CN110688834A - Method and equipment for rewriting intelligent manuscript style based on deep learning model - Google Patents
Method and equipment for rewriting intelligent manuscript style based on deep learning model Download PDFInfo
- Publication number
- CN110688834A CN110688834A CN201910780331.4A CN201910780331A CN110688834A CN 110688834 A CN110688834 A CN 110688834A CN 201910780331 A CN201910780331 A CN 201910780331A CN 110688834 A CN110688834 A CN 110688834A
- Authority
- CN
- China
- Prior art keywords
- style
- target
- deep learning
- learning model
- source
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013136 deep learning model Methods 0.000 title claims abstract description 89
- 238000000034 method Methods 0.000 title claims abstract description 60
- 239000013598 vector Substances 0.000 claims abstract description 71
- 230000011218 segmentation Effects 0.000 claims description 37
- 238000012549 training Methods 0.000 claims description 30
- 239000000463 material Substances 0.000 claims description 19
- 230000015654 memory Effects 0.000 claims description 5
- 238000010801 machine learning Methods 0.000 claims description 4
- 230000009193 crawling Effects 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 16
- 238000004422 calculation algorithm Methods 0.000 description 12
- 238000013528 artificial neural network Methods 0.000 description 11
- 239000011159 matrix material Substances 0.000 description 10
- 210000004027 cell Anatomy 0.000 description 8
- 230000008569 process Effects 0.000 description 6
- 230000000306 recurrent effect Effects 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 4
- 238000007792 addition Methods 0.000 description 3
- 235000019640 taste Nutrition 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000012217 deletion Methods 0.000 description 2
- 230000037430 deletion Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 235000019580 granularity Nutrition 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000006403 short-term memory Effects 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000001131 transforming effect Effects 0.000 description 2
- 241000288105 Grus Species 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000002360 explosive Substances 0.000 description 1
- 230000014509 gene expression Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Machine Translation (AREA)
Abstract
An exemplary aspect of the present disclosure relates to a method for intelligent document style rewriting based on a deep learning model, including receiving a source document and at least one target style associated with a source style; for each of the one or more natural sentences of the source document: generating, by a deep learning model, a semantic vector corresponding to the natural sentence of the source document based on the source style; and generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and sequentially merging one or more target natural sentences corresponding to the one or more natural sentences of the source document to generate at least one target document associated with the at least one target style. The disclosure also relates to a corresponding device and the like.
Description
Technical Field
The present application relates to artificial intelligence, and more particularly, to intelligent manuscript rewriting based on deep learning.
Background
With the explosive development of the media in the information age at present, public opinion propaganda means of various media are increasingly abundant, and reader preferences and tastes are increasingly diversified, so that great challenges are brought to writing propaganda and releasing work. The same article or a news manuscript is often modified into articles or news manuscripts with different styles according to the characteristics of various media and various readers, so that the method is suitable for readers with specific categories, levels, tastes and the like to improve the actual reading quantity, the interest degree of the readers and the influence of the articles or the news manuscripts. Thus, personalized adaptation of articles has a wide real need.
However, the categories, levels, tastes, etc. of the readers are quite different. Essentially the same article needs to be rewritten many times in order to accommodate a specific reader. This greatly increases the work load of writing.
Therefore, there is a need in the art for a technique that enables automated intelligent rewriting of a document.
Disclosure of Invention
An exemplary aspect of the present disclosure relates to a method for intelligent document style rewriting based on a deep learning model, including receiving a source document and at least one target style associated with a source style; for each of the one or more natural sentences of the source document: generating, by a deep learning model, a semantic vector corresponding to the natural sentence of the source document based on the source style; and generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and sequentially merging one or more target natural sentences corresponding to the one or more natural sentences of the source document to generate at least one target document associated with the at least one target style.
According to an exemplary embodiment, the deep learning model includes an encoder and a decoder, wherein semantic vectors corresponding to natural sentences of the source document are generated by the encoder of the deep learning model based on the source style, and target natural sentences corresponding to the semantic vectors are generated by the decoder of the deep learning model based on the at least one target style.
According to a further exemplary embodiment, the method further comprises tokenizing natural sentences of the source document, and wherein the encoder of the deep learning model comprises a plurality of cascaded first unit modules, wherein each word of the tokenized natural sentences is input to the plurality of cascaded first unit modules in sequence, respectively.
According to another exemplary embodiment, the method further includes generating, by the plurality of cascaded first unit modules, a present-level output based on an output of a previous-level first unit and a word input to the present-level in the participled natural sentence, wherein the first-level first unit is an output of the previous level in the source style, and a last-level first unit outputs a semantic vector corresponding to the natural sentence of the source document.
According to a further exemplary embodiment, the decoder of the deep learning model comprises a plurality of cascaded second unit modules, the method further comprises generating, by the plurality of cascaded second unit modules, target words corresponding to the semantic vectors, respectively, based on the at least one target style; and combining the target words generated by the plurality of cascaded second unit modules to form the target natural sentence.
According to an exemplary embodiment, the method further includes filling the input of the redundant first unit module with a blank when the number of words obtained after the natural sentence of the source document is participled is less than the number of the plurality of cascaded first unit modules.
According to an exemplary embodiment, the method further includes segmenting the natural sentence of the source document when the number of words obtained after the natural sentence is segmented is larger than the number of the plurality of cascaded first unit modules.
According to an exemplary embodiment, the source style is received externally or extracted directly from the source document.
According to an exemplary embodiment, the method further comprises training the deep learning model, wherein training the deep learning model comprises setting a feature library, the feature library comprising two or more features related to smart document style overwriting; generating a document material library comprising pairs of articles associated with at least two features in the feature library; and training the deep learning model based on the library of manuscript materials.
According to a further exemplary embodiment, generating the manuscript material library comprises one or more of the following or any combination thereof: for a particular feature in the feature library: capturing all articles with the specific characteristics from a characteristic website; retrieving articles with high relevance from a search engine based on the specific characteristics; and learning out the marking model by utilizing machine learning so as to find out articles related to the specific characteristics in texts crawled from the network.
Other aspects of the disclosure also relate to corresponding apparatuses and computer-readable media.
Drawings
Fig. 1 shows a diagram of an example Recurrent Neural Network (RNN).
FIG. 2 shows a diagram of an example Long Short Term Memory (LSTM) network.
FIG. 3 illustrates a diagram of a word segmentation system according to an exemplary aspect of the present disclosure.
FIG. 4 illustrates a block diagram of a deep learning model according to an exemplary aspect of the present disclosure.
FIG. 5 illustrates a flow chart of a method of offline training of a deep learning model according to an exemplary aspect of the present disclosure.
Fig. 6 illustrates a block diagram of a method of intelligent document rewrite using a trained deep learning model in accordance with an aspect of the present disclosure.
FIG. 7 illustrates a diagram of a scenario for intelligent document rewrite using a trained deep learning model, according to an exemplary aspect of the present disclosure.
Fig. 8 illustrates a block diagram of an intelligent document-rewriting device based on deep learning according to an exemplary aspect of the present disclosure.
Fig. 9 illustrates a block diagram of an apparatus for intelligent document rewrite using a trained deep learning model in accordance with an aspect of the disclosure.
Detailed Description
Conventional neural networks typically include an input layer, one or more hidden layers, and an output layer, where layers may be fully connected, but nodes between each layer are connectionless. This common neural network is powerless for many problems. For example, in speech recognition or natural language processing, it is generally necessary to base the prediction on the next word of a sentence, if it is desired to predict what the next word is. This is because the words before and after a sentence are not independent of each other.
In this case, a Recurrent Neural Network (RNN) arises. To process sequence data, the RNN correlates the current output of a sequence with the previous output. Thus, the network remembers the previous information and applies it to the calculation of the current output. In other words, the nodes between hidden layers are no longer connected but connected, and the input of a hidden layer includes not only the output of the input layer/previous hidden layer but also the output of the current hidden layer at the previous time.
Recurrent Neural Networks (RNNs) are most commonly used for time series data mining. Recurrent neural networks are a class of neural networks with memory and are therefore often used for mining of data with temporal correlation. In general, a neuron in a recurrent neural network can receive not only information from other neurons but also information of itself, thereby forming a cascaded network structure having loops.
Fig. 1 shows a diagram of an example simple Recurrent Neural Network (RNN) 100. As can be seen, the left part represents a simplified RNN structure, with the lowermost circle representing the input layer, the middle circle representing the hidden layer, and the uppermost circle representing the output layer. X in the input layer represents an input layer value vector. S in the hidden layer represents a hidden layer value vector. O in the output layer represents the output layer value vector. U, V and W each represent a weight matrix connecting the respective layers.
The functional relationship between the hidden layer value vector S and the input layer value vector X is shown in the following formula (1), for example:
St＝f(U·Xt+W·St-1+b) (1)
as can be seen, the hidden layer value vector S at time ttNot only dependent on the current input layer value vector XtAnd a weight matrix U connecting the input layer and the hidden layer, the hidden layer value vector S also depending on the previous time t-1t-1And the corresponding weight matrix W (and possibly also the bias term b).
On the other hand, the output layer value vector O at time ttAnd a hidden layer value vector StFor example, the functional relationship of (2) is shown as follows:
Ot＝g(V·St) (2)
as can be seen, the output tier value vector O at time ttDependent on the current hidden layer value vector StAnd a weight matrix V connecting the hidden layer and the output layer.
By spreading the RNN 100 in time, the structure of the right part of fig. 1 is obtained. As can be seen, the RNN adds a delay back to the hidden layer from the hidden layer compared to a fully connected structure such as a multi-layer perceptron (MLP), thereby giving the RNN a "memory" function. And is thus particularly suitable for processing time series data.
In theory, RNNs can process sequence data of any length. That is, the length of the input sequence may not be fixed. In practice, however, it is often assumed that the current state is only associated with a limited number of previous states for complexity reduction. RNNs and their various variants (e.g., bi-directional RNNs, LSTM, GRUs, etc.) have enjoyed great success and widespread use in numerous natural language processes.
RNNs can be classified into a variety of structures according to the number of input and output sequences, including: one-to-one structures, one-to-many structures, many-to-one structures, many-to-many structures, and the like. The many-to-many structure may also include structures where the input and output sequences are of equal length and structures where the input and output sequences are of unequal length. The many-to-many structure with unequal input and output sequences is called seq2seq (sequence-to-sequence) model. A common seq2seq model may include an encoder-decoder architecture, i.e., utilizing two RNNs, one RNN as an encoder and the other RNN as a decoder. According to one implementation, one word in the sequence may be input into the encoder at each time instant. The encoder is responsible for compressing the input sequence into a vector of specified length (i.e., embedding), which can be regarded as the semantics of the sequence, a process known as encoding. The decoder is responsible for generating the specified sequence from the semantic vector, a process known as decoding.
One of the ways that the encoder obtains the semantic vector may include directly taking the hidden state of the last input as the semantic vector C. Another way to obtain the semantic vector may include transforming the last hidden state to obtain the semantic vector C. The manner of obtaining the semantic vector may further include transforming all hidden states of the input sequence to obtain the semantic vector C. The decoder may input the semantic variables obtained by the encoder as initial states into the RNN as a decoder to obtain an output sequence. According to one exemplary implementation, the semantic vector C may participate in the operation only as an initial state. According to another exemplary implementation, the semantic vector C may participate in the operation of the sequence in the decoder at all time instants.
LSTM (long short term memory) networks are a special class of RNNs that are designed to address the above-mentioned issues with RNNs, among others. In general, an RNN may comprise a chain structure of repeating neural network element modules. In standard RNNs, this repeated unit block typically comprises only simple structures, such as a tanh layer. LSTM is also a chain structure, but the repeated unit modules have a more complex structure. Specifically, a general LSTM unit module may include an input gate, a forgetting gate, an output gate, and a cell, which interact in a special way, wherein the forgetting gate may be used to forget information that is not needed, the input gate and the output gate may be used for input and output of parameters, and the cell is used for storing states. LSTM also has various variations. The attention (attention) mechanism introduced in recent years can greatly improve the efficiency of the LSTM, thereby bringing a wider prospect for the LSTM.
Fig. 2 shows a diagram of an example LSTM network 200. LSTM network 200 may include a cascade of multiple cells. Each cell can obtain the output of the present stage by performing the embedding based on the output of the previous stage and the input of the present stage. Specifically, as can be seen, first, the current input may be cascaded with the previous h, noting the result as x. Then, x may be subjected to matrix point multiplication with the weight matrices w (f), w (i), w (j), w (o) to obtain result matrices f, i, j, o, respectively, where w (f), w (i), w (j), w (o) are the core weight parameters of the LSTM cells, and the purpose of training is to train the four weight matrix parameters. Then, sigmod operation is performed on matrix f, i.e. sigmod (f), sigmod operation is performed on matrix i, i.e. sigmod (i), tanh operation is performed on matrix j, i.e. tanh (j), sigmod operation is performed on matrix o, i.e. sigmod (o), and new c is calculated, wherein new c is old c sigmod (f) + (sigmod (i)) tanh (j). Finally, a new h is calculated, wherein new h is tanh (new c) sigmod (o). After the above operation, the cell is calculated to obtain a new c and a new h. Then h is taken as the output of the current time, and the binary group (c, h) is taken as the cell state of the current time to be stored for the next cell calculation.
One of the basic steps in the processing of natural language involves word segmentation. In the context of word-based languages such as English, spaces are used as natural delimiters between words. In scenarios such as chinese, only words, sentences and paragraphs can be simply delimited by distinct delimiters, without formally explicit delimiters between words. Therefore, when natural language processing of a Chinese is performed, word segmentation is generally required first. Existing word segmentation algorithms can be divided into three major categories: a word segmentation method based on character string matching, a word segmentation method based on understanding and a word segmentation method based on statistics. Whether the method is combined with the part-of-speech tagging process or not can be divided into a simple word segmentation method and an integrated method combining word segmentation and tagging.
FIG. 3 illustrates a word segmentation system 300 according to an exemplary aspect of the present disclosure. As can be seen, the word segmentation system 300 includes a word segmenter 301. When a sentence is input to the segmenter 301, the segmenter 301 outputs the segmented result. The tokenizer 301 may be implemented using any state-of-the-art or future-art tokenization algorithm.
For example, in a chinese word segmentation scenario, word segmentation methods based on string matching may be used, such as forward maximum matching, reverse maximum matching, least segmentation, and so on; an understanding-based word segmentation method; a statistical-based word segmentation method; rule-based word segmentation methods such as a minimum matching algorithm, a maximum matching algorithm, a word-by-word matching algorithm, a neural network word segmentation algorithm, an association-backtracking method, an N-shortest path word segmentation algorithm, and the like; an algorithm based on word frequency statistics, a word segmentation method based on expectation, a finite multi-level enumeration method and the like.
As can be seen, although one possible segmentation result is shown above, different segmentation results may be obtained in case different segmentation algorithms are used.
According to an exemplary embodiment, the source document may be input into the tokenizer sentence by sentence, and the tokenizer outputs the tokenized source sentence (or other unit) sentence by sentence or the tokenized source document in its entirety. According to another exemplary embodiment, the source document may be input integrally into the tokenizer, and the tokenizer may output the tokenized source document (or other element) sentence-by-sentence or integrally.
FIG. 4 illustrates a block diagram of a deep learning model 400 according to an exemplary aspect of the present disclosure. The deep learning model 400 may comprise, for example, a seq2seq model. According to an exemplary but non-limiting embodiment, the seq2seq model can comprise, for example, an encoder portion 410 and a decoder portion 420, wherein the encoder portion 410 can comprise a unit module h1，h2，……hn. Unit module h1，h2，……hnMay be implemented, for example, using the unit modules (spread out over time lines) discussed above in connection with RNNs (e.g., LSTM) and form the chain structure discussed above in connection with RNNs (e.g., LSTM) spread out over time lines. h is0May be a feature of the source script (or source style). Features may refer to categories, logos, genres, applicable demographics, etc. of paragraphs, examples of which may include, for example, the general body, the palace body, northeast, youth phrases, etc. … …. This will be further described below.
x1，x2，……xnIs a pair unit moduleh1，h2，……hnIs input. x is the number of1，x2，……xnMay include a sequence of words or phrases that may be obtained by tokenizing natural sentences in the source document. For example, the tokenization of natural sentences may be implemented using the tokenization system described above in connection with FIG. 2.
According to an exemplary but non-limiting embodiment, n can be a fixed window length. When the length of the natural sentence after word segmentation is less than n (for example, only x1，x2，……xiWherein i<n), the remaining portion may be filled with a blank (e.g., fill x)i+1，……xn). On the other hand, when the length of the natural sentence after word segmentation is larger than n, the sentence can be segmented. The segmentation criteria may include, for example, commas, discourse words, and/or other distinct segmented words, or the like, or any combination thereof.
According to an exemplary but nonlimiting embodiment, the unit module h1，h2，……hnWord/word sequence x based on this level input1，x2，……xnAnd the output of the preceding unit module is embedded to obtain the output vector h of the current stage1，h2，……hnWherein h is1，h2，……hn-1Are respectively input to the next unit module in the chain structure as its input.
Last unit module h which can be based on an encodernOutput of (i.e., hidden state) hnTo obtain a semantic vector C. For example, according to an exemplary but nonlimiting embodiment, h can be directly couplednAs semantic vector C. According to another exemplary but nonlimiting embodiment, h may be pairednA transformation is performed to obtain a semantic vector C. According to yet another exemplary but nonlimiting embodiment, the input sequence x may be input1，x2，……xnAll hidden states h of1，h2，……hnAggregated (and optionally transformed) to obtain a semantic vector C.
The decoder portion 420 may include a unit module h1’，h2’，……hm', where m can be greater than, equal to, or less than n. Unit module h1’，h2’，……hm' may be implemented, for example, using the unit modules discussed above in connection with RNNs (e.g., LSTMs), and form the chain structures discussed above in connection with RNNs (e.g., LSTMs).
According to an exemplary but nonlimiting embodiment, h0Is inputted to the unit module h1’，h0' may be a feature of the target manuscript (or target style), which may be different from the source manuscript feature. Examples of target script features may include, for example, a body in general, a body in palace, a word northeast, a word of youth, and so on … ….
According to an exemplary but non-limiting embodiment, the unit module h1’，h2’，……hm-1' the output vector h of the present stage can be obtained by embedding based on the semantic vector C and the output of the preceding stage unit module1’，h2’，……hm-1To the next unit module h2’，……hm'. In the example of fig. 4, a semantic vector C is input to each unit module h of the decoder 4201’，h2’，……hm-1'. In an alternative embodiment, the semantic vector C may also be input only to, for example, the first unit module of the decoder 420.
According to an exemplary but non-limiting embodiment, the unit module h1’，h2’，……hm' separately outputting y by embedding the output of the previous stage based on the semantic vector1，y2，……ym。y1，y2，……ymAnd may also include sequences of words or phrases. Combination y1，y2，……ymThe obtained sequence can be the natural sentence input sequence x with the source style1，x2，……xnAnd outputting the target natural sentence obtained after rewriting the characteristics of the target manuscript.
According to an exemplary but non-limiting embodiment, similarly, when the output sequence length is less than m (e.g., only y)1，y2，……yjWherein j is<m), the remaining portion of the output (e.g., y) may be usedj+1，……ym) May be filled with blanks. On the other hand, the sentence segmented at the encoder may be re-composed after being output at the decoder. The decoded sentences are combined in sequence to form the target document.
The various unit modules in the encoder 410 and decoder 420 may be implemented with various RNN structures, including, but not limited to, the network structures described above in connection with fig. 1 and 2, for example.
Although the embodiments are described in connection with outputting a target document corresponding to one target style, as can be appreciated, the present disclosure may encompass implementations in which multiple target documents corresponding to multiple target styles, respectively, are output simultaneously or sequentially.
By converting the content of the source style (e.g., sentence, paragraph, article) into the corresponding content of the target style (e.g., sentence, paragraph, article) through the encoder and decoder as described with reference to the exemplary embodiment of fig. 4, the rewrite flow of the deep learning assist can be realized, so that the work efficiency of manuscript rewrite can be greatly improved.
The training of the deep learning model may include, for example, offline training and online training. Fig. 5 illustrates a flow diagram of a method 500 of offline training of a deep-learning model (e.g., model 400 of fig. 4) according to an exemplary aspect of the present disclosure.
The method 500 may include setting up a feature library at block 502. As previously mentioned, features may refer to classification, identification, genre, applicable demographics, etc. of paragraphs, examples of which may include, for example, the common body, the palace body, northeast, youth words, etc. … …. Features may be represented by phrases or identifiers, etc. For example, according to an example, the natural sentence "this is good" may be characterized as a plain body. The corresponding palace body natural sentence is 'this is supposed to be excellent'. Setting the library of features may include, for example, setting different features according to different preferences for a population. The number of different article types that the model can adapt depends on the number of different features.
At block 504, the method 500 generates a library of document materials. The library of textual materials is associated with a particular feature. The document materials library of a specific feature may refer to a collection of documents having the specific feature. Generating a corpus of manuscript material associated with a particular feature may include, for example, grabbing all articles with the particular feature from a featured website (e.g., microblog, headline, banner, etc.). Generating a corpus of manuscript material associated with a particular feature may also include retrieving on a search engine using the feature to obtain articles with top-ranked relevance. Generating the manuscript material library associated with the specific feature may further include learning the marking model by using a machine learning method after obtaining certain article data through the foregoing various manners or any combination thereof, and then searching for more articles related to the feature in a text crawled on the web. According to an exemplary but non-limiting embodiment, learning the marking model can include learning a text classifier, which is then trained using the articles captured and/or acquired as described above as a training set. After training is finished, the trained text classifier can be used for searching for more articles related to corresponding features in the crawled text. After generating two or more document feature libraries associated with at least two or more of the feature types in the feature library, the method 500 may proceed to block 506.
At block 506, the method 500 trains a deep learning model, such as a seq2seq model or the like. According to an exemplary and non-limiting embodiment, training a deep learning model may include generating a training data set by finding pairs of two articles in each corpus of textual material that have the same or similar subject matter but different characteristics (e.g., classification, identification, genre, applicable population, etc. of paragraphs). The same or similar subject matter may include one or more of the following: contain the same keywords, and/or have sufficiently high similarity for each corresponding sentence in the article. The similarity may be calculated using various algorithms, including, but not limited to, for example, word vector-based similarity (e.g., cosine similarity, manhattan distance similarity, euclidean distance similarity, minuscule distance similarity, etc.), character-based similarity (e.g., edit distance similarity, simhash, number of common characters, etc.), probability statistics-based similarity (e.g., jackard similarity coefficient, etc.), word embedding model-based similarity (e.g., word2vec/doc2vec, etc.).
According to another exemplary but non-limiting embodiment, the training data set may also be obtained from other sources. For example, the training data set may be obtained directly from an external source. After the training data set is generated or obtained, a deep learning model, such as deep learning model 400 described with reference to fig. 4, may be trained using one article of each of two article pairs in the training data set as a source document and the other article as a target document. The above exemplary embodiments are trained on the granularity of articles, but the disclosure is not so limited. According to another exemplary but non-limiting embodiment, training may also be performed on sentences or other granularities.
Fig. 6 illustrates a block diagram of a method 600 of intelligent document rewrite using a trained deep learning model in accordance with an aspect of the present disclosure. The method 600 includes inputting a source document, a source style, and a target style, at block 602. The source style may represent a characteristic of the source document, including, for example, a category or style to which it belongs. The target style may represent a characteristic of the document to be rewritten, including, for example, a category or style to which it belongs. According to an exemplary but non-limiting embodiment, the source style may also be determined directly on the basis of the source document without further input.
At block 604, the method 600 includes outputting the target document with the trained deep learning model. For example, for the deep learning model described with reference to FIG. 4, the source document may be tokenized to x in natural sentences1，x2，……xn. The word segmentation may be implemented using, for example, the word segmentation system 300 described in conjunction with fig. 3. X after word segmentation1，x2，……xnUnit modules h that can be respectively inputted to the encoder 4101，h2，……hn. According to an exemplary and non-limiting embodiment, the length of a natural sentence after word segmentation is less than n (e.g., only x)1，x2，……xiWherein i<n), the remaining portion may be filled with a blank (e.g., fill x)i+1，……xn). On the other hand, when the length of the natural sentence after word segmentation is larger than n, the sentence can be segmented. The segmentation criteria may include, for example, commas, discourse words, and/or other distinct segmented words, or the like, or any combination thereof. Source style is input to h0. For example, sentence 1 in the source document is tokenized to x1，x2，……xn. Source style h0And the first word x1Is inputted to the first unit module h1. The first unit module is used for aligning a first word x1Embedding and outputting a vector h1. Vector h1With the second word x2Are inputted together to the second unit module h2. The second unit module h2For the second word x2Embedding and outputting a vector h2. Vector h2With a third word x3Are inputted together to the third unit module h3And so on. Vector hn-1With the nth word xnAre inputted together to the nth unit module hn. The nth unit module is used for the nth word xnEmbedding and outputting a vector hn. Through the unit modules h1，h2，……hnCascade between them, the word x of the sentence is established1，x2，……xnThe context relationship between them. As can be seen, although h is used for both the unit block and the vector of its outputnFormal expressions, but this is merely for the convenience of identifying their relatedness and their respective meanings are clear and should not be confused.
For the deep learning model described with reference to, for example, FIG. 4, it may be based on the last unit module h of the encodernOutput of (i.e., hidden state) hnTo obtain a semantic vector C. For example, according to an exemplary but nonlimiting embodiment, h can be directly couplednAs semantic vector C. According to another exemplary but nonlimiting embodiment, h may be pairednA transformation is performed to obtain a semantic vector C. According to yet another exemplary but nonlimiting embodiment, the input sequence x may be input1，x2，……xnAll hidden states h of1，h2，……hnPolymerized (and optionally modified)And) obtaining a semantic vector C. The semantic vector C may be input to each unit module h of the decoder portion 1201’，h2’，……hm', where m can be greater than, equal to, or less than n. h is0Is inputted to the unit module h1’，h0' may be a feature of the target manuscript (or target style), which may be different from the source manuscript feature. Unit module h1’，h2’，……hm-1' respective embedding of semantic vector C and output of vector h1’，h2’，……hm-1To the next unit module h2’，……hm'. Unit module h1’，h2’，……hm' also output y separately1，y2，……ym。y1，y2，……ymMay include a sequence of words or phrases. y is1，y2，……ymThe sequence of (2) can be a natural sentence input sequence x of the source style1，x2，……xnAnd outputting a sentence output sequence obtained by rewriting the target manuscript characteristics. According to an exemplary but non-limiting embodiment, similarly, when the output sequence length is less than m (e.g., only y)1，y2，……yjWherein j is<m), the remaining portion of the output (e.g., y) may be usedj+1，……ym) May be filled with blanks. On the other hand, the sentence segmented at the encoder may be re-composed after being output at the decoder.
At optional block 606, method 600 may include optionally making a human-assisted overwrite adjustment. The article manuscript automatically rewritten by the deep learning model may need to be checked and edited manually if the article manuscript meets the available standard. However, compared with pure manual rewriting, automatic rewriting of the deep learning model can greatly reduce the manual intervention degree and workload. For example, according to an exemplary but non-limiting embodiment, an editing interface can be provided to facilitate the addition, deletion, and/or modification of sentences in an edit box by a user. According to another exemplary but non-limiting embodiment, additions, deletions, and/or substitutions to paragraphs and/or sentences, etc. may also be supported. For example, the sentence may be manually input by the user at the time of addition, or may be automatically retrieved from the document material library. As another example, at the time of replacement, the user may manually input, or a sentence automatically generated by the deep learning model may be provided for selection.
At block 608, the method 600 may include outputting the result. Outputting the result may include, for example, outputting a rewritten sentence/article. After rewriting is completed, corresponding saving, previewing, auditing and/or publishing and the like can be performed.
FIG. 7 illustrates a diagram of a scenario 700 for intelligent document rewrite using a trained deep learning model, according to an exemplary aspect of the present disclosure.
As can be seen, the source document 701 is input into a participle module 702. The segmentation module 702 may, for example, include the segmentation system 300 described above in connection with fig. 3, and/or the like. The segmentation module 702 may output a segmented manuscript 703 based on the employed segmentation algorithm.
A tokenized script (e.g., a tokenized natural sentence or other unit) 703 is input into a trained deep learning model 704. The segmented manuscript 703 may be input into the deep learning model 704 sentence by sentence, or may be input into the deep learning model 704 by other units. For example, the tokenized script 703 may also be input into the deep learning model 704 in its entirety, or may be streamed into the deep learning model 704 while being generated.
A source style corresponding to the source document 701 and a target style corresponding to the target document may be input into the deep learning model 704. According to an alternative embodiment, the source style may also be automatically identified by the system from the source document. In this case, the system may also include, for example, a feature recognition module (not shown). For example, the source style may be a regular body and the target style may be a young language.
The deep learning model 704 may include, for example, the deep learning model 400 described above in connection with fig. 4, or variations thereof, and so forth. Based on the source style and the target style, the deep learning model 704 may rewrite the segmented source document 703 accordingly, for example, from a normal body to a body of a young person.
The deep learning model 704 outputs a target document 705 rewritten based on the target style. According to an exemplary embodiment, the target manuscript 705 may be output sentence by sentence, for example, a target sentence may be output every sentence of the source manuscript input, or a target sentence may be output sentence by sentence for the source manuscript input as a whole. According to another exemplary embodiment, the target manuscript 705 may be output in its entirety. For example, for a source document that is input in its entirety, all target sentences may be combined and output after they are obtained. For another example, for a source document input sentence by sentence, all target sentences may be obtained and then combined to output a target document. Combining the target sentences may include sequentially merging the target sentences to obtain the target document. The target document 705 may be output directly to the outside or may be output after being provided by the deep learning model 704 with further human-assisted overwrite adjustment.
FIG. 8 illustrates a block diagram of an offline training apparatus 800 for deep learning models, according to an exemplary aspect of the present disclosure. The deep learning model may include, for example, the deep learning model 400 described in conjunction with fig. 4, and the like.
According to an exemplary but non-limiting embodiment, the offline training device 800 of the deep learning model can include a module 802 for setting a feature library. The module for setting a feature library 802 may, for example, implement the functionality described above in connection with block 502 of fig. 5, and/or the like.
According to an exemplary but non-limiting embodiment, the offline training apparatus 800 of the deep learning model can further include a module 804 for generating a corpus of manuscript materials. The module 804 for generating a library of manuscript material may, for example, implement the functionality described above in connection with block 504 of fig. 5, and/or the like.
According to an exemplary but non-limiting embodiment, the offline training apparatus 800 for deep learning models can further include a module 806 for training the deep learning models. The module for training the deep learning model 806 may, for example, implement the functionality described above in connection with block 506 of fig. 5, and/or the like.
Fig. 9 illustrates a block diagram of an apparatus 900 for intelligent document rewrite using a trained deep learning model in accordance with an aspect of the disclosure.
According to an exemplary but non-limiting embodiment, intelligent document-rewriting device 900 may include a module 902 for inputting a document, a source style, and a target style. Module for inputting a manuscript, a source style, and a target style 902 may, for example, implement the functionality described above in connection with block 602 of fig. 6, and/or the like. Although direct input of the source style is described in the present embodiment, the source style may be extracted from the input source document according to an alternative embodiment.
According to an exemplary but non-limiting embodiment, intelligent document-rewriting device 900 may also include a module 904 for outputting the target document with the trained deep learning model. Module for outputting the target document with the trained deep learning model 904 may, for example, implement the functionality described above in connection with block 604 of fig. 6, and/or the like.
According to an exemplary but non-limiting embodiment, intelligent document rewrite apparatus 900 may also optionally include a module 906 for performing manually-assisted rewrite adjustments. Module for manually assisted overwrite adjustment 906 may, for example, implement the functionality described above in connection with block 606 of fig. 6, or the like.
According to an exemplary but non-limiting embodiment, intelligent document-rewriting device 900 may also include a module 908 for outputting results. The module for outputting results 908 may, for example, implement the functionality described above in connection with block 608 of fig. 6, and/or the like. The output result may be, for example, the target document or part thereof rewritten by the deep learning model, or may be the target document or part thereof adjusted by the manual-assisted rewriting.
Although the present application describes the intelligent document-rewriting method and apparatus based on deep learning by taking the seq2seq model as an example, the present application is not limited thereto, but can be applied to any deep learning model in the existing and future technologies.
Those of ordinary skill in the art appreciate that the benefits of the disclosure are not realized in full in any single embodiment. Various combinations, modifications, and alternatives will be apparent to one of ordinary skill in the art in light of this disclosure.
Furthermore, the term "or" is intended to mean an inclusive "or" rather than an exclusive "or". That is, unless specified otherwise or clear from context, the phrase "X" employing "a" or "B" is intended to mean any of the natural inclusive permutations. That is, the phrase "X" is satisfied using either "a" or "B" by any of the following examples: x is A; x is B; or X employs both A and B. The terms "connected" and "coupled" can mean the same thing, meaning that two devices are electrically connected. In addition, the articles "a" and "an" as used in this disclosure and the appended claims should generally be construed to mean "one or more" unless specified otherwise or clear from context to be directed to a singular form.
Various aspects or features will be presented in terms of systems that may include a number of devices, components, modules, and the like. It is to be understood and appreciated that the various systems may include additional devices, components, modules, etc. and/or may not include all of the devices, components, modules etc. discussed in connection with the figures. Combinations of these approaches may also be used.
The various illustrative logics, logical blocks, modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. Further, at least one processor may comprise one or more modules operable to perform one or more of the steps and/or actions described above. For example, the embodiments described above in connection with the various methods may be implemented by a processor and a memory coupled to the processor, wherein the processor may be configured to perform any of the steps of any of the methods described above, or any combination thereof.
Further, the steps and/or actions of a method or algorithm described in connection with the aspects disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. For example, the embodiments described above in connection with the various methods may be implemented by a computer readable medium having stored thereon computer program code which, when executed by a processor/computer, performs any of the steps of any of the methods described above, or any combination thereof.
All structural and functional equivalents to the elements of the various aspects described throughout this disclosure that are known or later come to be known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be encompassed by the claims. Moreover, nothing disclosed herein is intended to be dedicated to the public regardless of whether such disclosure is explicitly recited in the claims.
Claims (22)
1. A method for intelligent manuscript style rewriting based on deep learning model is characterized by comprising the following steps:
receiving a source document and at least one target style associated with a source style;
for each of the one or more natural sentences of the source document:
generating, by a deep learning model, a semantic vector corresponding to the natural sentence of the source document based on the source style; and
generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and
sequentially merging one or more target natural sentences corresponding to the one or more natural sentences of the source document to generate at least one target document associated with the at least one target style.
2. The method of claim 1, wherein the deep learning model comprises an encoder and a decoder, wherein
Semantic vectors corresponding to natural sentences of the source document are generated by an encoder of the deep learning model based on the source style, and
a target natural sentence corresponding to the semantic vector is generated by a decoder of the deep learning model based on the at least one target style.
3. The method of claim 2, further comprising:
segmenting natural sentences of the source document, and wherein
The encoder of the deep learning model includes a plurality of cascaded first unit modules, wherein each word in the participled natural sentence is sequentially input to the plurality of cascaded first unit modules, respectively.
4. The method of claim 3, further comprising:
generating, by the plurality of cascaded first unit modules, a present-level output based on an output of a previous-level first unit and a word input to the present-level in the participled natural sentence, wherein the first-level first unit is an output of the previous level in the source style, and a last-level first unit outputs a semantic vector corresponding to the natural sentence of the source document.
5. The method of claim 3, wherein the decoder of the deep learning model comprises a plurality of cascaded second unit modules, the method further comprising:
generating, by the plurality of cascaded second unit modules, target words corresponding to the semantic vectors based on the at least one target style, respectively; and
and combining the target words generated by the plurality of cascaded second unit modules to form the target natural sentence.
6. The method of claim 3, further comprising filling in the input of the redundant first unit modules with blanks when the number of words resulting from the tokenization of the natural sentence of the source document is less than the number of the plurality of cascaded first unit modules.
7. The method of claim 3, further comprising segmenting natural sentences of the source document when a number of words resulting from the segmentation of the natural sentences is greater than a number of the plurality of cascaded first unit modules.
8. The method of claim 1, wherein the source style is received externally or extracted directly from the source document.
9. The method of claim 1, further comprising training the deep learning model, wherein training the deep learning model comprises:
setting a feature library, wherein the feature library comprises two or more features related to intelligent manuscript style rewriting;
generating a document material library comprising pairs of articles associated with at least two features in the feature library; and
training the deep learning model based on the library of textual materials.
10. The method of claim 9, wherein generating a library of manuscript materials comprises one or more of, or any combination of:
for a particular feature in the feature library:
(i) capturing all articles with the specific characteristics from a characteristic website;
(ii) retrieving articles with high relevance from a search engine based on the specific characteristics; and
(iii) and learning out the marking model by utilizing machine learning so as to find out articles related to the specific characteristics in texts crawled from the network.
11. An apparatus for intelligent manuscript style rewrite based on deep learning model, comprising:
means for receiving a source document and at least one target style associated with a source style;
for each of the one or more natural sentences of the source document:
means for generating, by a deep learning model, semantic vectors corresponding to natural sentences of the source document based on the source style; and
means for generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and
means for sequentially merging the target natural sentences to generate at least one target manuscript associated with the at least one target style.
12. The apparatus of claim 11, wherein the deep learning model comprises an encoder and a decoder, wherein
Semantic vectors corresponding to natural sentences of the source document are generated by an encoder of the deep learning model based on the source style, and
a target natural sentence corresponding to the semantic vector is generated by a decoder of the deep learning model based on the at least one target style.
13. The apparatus of claim 12, further comprising:
means for tokenizing natural sentences of the source document, and wherein
The encoder of the deep learning model includes a plurality of cascaded first unit modules, wherein each word in the participled natural sentence is sequentially input to the plurality of cascaded first unit modules, respectively.
14. The apparatus of claim 13, further comprising:
a module for generating an output of a current stage by the plurality of cascaded first unit modules based on an output of a previous stage first unit and a word input to the current stage in the participled natural sentence, wherein the first stage first unit takes the source style as an output of the previous stage, and a last stage first unit outputs a semantic vector corresponding to the natural sentence of the source document.
15. The apparatus of claim 13, wherein the decoder of the deep learning model comprises a plurality of cascaded second unit modules, the apparatus further comprising:
means for generating, by the plurality of cascaded second unit modules, target words corresponding to the semantic vectors based on the at least one target style, respectively; and
and a module for combining the target words generated by the plurality of cascaded second unit modules to form a target natural sentence.
16. The apparatus of claim 13, further comprising means for filling in the input of redundant first unit modules with blanks when the number of words resulting from the tokenization of the natural sentence of the source document is less than the number of the plurality of cascaded first unit modules.
17. The apparatus of claim 13, further comprising means for segmenting natural sentences of the source document when a number of words resulting from the segmentation of the natural sentences is greater than a number of the plurality of cascaded first unit modules.
18. The apparatus of claim 11, wherein the source style is received externally or extracted directly from the source document.
19. The apparatus of claim 11, further comprising means for training the deep learning model, wherein the means for training the deep learning model comprises:
a module for setting a feature library, the feature library comprising two or more features related to intelligent document style overwriting;
means for generating a library of textual material comprising pairs of articles associated with at least two features in the library of features; and
means for training the deep learning model based on the library of manuscript materials.
20. The apparatus of claim 19, wherein the means for generating the library of manuscript materials comprises one or more of, or any combination of:
for a particular feature in the feature library:
(i) a module for crawling all articles with the specific feature from a featured website;
(ii) a module for retrieving articles with high relevance from a search engine based on the specific features; and
(iii) and a module for learning the marking model by machine learning to find articles related to the specific characteristics in texts crawled from the network.
21. An apparatus for intelligent manuscript style rewrite based on deep learning model, comprising:
a memory; and
a processor coupled to the memory, the processor configured to:
receiving a source manuscript and at least one target style which are associated with a source style, wherein the source manuscript comprises one or more natural sentences;
for each of the one or more natural sentences of the source document:
generating, by a deep learning model, semantic vectors corresponding to natural sentences of the source document based on the source style; and
generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and
sequentially merging the target natural sentences to generate at least one target manuscript associated with the at least one target style.
22. A computer readable medium storing processor-executable instructions for intelligent manuscript style rewrite based on a deep learning model, the processor-executable instructions when executed by a processor cause the processor to:
receiving a source document and at least one target style associated with a source style;
for each of the one or more natural sentences of the source document:
generating, by a deep learning model, semantic vectors corresponding to natural sentences of the source document based on the source style; and
generating, by the deep learning model, a target natural sentence corresponding to the semantic vector based on the at least one target style; and
sequentially merging the target natural sentences to generate at least one target manuscript associated with the at least one target style.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910780331.4A CN110688834B (en) | 2019-08-22 | 2019-08-22 | Method and equipment for carrying out intelligent manuscript style rewriting based on deep learning model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910780331.4A CN110688834B (en) | 2019-08-22 | 2019-08-22 | Method and equipment for carrying out intelligent manuscript style rewriting based on deep learning model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110688834A true CN110688834A (en) | 2020-01-14 |
| CN110688834B CN110688834B (en) | 2023-10-31 |
Family
ID=69108564
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910780331.4A Active CN110688834B (en) | 2019-08-22 | 2019-08-22 | Method and equipment for carrying out intelligent manuscript style rewriting based on deep learning model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110688834B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111553146A (en) * | 2020-05-09 | 2020-08-18 | 杭州中科睿鉴科技有限公司 | News writing style modeling method, writing style-influence analysis method and news quality evaluation method |
| CN111768755A (en) * | 2020-06-24 | 2020-10-13 | 华人运通(上海)云计算科技有限公司 | Information processing method, information processing apparatus, vehicle, and computer storage medium |
| CN111931496A (en) * | 2020-07-08 | 2020-11-13 | 广东工业大学 | Text style conversion system and method based on recurrent neural network model |
| CN113742461A (en) * | 2020-05-28 | 2021-12-03 | 阿里巴巴集团控股有限公司 | Dialogue system test method and device and statement rewriting method |
| WO2022105229A1 (en) * | 2020-11-20 | 2022-05-27 | 北京搜狗科技发展有限公司 | Input method and apparatus, and apparatus for inputting |
| WO2023115914A1 (en) * | 2021-12-20 | 2023-06-29 | 山东浪潮科学研究院有限公司 | Method and device for generating document having consistent writing style, and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106997370A (en) * | 2015-08-07 | 2017-08-01 | 谷歌公司 | Text classification and conversion based on author |
| CN108304436A (en) * | 2017-09-12 | 2018-07-20 | 深圳市腾讯计算机系统有限公司 | The generation method of style sentence, the training method of model, device and equipment |
| CN109344391A (en) * | 2018-08-23 | 2019-02-15 | 昆明理工大学 | Multiple features fusion Chinese newsletter archive abstraction generating method neural network based |
| CN109583952A (en) * | 2018-11-28 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Advertising Copy processing method, device, equipment and computer readable storage medium |
| CN109885811A (en) * | 2019-01-10 | 2019-06-14 | 平安科技（深圳）有限公司 | Written style conversion method, device, computer equipment and storage medium |
-
2019
- 2019-08-22 CN CN201910780331.4A patent/CN110688834B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106997370A (en) * | 2015-08-07 | 2017-08-01 | 谷歌公司 | Text classification and conversion based on author |
| CN108304436A (en) * | 2017-09-12 | 2018-07-20 | 深圳市腾讯计算机系统有限公司 | The generation method of style sentence, the training method of model, device and equipment |
| CN109344391A (en) * | 2018-08-23 | 2019-02-15 | 昆明理工大学 | Multiple features fusion Chinese newsletter archive abstraction generating method neural network based |
| CN109583952A (en) * | 2018-11-28 | 2019-04-05 | 深圳前海微众银行股份有限公司 | Advertising Copy processing method, device, equipment and computer readable storage medium |
| CN109885811A (en) * | 2019-01-10 | 2019-06-14 | 平安科技（深圳）有限公司 | Written style conversion method, device, computer equipment and storage medium |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111553146A (en) * | 2020-05-09 | 2020-08-18 | 杭州中科睿鉴科技有限公司 | News writing style modeling method, writing style-influence analysis method and news quality evaluation method |
| CN113742461A (en) * | 2020-05-28 | 2021-12-03 | 阿里巴巴集团控股有限公司 | Dialogue system test method and device and statement rewriting method |
| CN111768755A (en) * | 2020-06-24 | 2020-10-13 | 华人运通(上海)云计算科技有限公司 | Information processing method, information processing apparatus, vehicle, and computer storage medium |
| CN111931496A (en) * | 2020-07-08 | 2020-11-13 | 广东工业大学 | Text style conversion system and method based on recurrent neural network model |
| WO2022105229A1 (en) * | 2020-11-20 | 2022-05-27 | 北京搜狗科技发展有限公司 | Input method and apparatus, and apparatus for inputting |
| WO2023115914A1 (en) * | 2021-12-20 | 2023-06-29 | 山东浪潮科学研究院有限公司 | Method and device for generating document having consistent writing style, and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110688834B (en) | 2023-10-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109753566B (en) | Model training method for cross-domain emotion analysis based on convolutional neural network | |
| CN110119765B (en) | Keyword extraction method based on Seq2Seq framework | |
| CN110688834B (en) | Method and equipment for carrying out intelligent manuscript style rewriting based on deep learning model | |
| US10650188B2 (en) | Constructing a narrative based on a collection of images | |
| US11238093B2 (en) | Video retrieval based on encoding temporal relationships among video frames | |
| CN109858041B (en) | Named entity recognition method combining semi-supervised learning with user-defined dictionary | |
| CN109165380B (en) | Neural network model training method and device and text label determining method and device | |
| CN110263325B (en) | Chinese word segmentation system | |
| CN111950287B (en) | Entity identification method based on text and related device | |
| CN112131350A (en) | Text label determination method, text label determination device, terminal and readable storage medium | |
| CN111125333B (en) | Generation type knowledge question-answering method based on expression learning and multi-layer covering mechanism | |
| CN109977220B (en) | Method for reversely generating abstract based on key sentence and key word | |
| CN112163092B (en) | Entity and relation extraction method, system, device and medium | |
| CN110807324A (en) | Video entity identification method based on IDCNN-crf and knowledge graph | |
| CN110750635A (en) | Joint deep learning model-based law enforcement recommendation method | |
| Liu et al. | Uamner: uncertainty-aware multimodal named entity recognition in social media posts | |
| CN112966525B (en) | Law field event extraction method based on pre-training model and convolutional neural network algorithm | |
| CN114417851B (en) | Emotion analysis method based on keyword weighted information | |
| CN115794999A (en) | Patent document query method based on diffusion model and computer equipment | |
| CN112287687A (en) | Case tendency extraction type summarization method based on case attribute perception | |
| Raphal et al. | Survey on abstractive text summarization | |
| CN114387537A (en) | Video question-answering method based on description text | |
| CN115510236A (en) | Chapter-level event detection method based on information fusion and data enhancement | |
| CN113127604B (en) | Comment text-based fine-grained item recommendation method and system | |
| Abolghasemi et al. | HTS-DL: hybrid text summarization system using deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20200923 Address after: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman, British Islands Applicant after: Innovative advanced technology Co.,Ltd. Address before: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman, British Islands Applicant before: Advanced innovation technology Co.,Ltd. Effective date of registration: 20200923 Address after: Cayman Enterprise Centre, 27 Hospital Road, George Town, Grand Cayman, British Islands Applicant after: Advanced innovation technology Co.,Ltd. Address before: A four-storey 847 mailbox in Grand Cayman Capital Building, British Cayman Islands Applicant before: Alibaba Group Holding Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |