from the text, wherein the plurality of segmented words are sequentially arranged according to their forward and backward positions in the text, and the conversion unit arranges the sequentially arranged plurality of Word segmentation is converted into a corresponding sequence of multiple word vectors, and the recurrent neural network outputs the writing based on the sequence of multiple word vectors Type output value.

In one embodiment, in the apparatus for training a writing model, the training unit is further configured to use the samples to train the writing model by a back propagation algorithm.

Another aspect of the present specification provides a writing device, including: an obtaining unit configured to obtain n first candidate texts of a first to-be-written text of an article to be written and m second candidate texts of a second to-be-written text, Where n and m are a predetermined number, and in the article to be written, the first to-be-written text is located in front of and adjacent to the second to-be-written text; combination A unit configured to combine each text in the first candidate text with each text in the second candidate text to obtain n × m input texts, wherein, in each of the input In the text, the first candidate text is located in front of the second candidate text; the input unit is configured to input the n × m input texts respectively through a writing model trained according to the above training method to obtain Describing n × m model output values corresponding to the n × m input texts respectively, the model output values predicting the smoothness of the corresponding input texts; and a determining unit configured to be based on the n × m model output values, From the n Determining a first candidate text to be written in the text of the selected text, and determining the selected text written text from the second to be the second candidate text m.

Another aspect of the present specification provides a writing device, including: a first obtaining unit configured to obtain q candidate texts of to-be-written text in a to-be-completed article, where q is a predetermined number; and a second obtaining unit configured to: Obtaining a predetermined length of existing text adjacent to the to-be-written text and in front of the to-be-written text from the to-be-completed article; a combination unit configured to combine the existing text with the The q candidate texts are respectively combined to obtain q input texts, wherein in each of the input texts, the existing text is located in front of the candidate texts; the input unit is configured to input the q input texts The text is input respectively through a writing model trained according to the above training method to obtain q model output values corresponding to the q input texts respectively, the model output values predicting the smoothness of the corresponding input texts; and a determining unit, configuration To determine the selected text of the text to be written from the q candidate texts based on the q model output values.

In one embodiment, the writing device further includes a replacement unit configured to, after determining the selected text, perform synonymous replacement of words and / or sentences in the selected text.

In one embodiment, in the writing device, the first acquisition unit includes a first acquisition subunit configured to acquire at least one keyword of the text to be written; and a second acquisition subunit, Based on the keywords, q candidate texts of the text to be written are obtained.

In one embodiment, the writing device further includes a third acquisition unit configured to acquire keywords of the subject of the article to be completed before acquiring at least one keyword of the text to be written.

In one embodiment, in the writing device, the second acquisition subunit is further configured to search based on the keywords in a corpus corresponding to the text to be written according to a predetermined search and sorting method, Thereby, the q candidate texts are obtained.

The intelligent drafting scheme according to the embodiment of the present specification proposes an algorithmic framework that can efficiently solve the intelligent drafting problem; solves the problem of corpus reuse by using synonymous replacement; and uses the method of inserting sentences after paragraphs to enrich the content of the article. Compared with the method of writing manuscripts that is too automated, this solution adds manual intervention, which greatly improves the practicability and accuracy; compared with the completely template-based method, it greatly enhances flexibility and scalability.

BRIEF DESCRIPTION OF THE DRAWINGS

By describing the embodiments of this specification with reference to the drawings, the embodiments of this specification can be made more clear:

FIG. 1 shows a schematic diagram of a system 100 according to an embodiment of the present specification;

FIG. 2 shows a method for training a writing model according to an embodiment of the present specification;

FIG. 3 illustrates a writing method according to an embodiment of the present specification;

FIG. 4 shows a writing method according to an embodiment of the present specification;

FIG. 5 illustrates an apparatus 500 for training a writing model according to an embodiment of the present specification;

FIG. 6 illustrates a writing device 600 according to an embodiment of the present specification; and

FIG. 7 illustrates a writing device 700 according to an embodiment of the present specification.

Detailed ways

The embodiments of the present specification will be described below with reference to the drawings.

FIG. 1 shows a schematic diagram of a system 100 according to an embodiment of the present specification. As shown in FIG. 1, the system 100 includes a writing model 11 and a selection unit 12. Among them, the writing model 11 further includes a word segmentation unit 111, a conversion unit 112, and a recurrent neural network 113. In the stage of training the writing model 11, first, a training sample is obtained, the training sample includes a text and a calibration value for the text, and the calibration value indicates whether the text is smooth. Then, using the training samples, the writing model 11 is trained by, for example, a BPTT algorithm (back propagation algorithm), so that after the text is input to the writing model, the writing model is compared with that before training. The absolute value of the difference between the output value after training and the calibration value decreases. After the text in the training sample is input into the writing model 11, in the writing model 11, the word segmentation unit 111 performs word segmentation on the text in the sample to obtain a plurality of sequenced word segmentations. After that, the conversion unit 112 converts the plurality of word segmentation into word vectors arranged in sequence corresponding to the plurality of word segmentations and sequentially inputs the word vectors to the recurrent neural network 113. Then, the recurrent neural network 113 calculates the word vectors that are sequentially input, thereby outputting a calculation result. In the process of training the model, in the recurrent neural network 113, based on the sequentially input word vectors and calibration values, parameters are adjusted according to, for example, the BPTT algorithm, so that the plurality of recurrent neural networks 113 are sequentially input. After the word vector, the output value of the network is closer to the calibration value. Multiple training samples may be obtained to train the writing model multiple times, that is, perform multiple parameter adjustments on the recurrent neural network 113, so that the output value of the writing model corresponding to the input sample is infinitely close to the input The calibration value of the sample.

In the stage of writing with the writing model 11, a plurality of input texts are first obtained, and a plurality of input texts are input to the writing model 11. For example, for the first paragraph and the second paragraph of the article to be written, n first candidate paragraphs for the first paragraph and m second candidate paragraphs for the second paragraph can be obtained by searching, Combine to obtain n × m input texts, and input the n × m input texts to the writing model 11 respectively. Each time an input text is input to the writing model 11, similarly, the word segmentation unit 111 first performs word segmentation on the text to obtain the sequenced word segmentation, and then the conversion unit 112 converts the group of sequenced word segmentation into a sequenced word Vector and input it to the recurrent neural network 113 in sequence. The recurrent neural network 113 is a network trained by the above-mentioned model. It calculates word vectors that are sequentially input, thereby outputting output values corresponding to the word vectors that are sequentially input, and sends the output values to the selection unit 12. After obtaining the n × m output values corresponding to the n × m input texts, the selection unit 12 may select the input text with the largest output value as the first paragraph and the second paragraph of the article to be written.

For the third paragraph of the article to be written, similarly, q third candidate paragraphs of the third paragraph can be obtained by searching. Then, the text of the selected second paragraph and the q third candidate paragraphs are respectively combined to obtain q input text. Similarly, by inputting the q input texts to the writing model 11 respectively, q model output values corresponding to the q input texts can be obtained. Finally, the selection unit 12 selects, from the q third candidate paragraphs, the paragraph corresponding to the largest model output value as the third paragraph. Similarly, subsequent paragraphs of the article to be written can be obtained.

The system 100 according to the embodiment of the present specification shown in FIG. 1 is merely exemplary. The system according to the embodiment of the present specification is not limited to the system 100. For example, the system 100 may further include a search unit for searching for text, and a combination unit. For combining text. In addition, the system 100 may also include other units, such as a replacement unit, for synonymous replacement of words and / or sentences in selected text of the article to be written, and so on. The manuscript writing model 11 is not limited to processing paragraphs of articles to be written. For example, based on the corresponding corpus, sentences can be inserted into existing articles, supplementary content, and the like.

FIG. 2 illustrates a method for training a writing model according to an embodiment of the present specification. The writing model includes a word segmentation unit, a transformation unit, and a recurrent neural network. The method includes:

In step S21, a sample for training the writing model is obtained, the sample includes a text and a calibration value for the text, and the calibration value indicates whether the text is smooth; and

In step S22, using the sample to train the writing model, so that after the text is input to the writing model, compared to before training, the output value of the writing model after the training is compared with the calibration The absolute value of the difference in values decreases, wherein after the text is input to the writing model, the word segmentation unit obtains a plurality of sequenced word segmentations from the text, where the plurality of word segmentations are based on The text is arranged sequentially in the front and rear positions, the conversion unit converts the sequentially arranged word segments into correspondingly arranged multiple word vectors, and the recurrent neural network outputs based on the sequentially arranged multiple word vectors An output value of the writing model.

First, in step S21, a sample for training the writing model is obtained, the sample includes a text and a calibration value for the text, and the calibration value indicates whether the text is smooth. In the embodiment of the present specification, the writing model includes a word segmentation unit, a conversion unit, and a recurrent neural network. The recurrent neural network is, for example, RNN, LSTM, GRU, or the like. The common feature of these recurrent neural networks is that the final output value is obtained by sequentially inputting a set of vectors, wherein the calculation of the vector input at a certain time point takes into account one or more vectors input before the time point. That is, the recurrent neural network characterizes the relationship between the current output of a sequence and the previous information.

In one embodiment, two non-contiguous texts from an existing article can be combined into a sample text, and in this case, since the two texts are non-contiguous in the original article, that is, its It is not smooth to combine them, so mark the sample text as 0, which means that it is not smooth. In one embodiment, one text may be obtained from two different existing articles and combined into a sample text. Similarly, the sample text is not smooth and its calibration value is also 0. In another embodiment, two adjacently arranged texts are obtained from an existing article, and the two texts are combined into a sample text in the original order. Since the two texts in the sample text are in the existing article, The middle is adjacent and arranged in the original order, that is, the sample text is smooth, so the sample text is marked as 1, which means smooth. Among them, the text selected from the existing article can be a sentence or a paragraph, which can be used to train a writing model.

Therefore, a large number of training samples can be easily obtained by acquiring existing articles. In the embodiment of the present specification, the writing model is repeatedly trained by acquiring samples of the order of 100,000 to one million, that is, the method shown in FIG. 2 is repeatedly repeated, thereby training a more accurate writing model. The existing articles can be obtained from a website, can be obtained from a search, can be obtained from a database, and the like. In one embodiment, existing articles in a selected field may be obtained to train a writing model for writing articles in the selected field. The field can be divided in different dimensional spaces. For example, in the dimensional space of content, the field can include entertainment, sports, technology, finance and other categories. In the dimensional space of style, the field can include micro Bo, WeChat, Zhihu, headlines and many other categories. For example, in the case of a selected field of entertainment news, multiple existing articles can be obtained from existing entertainment news, and training samples can be obtained from them to train a writing model. In this case, the trained writing model can be used to write articles in the field of entertainment news.

In step S22, using the sample to train the writing model, so that after the text is input to the writing model, compared to before training, the output value of the writing model after the training is compared with the calibration The absolute value of the difference in values decreases.

Wherein, after the text is input to the writing model, the segmentation unit obtains a plurality of segmented words arranged sequentially from the text, wherein the plurality of segmented words are sequentially arranged according to their forward and backward positions in the text. The word segmentation method may, for example, perform word segmentation based on thesaurus comparison, word segmentation based on a deep learning model, and the like. The conversion unit converts the sequentially arranged plurality of word segmentation into correspondingly arranged plurality of word vectors, wherein the segmentation can be converted into the corresponding by a specific model, for example, a word2vec (word-vector) model. Word vector, that is, a computer-understood dense vector. Thereby, a sequenced word vector corresponding to a plurality of sequenced participles can be obtained. The recurrent neural network outputs the output value of the writing model based on the plurality of word vectors arranged in order, that is, the recurrent neural network performs a calculation for each input word vector, except that the calculation is based on the word vector. In addition to the eigenvalues of, the calculation is based on the previous word vector. Therefore, after performing multiple corresponding calculations on multiple word vectors that are input sequentially, the output results reflect the Dependency, which can be used to calculate the smoothness of the text.

In one embodiment, the model is trained by a back-propagation algorithm (BPTT algorithm). The training process can be divided into three steps: forward calculation, calculation of loss value and backward calculation. First, an existing model is used to calculate an output value after sequentially inputting the plurality of word vectors to the model, and an intermediate result at each step of calculating each word vector. Then, calculate the loss value at each step, accumulate the loss value at each step, and average them. For example, for a sample with a calibration value of 1, subtract 1 from the model output value to obtain the loss value for the last step of the model calculation. The parameters of the neurons in the last layer are pushed forward to obtain the loss value of each neuron in the previous layer. Finally, the back propagation calculation is performed based on the results of the forward propagation, and the gradient value of each step is accumulated. Then, the gradient descent method is used to adjust the parameters of the existing model, so that after the adjustment, the output results of the model on the multiple word vectors are closer to the calibration value (for example, "1"). By inputting a large number of training samples to the model to adjust the parameters multiple times, the output result of the model on the training samples is infinitely close to its calibration value.

However, the training method is not limited to the above-mentioned back-propagation algorithm, and can also be trained by the existing improved algorithm of BPTT, or the model can be trained by using deep learning libraries such as Tensorflow, Torch, Theano, and the like.

FIG. 3 illustrates a writing method according to an embodiment of the present specification, including:

In step S31, n first candidate texts of the first to-be-written text of the article to be written and m second candidate texts of the second to-be-written text are obtained, where n and m are a predetermined number, and wherein In writing an article, the first text to be written is located in front of the second text to be written and is adjacent to the second text to be written;

In step S32, each text in the first candidate text is combined with each text in the second candidate text, thereby obtaining n × m input texts, wherein, in each of the input texts, The first candidate text is located in front of the second candidate text;

In step S33, the n × m input texts are respectively input into a writing model trained according to the model training method described above to obtain n × m model output values corresponding to the n × m input texts respectively. How well the model output predicts the corresponding input text; and

In step S34, based on the n × m model output values, the selected text of the first to-be-written text is determined from the n first candidate texts, and the selected text is determined from the m second candidate texts. The selected text of the second text to be written is described.

First, in step S31, n first candidate texts of the first to-be-written text of the article to be written and m second candidate texts of the second to-be-written text are obtained, where n and m are a predetermined number, and In the article to be written, the first text to be written is located in front of the second text to be written and is adjacent to the second text to be written. The writing method shown in FIG. 3 is implemented by a writing model trained with the training method shown in FIG. 2. For example, during the training phase, the writing model is trained through a corpus (training text) in the field of entertainment news, that is, the writing model is A model for writing entertainment news. When writing a new entertainment news through the above writing model, for the first paragraph (that is, the first text to be written) and the second paragraph (that is, the second text to be written) of the article to be written, the user can provide the first paragraph At least one keyword, such as keywords about people, time, background, and at least one keyword in the second paragraph, such as keywords about events, passing. Then, according to a predetermined search and ranking strategy, n candidate texts corresponding to the first paragraph and m candidate texts corresponding to the second paragraph may be searched in a corpus corresponding to the entertainment news field and corresponding to the paragraphs. Among them, n and m are predetermined numbers, which may be on the order of tens or hundreds, for example. A corpus in a specific domain and a specific form may be formed by obtaining articles in a specific domain and obtaining text in a specific form (for example, paragraphs, sentences, etc.) from the articles. Articles in the specific field (for example, entertainment news field) can be obtained from a specific field column of the website, can be obtained by searching, or can be obtained from a database.

The above method for obtaining candidate text is only schematic. The first to-be-written text and the second to-be-written text are not limited to paragraphs, and are not limited to the first and second paragraphs of the article to be written, and may be, for example, two sentences or two parts of text in a paragraph. The method for obtaining candidate text is not limited to search and acquisition through keywords, for example, it can be obtained through manual finishing, or candidate text can be obtained through machine learning, and so on. In addition, the corpus is not limited to the above-mentioned corpus in a specific field and a specific form. For example, the corpus may be a corpus corresponding only to a domain, and correspondingly, the candidate text searched by the corpus is not limited to a specific form, for example, it may be a sentence or a paragraph. For another example, the corpus may be a corpus corresponding only to the form of the text to be written. Correspondingly, the candidate text searched by the corpus is not limited to a specific field, for example, it may be in the entertainment field or the financial field. The corpus included in the corpus can be set according to the needs of specific scenarios. In one embodiment, the field of the corpus is selected according to the applicable field of the writing model. For example, the writing model is a model for writing articles in the field of entertainment news. When writing using this writing model, the corpus of the field of entertainment news is selected. Search for candidate text for text to write.

In step S32, each text in the first candidate text is combined with each text in the second candidate text, thereby obtaining n × m input texts, wherein, in each of the input texts, The first candidate text is located in front of the second candidate text. For example, in the above-mentioned case of obtaining the candidate text of the first paragraph and the candidate text of the second paragraph of the article to be written, as an illustration, the sentences in one paragraph are represented by letters. For example, a candidate text in the first paragraph is "a, b, c." And a candidate text in the second paragraph is "d, e." By combining the previous two candidate texts in order, the candidate for the first paragraph is The text is placed in front of the candidate text in the second paragraph, and an input text "a, b, c. D, e." Is obtained. By combining each of the n candidate texts of the first paragraph with each of the m candidate texts of the second paragraph as described above, n × m input texts can be obtained.

In step S33, the n × m input texts are respectively input into a writing model trained according to the model training method described above to obtain n × m model output values corresponding to the n × m input texts respectively. The model output value predicts the smoothness of the corresponding input text.

For each of the n × m input texts, after inputting it into a writing model trained according to the above-mentioned model training method, the writing model outputs a calculation result through a process as described above with reference to FIG. 2. That is, the word segmentation unit in the writing model obtains a plurality of word segmentations arranged sequentially from the text, wherein the plurality of word segmentations are sequentially arranged according to their forward and backward positions in the text. The conversion unit converts the sequentially arranged plurality of word segmentation into a correspondingly arranged plurality of word vectors through a word2vec (word-vector) model, for example. The recurrent neural network outputs an output value of the writing model based on the sequentially arranged plurality of word vectors.

Because during the training of the writing model, the text including consecutive paragraphs is marked as 1, which means smooth, and the text including discontinuous paragraphs is marked as 0, which means not smooth. Therefore, the output value of the writing model relative to the input text indicates the smoothness of the input text. The closer the output value is to 1, the smoother the input text is. The closer the output value is to 0, the less smooth the input text is. .

In step S34, based on the n × m model output values, the selected text of the first to-be-written text is determined from the n first candidate texts, and the selected text is determined from the m second candidate texts. The selected text of the second text to be written is described. After obtaining n × m model output values corresponding to the n × m input texts from the writing model, it can be determined that the input text corresponding to the maximum value of the n × m model output values is the selected text, that is, The candidate text included in the first paragraph is the selected text of the first paragraph, and the candidate text included in the second paragraph is the selected text of the second paragraph. That is, the smoothest combination of the pairwise combinations of the n first candidate texts and the m second candidate texts is used as the selected text of the article to be written.

FIG. 4 illustrates a writing method according to an embodiment of the present specification, including:

In step S41, q candidate texts of the text to be written in the article to be completed are acquired, where q is a predetermined number;

In step S42, obtaining a predetermined length of existing text that is adjacent to the text to be written and is located in front of the text to be written from the to-be-completed article;

In step S43, the existing text and the q candidate texts are respectively combined to obtain q input texts, wherein in each of the input texts, the existing texts are located in front of the candidate texts ;

In step S44, the q input texts are respectively input into a writing model trained according to the above model training method to obtain q model output values corresponding to the q input texts respectively, and the model output values predict corresponding The fluency of the input text; and

In step S45, based on the q model output values, a selected text of the text to be written is determined from the q candidate texts.

First, in step S41, q candidate texts of the text to be written in the article to be completed are acquired, where q is a predetermined number. For example, for the to-be-completed articles in the field of entertainment news in which the selected texts of the first paragraph and the second paragraph have been determined, at least one keyword of the third paragraph may be provided by the user, for example, keywords related to event impact and public opinion. Then, according to a predetermined search and ranking strategy, q candidate texts corresponding to the third paragraph are searched in a corpus corresponding to the entertainment news and corresponding to the paragraphs, where q is a predetermined number, which can be, for example, dozens, several Hundreds of magnitude.

Here, the method for obtaining the candidate text is only schematic. The text to be written is not limited to a paragraph, and it may be, for example, a sentence or a part of text in a paragraph. The method for obtaining candidate text is not limited to search and acquisition through keywords, for example, it can be obtained through manual finishing, or candidate text can be obtained through machine learning, and so on.

In step S42, a predetermined length of existing text that is adjacent to the text to be written and located in front of the text to be written is obtained from the to-be-completed article. For example, in the above-mentioned article to be completed in the field of entertainment news, the text to be written is the third paragraph of the article, and the existing text of a predetermined length adjacent to the text to be written and in front of the text to be written may be the article The second paragraph. The predetermined length may be determined according to the length of the text to be written. For example, if the text to be written is a sentence, a sentence before the sentence to be written may be obtained in the article to be completed as the existing text.

In step S43, the existing text and the q candidate texts are respectively combined to obtain q input texts, wherein in each of the input texts, the existing texts are located in front of the candidate texts . For example, for the above-mentioned articles in the field of entertainment news, the selected text of the second paragraph and the q candidate texts of the third paragraph are respectively combined to obtain q input text.

Step S44 is basically the same as the description of step S33 shown in FIG. 3, and is not repeated here.

In step S45, based on the q model output values, a selected text of the text to be written is determined from the q candidate texts. For example, in the above example, after obtaining q model output values corresponding to q input texts respectively from the writing model, it may be determined that the candidate text included in the input text corresponding to the maximum value among the q model output values is The selected text, that is, the candidate text of the third paragraph included therein is the selected text of the third paragraph. That is, the candidate text of the third paragraph that is most smoothly combined with the selected text of the second paragraph is used as the selected text of the third paragraph.

In this method, the text to be written can be a sentence, a paragraph, or a part of a paragraph. By using the above writing model to determine the most smooth insertion text that is connected to the existing paragraphs of the article, it can be used for other content. Existing articles are enriched and supplemented.

In an embodiment, after the selected text of the text to be written is determined through the above writing model, in order to avoid simple repeated use, the words and sentences in the selected text may be replaced synonymously, for example, in the selected text. Replace words with synonyms, replace sentences in the selected text with another sentence with the same meaning but different expression, etc. Among them, the above replacement can be performed through a synonym thesaurus and a synonym corpus.

In one embodiment, in obtaining candidate text of the text to be written, in addition to acquiring at least one keyword of the text to be written, a keyword of a topic of the article to be written may also be obtained. Therefore, based on the keywords of the topic and the keywords of the text to be written, a search strategy is set to search for candidate text. For example, in the corpus, the corpus can be filtered first by using the topic keywords, and then the candidate text is searched from the filtered corpus by the keywords of the text to be written.

FIG. 5 illustrates an apparatus 500 for training a writing model according to an embodiment of the present specification. The writing model includes a word segmentation unit, a transformation unit, and a recurrent neural network. The apparatus 500 includes:

The obtaining unit 51 is configured to obtain a sample for training the writing model, the sample including a text and a calibration value for the text, the calibration value indicating whether the text is smooth or not; and

The training unit 52 is configured to use the sample to train the writing model such that after the text is input to the writing model, compared to before training, the output value of the writing model after the training is equal to The absolute value of the difference between the calibration values is reduced, wherein after the text is input to the writing model, the word segmentation unit obtains a plurality of word segmentations arranged in sequence from the text, wherein the plurality of word segmentations are based on It is arranged sequentially in front and back in the text, the conversion unit converts the sequentially arranged plurality of word segmentation into correspondingly arranged multiple word vectors, and the recurrent neural network is based on the sequentially arranged multiple The word vector outputs the output value of the writing model.

In one embodiment, in the apparatus for training a writing model, the training unit is further configured to use the samples to train the writing model by a back propagation algorithm.

FIG. 6 illustrates a writing device 600 according to an embodiment of the present specification, including:

The obtaining unit 61 is configured to obtain n first candidate texts of the first to-be-written text of the article to be written and m second candidate texts of the second to-be-written text, where n and m are a predetermined number, and wherein In the article to be written, the first to-be-written text is located in front of the second to-be-written text and is adjacent to the second to-be-written text;

The combining unit 62 is configured to combine each text in the first candidate text with each text in the second candidate text, thereby obtaining n × m input texts. In the input text, the first candidate text is located in front of the second candidate text;

The input unit 63 is configured to input the n × m input texts respectively into a writing model trained according to the training method to obtain n × m model output values corresponding to the n × m input texts, respectively. The model output value predicts the smoothness of the corresponding input text; and

The determining unit 64 is configured to determine the selected text of the first to-be-written text from the n first candidate texts based on the n × m model output values, and from the m second candidates The text determines the selected text of the second text to be written.

FIG. 7 illustrates a writing device 700 according to an embodiment of the present specification, including:

The first obtaining unit 71 is configured to obtain q candidate texts of a text to be written in a to-be-completed article, where q is a predetermined number;

A second obtaining unit 72 configured to obtain, from the to-be-completed article, an existing text of a predetermined length that is adjacent to the to-be-written text and is located in front of the to-be-written text;

The combining unit 73 is configured to combine the existing text and the q candidate texts separately to obtain q input texts, where in each of the input texts, the existing texts are located in the candidate Before the text

The input unit 74 is configured to respectively input the q input texts into a writing model trained according to the training method to obtain q model output values corresponding to the q input texts respectively, and the model output values are predicted The fluency of the corresponding input text; and

The determining unit 75 is configured to determine the selected text of the text to be written from the q candidate texts based on the q model output values.

In an embodiment, the writing device 600 (700) further includes a replacement unit 65 (76) configured to, after determining the selected text, perform the same operation on the words and / or sentences in the selected text. Meaning replacement.

In one embodiment, in the writing device 700, the first obtaining unit 71 includes a first obtaining subunit 711 configured to obtain at least one keyword of the text to be written; and a second obtaining The subunit 712 obtains q candidate texts of the text to be written based on the keywords.

In one embodiment, the writing device 700 further includes a third obtaining unit 77 configured to obtain keywords of a topic of the article to be completed before obtaining at least one keyword of the text to be written.

In one embodiment, in the writing device, the second acquisition subunit 712 is further configured to perform a search based on the keywords in a corpus corresponding to the text to be written according to a predetermined search and sorting method. To obtain the q candidate texts.

The intelligent drafting scheme according to the embodiment of the present specification proposes an algorithm framework capable of efficiently solving the intelligent drafting problem; solving the problem of corpus reuse by using synonymous replacement; and using the method of inserting sentences after paragraphs to enrich the content of the article. Compared with the method of writing manuscripts that is too automated, this solution adds manual intervention, which greatly improves the practicability and accuracy; compared with the completely template-based method, it greatly enhances flexibility and scalability.

Those of ordinary skill in the art should be further aware that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the hardware Interchangeability with software. In the above description, the composition and steps of each example have been described generally in terms of functions. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present application.

The steps of the method or algorithm described in connection with the embodiments disclosed herein may be implemented by hardware, a software module executed by a processor, or a combination of the two. Software modules can be placed in random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or in technical fields Any other form of storage medium is known.

The specific embodiments described above further describe the objectives, technical solutions, and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. The scope of protection, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention.

Claims (28)

A method for training a writing model. The writing model includes a word segmentation unit, a transformation unit, and a recurrent neural network. The method includes: Obtaining a sample for training the writing model, the sample including a text and a calibration value for the text, the calibration value indicating whether the text is smooth; and Using the sample to train the writing model such that after the text is input to the writing model, compared to before training, the difference between the output value of the writing model after the training and the calibration value is The absolute value decreases, wherein after the text is input to the writing model, the word segmentation unit obtains a plurality of word segmentations arranged in sequence from the text, wherein the plurality of word segmentations are based on their The front and back positions are arranged sequentially, the conversion unit converts the sequentially arranged plurality of word segments into correspondingly arranged multiple word vectors, and the recurrent neural network outputs the writing model based on the sequentially arranged multiple word vectors Output value. The method of training a writing model according to claim 1, wherein the text comprises two non-adjacent texts in an existing article, and the calibration value is 0. The method of training a writing model according to claim 1, wherein the text includes two texts respectively belonging to two existing articles, and the calibration value is 0. The method of training a writing model according to claim 1, wherein the text order comprises two texts arranged next to each other in an existing article, and the calibration value is 1. The method for training a writing model according to any one of claims 2-4, wherein the existing article is an existing article belonging to a selected field, and the writing model is used for writing after the training Articles belonging to the selected area. The method of training a writing model according to claim 1, wherein training the writing model using the samples comprises using the samples to train the writing model by a back propagation algorithm. The method of training a writing model according to claim 1, wherein the recurrent neural network comprises one of the following networks: RNN, LSTM, and GRU. A method of writing, including: Obtaining n first candidate texts of a first to-be-written text of an article to be written and m second candidate texts of a second to-be-written text, where n, m are a predetermined number, and wherein, in the to-be-written article, The first text to be written is located in front of the second text to be written and is adjacent to the second text to be written; Combining each text in the first candidate text with each text in the second candidate text to obtain n × m input texts, wherein, in each of the input texts, the The first candidate text is located in front of the second candidate text; Input the n × m input texts respectively into a writing model trained by the method according to any one of claims 1-7 to obtain n × m models corresponding to the n × m input texts respectively An output value that predicts the degree of fluency of the corresponding input text; and Determine the selected text of the first to-be-written text from the n first candidate texts based on the n × m model output values, and determine the second to-be-written text from the m second candidate texts Write selected text for text. A method of writing, including: Obtaining q candidate texts of to-be-written text in a to-be-completed article, where q is a predetermined number; Obtaining an existing text of a predetermined length that is adjacent to the text to be written and is located in front of the text to be written from the to-be-completed article; Combining the existing text and the q candidate texts separately to obtain q input texts, wherein in each of the input texts, the existing texts are located in front of the candidate texts; Input the q input texts respectively into a writing model trained by the method according to any one of claims 1 to 7 to obtain q model output values corresponding to the q input texts respectively, the model How well the output value predicts the corresponding input text; and Based on the q model output values, a selected text of the text to be written is determined from the q candidate texts. The writing method according to claim 8 or 9, further comprising, after the selected text is determined, synonymous replacement of words and / or sentences in the selected text. The writing method according to claim 9, wherein the obtaining the q candidate texts of the text to be written in the article to be completed comprises acquiring at least one keyword of the text to be written; and, based on the keywords, Acquiring q candidate texts of the text to be written. The writing method according to claim 11, further comprising, before acquiring at least one keyword of the text to be written, acquiring a keyword of a theme of the article to be completed. The writing method according to claim 11 or 12, wherein obtaining the q candidate texts of the text to be written based on the keywords comprises, according to a predetermined search and sorting method, corresponding to the text to be written The corpus is searched based on the keywords to obtain the q candidate texts. The writing method according to claim 13, wherein the corpus corresponding to the text to be written comprises at least one of the following: a corpus corresponding to a field of the text to be written, and a form corresponding to the form of the text to be written Corpus. A device for training a writing model. The writing model includes a word segmentation unit, a conversion unit, and a recurrent neural network. The device includes: An obtaining unit configured to obtain a sample for training the writing model, the sample including a text and a calibration value for the text, the calibration value indicating whether the text is smooth; and A training unit configured to use the sample to train the writing model such that, after the text is input to the writing model, compared to before training, the output value of the writing model after The absolute value of the difference between the calibration values is reduced, wherein after the text is input to the writing model, the word segmentation unit obtains a plurality of sequenced word segments arranged in sequence from the text, wherein the plurality of word segmentations are based on Are arranged sequentially in front and back positions in the text, the conversion unit converts the sequentially arranged plurality of word segmentation into correspondingly arranged multiple word vectors, and the recurrent neural network is based on the sequentially arranged multiple words The vector outputs the output value of the writing model. The apparatus for training a writing model according to claim 15, wherein the text includes two non-adjacent texts in an existing article, and the calibration value is 0. The apparatus for training a writing model according to claim 15, wherein the text includes two texts respectively belonging to two existing articles, and the calibration value is 0. The apparatus for training a writing model according to claim 15, wherein the text order comprises two texts arranged next to each other in an existing article, and the calibration value is 1. The apparatus for training a writing model according to any one of claims 16 to 18, wherein the existing article is an existing article belonging to a selected field, and the writing model is used for writing after the training Articles belonging to the selected area. The apparatus for training a writing model according to claim 15, wherein the training unit is further configured to use the samples to train the writing model by a back propagation algorithm. The apparatus for training a writing model according to claim 15, wherein the recurrent neural network comprises one of the following networks: RNN, LSTM, and GRU. A writing device including: The obtaining unit is configured to obtain n first candidate texts of the first to-be-written text of the article to be written and m second candidate texts of the second to-be-written text, where n and m are a predetermined number, and In the article to be written, the first to-be-written text is located in front of the second to-be-written text and is adjacent to the second to-be-written text; A combining unit configured to combine each text in the first candidate text with each text in the second candidate text to obtain n × m input texts, where In the input text, the first candidate text is located in front of the second candidate text; An input unit configured to respectively input the n × m input texts into a writing model trained by the method according to any one of claims 1 to 7, so as to obtain respective correspondences to the n × m input texts N × m model output values that predict the degree of smoothness of the corresponding input text; and A determining unit configured to determine a selected text of the first to-be-written text from the n first candidate texts based on the n × m model output values, and from the m second candidate texts A selected text of the second to-be-written text is determined. A writing device including: A first obtaining unit configured to obtain q candidate texts of a text to be written in a to-be-completed article, where q is a predetermined number; A second obtaining unit configured to obtain, from the to-be-completed article, an existing text of a predetermined length that is adjacent to the to-be-written text and located in front of the to-be-written text; A combining unit configured to combine the existing text and the q candidate texts separately to obtain q input texts, wherein in each of the input texts, the existing texts are located in the candidate texts In front of An input unit configured to respectively input the q input texts into a writing model trained by the method according to any one of claims 1 to 7 to obtain q models corresponding to the q input texts respectively An output value that predicts the degree of fluency of the corresponding input text; and The determining unit is configured to determine the selected text of the text to be written from the q candidate texts based on the q model output values. The writing device according to claim 22 or 23, further comprising a replacement unit configured to perform synonymous replacement of words and / or sentences in the selected text after determining the selected text. The writing device according to claim 23, wherein the first acquisition unit comprises a first acquisition subunit configured to acquire at least one keyword of the text to be written; and a second acquisition subunit based on The keywords obtain q candidate texts of the text to be written. The writing device according to claim 25, further comprising a third acquisition unit configured to acquire keywords of a subject of the article to be completed before acquiring at least one keyword of the text to be written. The writing device according to claim 25 or 26, wherein the second acquisition subunit is further configured to perform a search based on the keywords in a corpus corresponding to the text to be written according to a predetermined search and sorting method To obtain the q candidate texts. The writing device according to claim 27, wherein the corpus corresponding to the text to be written comprises at least one of the following: a corpus corresponding to a field of the text to be written, and a form corresponding to the form of the text to be written Corpus.

Images