CN110362797B - Research report generation method and related equipment - Google Patents
Research report generation method and related equipment Download PDFInfo
- Publication number
- CN110362797B CN110362797B CN201910513763.9A CN201910513763A CN110362797B CN 110362797 B CN110362797 B CN 110362797B CN 201910513763 A CN201910513763 A CN 201910513763A CN 110362797 B CN110362797 B CN 110362797B
- Authority
- CN
- China
- Prior art keywords
- report
- outline
- research
- dictionary
- word
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/10—Text processing
- G06F40/166—Editing, e.g. inserting or deleting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/237—Lexical tools
- G06F40/242—Dictionaries
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Machine Translation (AREA)
Abstract
The application discloses a research report generating method and related equipment, wherein a research report dictionary is built by utilizing a plurality of collected research reports, and corresponding research reports are automatically output according to event texts, the research report dictionary, outline generating models and report generating models, wherein the outline generating models and the report generating models select word forming word sequences from the research report dictionary as report outline and research report according to a probability optimal principle, so that the technical problems that labor writing reports consumes a lot of effort and labor cost in the prior art are overcome, and the quality of the research reports generated according to the report outline is higher.
Description
Technical Field
The application relates to the field of report generation, in particular to a research report generation method and related equipment.
Background
LSTM (Long Short-Term Memory) is a Long and Short Term Memory network, a time-cycled neural network, suitable for processing and predicting important events with relatively Long intervals and delays in a time series.
A Variational Auto-Encoder (VAE) is a depth generation model.
In the financial field, a large number of reports in fixed format are involved, such as research reports, a stock instruction book and investment intention books, and reports of different companies in different industries have different requirements. Such report writing often requires high timeliness and a lot of data collection and analysis, and in the prior art, the data is generally collected manually, analyzed and the report is written, so that the labor cost is high and a lot of effort is required.
Disclosure of Invention
The present application aims to solve at least one of the technical problems in the related art to some extent. To this end, an object of the present application is to provide a study report generating method and related apparatus for automatically generating a study report from event text.
The technical scheme adopted by the application is as follows:
in a first aspect, the present application provides a method for generating a research report, including:
study report acquisition step: acquiring a plurality of research reports from a plurality of information sources;
dictionary acquisition step: performing data preprocessing and feature selection on a plurality of the study reports to construct a study report dictionary;
outline acquisition: acquiring a report outline corresponding to the event text according to the event text, the research report dictionary and an outline generation model, wherein the outline generation model selects a word sequence formed by a plurality of words from the research report dictionary as the report outline according to a probability optimal principle;
report generation: and acquiring a research report according to the event text, the report outline, the research report dictionary and a report generation model, wherein the report generation model selects a plurality of word composition word sequences from the research report dictionary as the research report according to a probability optimal principle.
Further, the report generation model selects and outputs words one by one from the study report dictionary to generate the study report.
Further, the outline generating model updates the report outline according to the event text, the report outline, the research report dictionary, and the last word output by the report generating step.
Further, the dictionary obtaining step further includes: a beginning tag and an ending tag are added to the study report dictionary.
Further, the outline generation model includes:
vector representation is carried out on the event text and the beginning marks according to the research report dictionary so as to obtain event vectors and beginning mark vectors;
acquiring the hidden layer state of the event text and the hidden layer state of the beginning mark according to the event vector, the beginning mark vector and an LSTM network;
and acquiring the report outline according to the hidden layer state of the event text, the hidden layer state of the beginning mark and the attention mechanism.
Further, the outline generating model further includes:
performing vector representation on the last word output by the report generating step according to the research report dictionary to obtain a word vector;
acquiring the hidden layer state of the word according to the word vector and the LSTM network;
and updating the report outline according to the hidden layer state of the word, the hidden layer state of the event text and the attention mechanism.
Further, the outline generation model includes:
acquiring the report outline according to the event text, the beginning mark and a transducer model;
and updating the report outline according to the last word, the event text and the transducer model which are output by the report generating step.
Further, the report generation model includes a VAE generation model.
Further, the research report generating method further comprises:
and carrying out event entity recognition according to the research reports obtained in the research report acquisition step to obtain corresponding event texts, wherein a plurality of research reports and the corresponding event texts form a training data set, and the training data set is used for training the outline generation model and the report generation model.
In a second aspect, the present application provides a study report generating apparatus comprising:
a study report acquisition module for acquiring a plurality of study reports from a plurality of information sources;
a dictionary acquisition module for performing data preprocessing and feature selection on a plurality of the study reports to construct a study report dictionary;
the outline acquisition module is used for acquiring a report outline corresponding to the event text according to the event text, the research report dictionary and an outline generation model, wherein the outline generation model selects a plurality of word forming word sequences from the research report dictionary as the report outline according to a probability optimal principle;
and the report generation module is used for acquiring a research report according to the event text, the report outline, the research report dictionary and a report generation model, and the report generation model selects a plurality of word composition word sequences from the research report dictionary as the research report according to a probability optimal principle.
In a third aspect, the present application provides a computer-readable storage medium storing computer-executable instructions for causing the computer to perform the study report generating method.
The beneficial effects of the application are as follows:
according to the application, the collected multiple study reports are utilized to construct a study report dictionary, and then corresponding study reports are automatically output according to the event text, the study report dictionary, the outline generation model and the report generation model, wherein the outline generation model and the report generation model select word sequences from the study report dictionary to form word sequences as the report outline and the study report according to the probability optimization principle, so that the technical problems that labor writing and reporting consume a great deal of effort and labor cost in the prior art are overcome, and the quality of the study report generated according to the report outline is higher.
Drawings
FIG. 1 is a method flow diagram of one embodiment of a method of generating a research report in the present application;
FIG. 2 is a method flow diagram of a first embodiment of a method of generating a research report in accordance with the present application;
FIG. 3 is a schematic diagram of the training process of FIG. 2;
FIG. 4 is a schematic diagram of an example of generating a research report using the research report generation method of the present application;
FIG. 5 is a method flow diagram of a second embodiment of a method of generating a research report in accordance with the present application;
FIG. 6 is a schematic diagram of the training process of FIG. 5;
fig. 7 is a block diagram showing the structure of an embodiment of the study report generating apparatus of the present application.
Detailed Description
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other.
Example 1
The idea of the research report generation method in this embodiment is mainly: as a outline is drawn before writing, an outline of a research report (hereinafter referred to as a research report) is generated according to the event text, and then a final research report is generated according to the outline and the event text. Event text refers herein to an event-based narrative text, that is, a collection of words that recite one or more events (or questions), such as event-based financial news (hereinafter referred to as news) in the financial field, from which people perform macro analysis to generate macro research reports (hereinafter referred to as macro research reports); for another example, the treatise in the college entrance examination paper is written, the text of a certain problem or a certain event is given in the title, the writing is required according to the text, and the written writing is equivalent to the research report in the text. In particular, with reference to FIG. 1, FIG. 1 is a method flow diagram of one embodiment of a method of generating a research report in the present application; the research report generation method comprises the following steps:
study report acquisition step: the method comprises the steps of obtaining a plurality of research reports from a plurality of information sources, wherein the information sources can be Internet or paper data sources (such as libraries), the financial field is taken as an example, the information sources for obtaining the research reports can be financial websites such as new wave financial, eastern financial and common flowers and stars, and the information sources for composition can be library or composition resources on the network;
dictionary acquisition step: performing data preprocessing and feature selection on the obtained multiple research reports to construct a research report dictionary, wherein the data preprocessing comprises word segmentation processing on the research reports, counting the occurrence times of words in the research reports to perform feature selection, and determining whether to put the words into the research report dictionary according to the occurrence times;
outline acquisition: according to the event text, the research report dictionary and the outline generation model, the outline generation model selects a plurality of word forming word sequences from the research report dictionary to be used as the report outline according to the probability optimal principle, and the word selecting method according to the probability optimal principle can be a word output method adopting global search optimal probability, namely only one word is output at a time, and the word forming word sequences of the plurality of words are used as the report outline; or a method of searching for word output with preset number of optimal probability by adopting a cluster search algorithm, namely outputting preset number of words at one time, and selecting a word sequence formed by a plurality of words according to the principle of optimal probability by the words output for multiple times as a report outline;
report generation: according to the event text, the report outline, the research report dictionary and the report generation model, the report generation model selects word sequences formed by a plurality of words from the research report dictionary according to the principle of probability optimization to serve as the research report, and the method is similar to outline generation, can adopt word output of global search optimal probability or can adopt a cluster search algorithm to select words for output, and is not repeated.
According to the application, the research report dictionary is constructed by utilizing the collected multiple research reports, and the corresponding research reports are automatically output according to the event text, the research report dictionary, the outline generation model and the report generation model, so that the technical problems that a great deal of effort and labor cost are consumed in manually compiling the report in the prior art are overcome, the manpower is liberated, the output efficiency of the research report is improved, and the manpower cost is reduced. And the quality of the research report generated according to the report outline is higher. Further, the overall framework of the research report generation method is an Encoder-Decoder (so-called encoding, which is to convert an input sequence into a vector of a fixed length, and decoding, which is to convert a previously generated fixed vector into an output sequence), i.e., to input a sequence and output a sequence. The procedure of the Encoder is to change the sequence of event text into a fixed length vector representation, and the procedure of the Encoder is to change the fixed length vector representation into a variable length text sequence, i.e., a study report.
Taking a report in the financial field as an example, the following describes the process of collecting a study report and constructing a study report dictionary:
first, the research report text (11 ten thousand pieces of data in total) of websites such as new and wasted financial, eastern financial and the same flower in the macroscopic economic plate is crawled. And extracting Chinese characters and Chinese punctuation marks in the research report through the regular expression. The digital symbols are not used because they are difficult to directly generate high-accuracy results. Then, the text data is segmented into single words by adopting the jieba segmentation to obtain word sets, and words which repeatedly appear more than 5 times in the word sets are selected and put into a research report dictionary in order to improve the calculation efficiency. In the dictionary obtaining step, a beginning flag and an ending flag are added to the study report dictionary. Specifically, the first four keys of the research report dictionary (a key may be understood as an element in the dictionary) are a mask flag (mask), an unknown flag (unk), a start flag (start), and an end flag (end), respectively. Wherein a mask flag (mask) is used to indicate that unwanted information is masked out. Unknown tokens (unk) are used to represent words in a word set that are not in a dictionary, typically words that occur infrequently but are of significance, such as institution names, person names. A start marker (start) and an end marker (end) are used to be added to the start and end of each piece of text data, respectively, to designate the start and end of the text data. The keys of the dictionary are words (containing individual punctuation marks) and the value of the dictionary is the sequence number of the word.
In addition, the research report generation method further comprises the following steps:
and carrying out event entity recognition according to the research reports obtained in the research report acquisition step to obtain corresponding event texts, wherein a plurality of research reports and the corresponding event texts form a training data set, and the training data set is used for training a outline generation model and a report generation model. Taking a research report in the financial field as an example, when a macroscopic research report is obtained by crawling, processing the macroscopic research report by utilizing rule matching and part-of-speech matching to obtain a news text, extracting the text containing event related contents as news, and changing the data into a one-to-one set of news and macroscopic research report as a training data set. Specifically, the research report may be segmented and saved as an array, and then the text in the array is matched through the set event rule to obtain the event paragraph as the news text.
In this embodiment, two outline generating models are provided, wherein the first outline generating model refers to fig. 2, and fig. 2 is a flowchart of a method of a first specific embodiment of the study report generating method in the present application; taking news in the financial field as an example for an event text, the first outline generation model includes:
vector representation is carried out on the event text and the beginning mark (< start >) according to the research report dictionary so as to obtain an event vector and a beginning mark vector; in this embodiment, a text may be converted into a fixed length vector representation by word embedding.
Acquiring hidden layer states of event texts according to event vectors and a bidirectional LSTM network, and acquiring hidden layer states of the beginning marks according to the beginning mark vectors and the unidirectional LSTM network, wherein the unidirectional LSTM network has less processing time, but the bidirectional LSTM network can better capture bidirectional semantic dependence;
acquiring a report outline according to the hidden layer state of the event text, the hidden layer state of the beginning mark and the attention mechanism;
then, carrying out vector representation on the last word output by the report generating step according to the research report dictionary to obtain a word vector, and converting the word into the word vector by word casting;
acquiring the hidden layer state of the word according to the word vector and the unidirectional LSTM network;
the report outline is updated according to the hidden state of the word, the hidden state of the event text and the attention mechanism.
In this embodiment, the report generating model includes a VAE generating model, in which a probability distribution sequence of a word can be obtained after an attention mechanism is performed in a first outline generating model, and the outline generating model selects a vocabulary combination for researching an optimal probability in a report dictionary according to probability distribution information of the word, and outputs the vocabulary as a report outline. And in the same way, in the report generation model, the other probability distribution information can be obtained after the processing of the VAE generation model, and the report generation model selects vocabulary combinations with optimal probability in the research report dictionary to output and form a research report according to the probability distribution information. Specifically, referring to FIG. 2, the beginning of the datagram is marked with a < start > tag and the end is marked with an < end > tag. The probability distribution of the next word of the study report is predicted from the < start > tag using the first outline generation model and the VAE generation model, and word output is selected from the study report dictionary based on the probability distribution to generate the study report. In brief, when the research report generating method is executed, firstly, news and a first < start > mark are input, a first word of a research report is output after being processed by a outline generating model and a report generating model, the first word is returned to an input end to replace the first mark and is regenerated into a second output, the second output is returned to the input end to replace the first output, and the words are output one by one until the output is an end mark < end >, and finally, the generation of the research report is finished. That is, the word currently output needs to depend on the word output last, and the report outline is updated according to the word output last and the news input, so that the quality of the generated research report is effectively improved. In the case of selecting words from the study report dictionary according to the probability distribution to generate the study report, the word with the highest probability (i.e., only one word is output at a time) can be selected from the study report dictionary by the global search of the optimal solution, but when the study report dictionary is huge, the space efficiency of the global search of the optimal solution is low. Thus, a bundle search algorithm may be employed to increase search efficiency, which uses a beam size parameter to limit the number of possible words that remain at each step, taking into account not only the probability of a single word, but also the probability of the preceding and following words coming together. Therefore, after the VAE generating model is processed, a cluster search algorithm is used to obtain a preset number of words with the highest probability, in this embodiment, the beam size=3, that is, the preset number is 3, and three maximum possible results are reserved in each step. 3 outputs can be obtained in each prediction, the 3 outputs are returned to the input end for the next prediction, and after the prediction is finished, one output is selected as the final output according to all the words used for generating the research report and the probability optimal principle, so that the final research report is obtained.
The outline generating model and the VAE generating model need to be trained before formal use, and referring to FIG. 3, FIG. 3 is a schematic diagram of the training process of FIG. 2; taking news in the financial field as event text, the following describes a training process of a first outline generation model:
firstly, defining input news as X, wherein X is word in news, the expression of news is as follows,
X＝(x 1 ,…,x m ) (1)
the decoding of the potential vector has two stages, wherein O is a word in the outline of the research report generated in the first stage, O is a word in the outline of the research report generated in the second stage, Y is a word in the research report generated in the second stage, the length of a definition generated text is L, and the expression is as follows:
O＝(o 1 ,…,o L ) (2)
Y＝(y 1 ,…,y L ) (3)
extracting Chinese characters and Chinese characters from news and macroscopic research report of the training data set through regular matching, then performing preliminary length statistics on texts by taking words as units after jieba word segmentation, wherein the finally obtained length statistics data are shown in table 1:
TABLE 1 news and report Length statistics
The news and the research report are then "truncated and supplemented" to the same length set for training purposes, e.g., the news length is 30 words and the research report length is 200 words. The input news text and the ground report text are changed into vector representations through word scrolling to obtain news vectors and ground report vectors, and a layer of Embedding network is directly adopted here. Since LSTM networks are unable to encode back-to-front information when modeling sentences, bi-directional LSTM networks are better able to capture bi-directional semantic dependencies. The news vector after the Embedding is then entered into the bi-directional LSTM network. Wherein, after the news vector is input into the bidirectional LSTM network, the hidden layer state of the news can be obtained and expressed as H, H is the hidden layer sub-state, t is time, and the expression of H is as follows:
in order to generate a high-quality report, the model encoding and decoding process needs to fully absorb the structure and content of the corresponding report, and the Decoder process predicts the probability distribution of the next word of the report in the vocabulary through the LSTM network. The hidden state of the macro report is denoted as S, and the hidden state of each time step depends on the previous time input and the hidden state of the previous time, and the expression is as follows:
the hidden state of news and the hidden state of macroscopic studies calculate the attention score by the attention mechanism (as in equation (6)), where the attention is calculated using the softmax function using the multiplicative attention score to accumulate the attention score (as in equation (7)). The context vector is obtained by a weighted average of the attention weight and the hidden layer state of the news, as in equation (8), i.e. multiplying the attention weight and the hidden layer state of the news. The expressions of the formulas (6), (7) and (8) are as follows:
then, the context vector and the hidden layer state of the macro report are connected in series to update and obtain the attention hidden layer state. The predicted output of a word is calculated from the hidden state of attention. The expression is as follows:
where Wc is a model parameter.
The first stage objective function of decoding is as follows:
where P is probability.
To this end, a probability distribution of words in the study report dictionary is obtained from the news input, and a synopsis of the report can be generated from the probability distribution.
In the second stage of decoding, news and a synopsis generated according to the news need to be input in order to generate the final report.
The decoding model adopts a variational self-encoder model (VAE), generates target variables by fusing double-input news X and outline O, learns posterior probability distribution of hidden variables z, and can be rewritten as the following expression:
assuming that the posterior distribution is a standard normal distribution, the original text is re-decoded from the random sampling of the distribution. Training with the weight reconstruction loss and regularization loss, then ELBO can be expressed as:
logP(X,O)≥E q(z|x,o) [logp(x,o|z)]-KL(q(z|x,o)||p(z)) (12)
to the right of the above inequality is ELBO, where the first term is to sample attention from P (z|x, O), calculate cross entropy loss using the sampled attention as input to the decoder, and the second term is to measure the similarity of the two probability distributions by KL-divergence, ensuring that the posterior distribution is close to the a priori distribution.
The second stage objective function of decoding is as follows:
so far, the probability distribution information of the words can be obtained again according to the news and report outline, and the final research report can be obtained according to the probability distribution information.
The global objective function is obtained by summing the loss functions of the two decoding stages, expressed as follows:
after inputting the training data set of the one-to-one news and macro report into the model of fig. 3 and training for 50 rounds, the final model parameters can be adjusted and determined. Specifically, taking global optimal search as an example, only one word is output at a time when the research report is generated, each macroscopic research report is input into the model in the sequence of the beginning mark, macroscopic research report and ending mark, each time the training is carried out by inputting the words one by one, for example, the beginning mark can be input at the beginning to obtain an output at the output end of the model, the output is the first word of the predicted research report, the output word is compared with the first word of the real research report, and the model parameters are modified according to the comparison result; and then inputting the first word of the macroscopic report into the model to obtain an output, wherein the output is the predicted second word of the macroscopic report, comparing the second word with the real macroscopic report, and adjusting the model parameters according to the comparison result again to continuously reduce the error between the model output and the real word. After the model is trained with a plurality of news-macro study data, the structure and parameters of the model are preserved after training. The newly input news can be predicted according to the model with the final determined model parameters, the news can be input by using the model of fig. 2, the research report of fig. 4 can be obtained, and fig. 4 is an example schematic diagram of the research report generated by using the research report generating method in the application.
Referring to fig. 5, fig. 5 is a method flow diagram of a second embodiment of a method of generating a research report in accordance with the present application; in fig. 5, the event text is exemplified by news in the financial domain, and the second outline generation model includes:
acquiring a report outline according to the event text, the beginning mark and the transducer model;
and updating the report outline according to the last word, the event text and the transducer model output by the report generating step. Wherein, the report generation model also adopts a VAE generation model. The transducer model is equivalent to replacing the LSTM network and the attention mechanism, the attention score of the transducer model is solved as in formula (15),
where Q, K, V are the three matrix vectors into which the input X is transformed. Similar to fig. 2, the model of fig. 5 is implemented by inputting news and a beginning < start > tag, outputting the first word of the report after processing by the outline generation model and the report generation model, returning the first word to the input end to replace the beginning tag, and so on, and the model of fig. 5 outputs the words of the report one by one. In addition, similarly, the outline generation model and the report generation model can select words by selecting the global search optimal solution, and the cluster search algorithm can also be applied to the model of fig. 5 to improve the search efficiency.
Referring to fig. 6, fig. 6 is a schematic diagram of the training process of fig. 5; the training data set is sequentially input into the model for training, and similar to fig. 4, the macroscopic report is input into the model in the form of a beginning mark-macroscopic report-ending mark, and the model parameters are adjusted according to the comparison of the model output and the real report words.
Example 2
Based on embodiment 1 providing embodiment 2, embodiment 2 provides a research report generating apparatus, and referring to fig. 7, fig. 7 is a block diagram of an embodiment of a research report generating apparatus in the present application, the research report generating apparatus includes:
a study report acquisition module for acquiring a plurality of study reports from a plurality of information sources;
a dictionary acquisition module for performing data preprocessing and feature selection on a plurality of study reports to construct a study report dictionary;
the outline acquisition module is used for acquiring a report outline corresponding to the event text according to the event text, the research report dictionary and the outline generation model, and the outline generation model selects a word sequence formed by a plurality of words from the research report dictionary as the report outline according to the probability optimal principle;
and the report generation module is used for acquiring a research report according to the event text, the report outline, the research report dictionary and the report generation model, and the report generation model selects a plurality of word forming word sequences from the research report dictionary to serve as the research report according to the probability optimal principle.
The specific operation of the study report generating apparatus will be described with reference to embodiment 1, and will not be described in detail. The research report generating device can automatically generate the research report, liberate manpower and improve the output efficiency of the research report.
Example 3
Embodiment 3 is provided based on embodiment 1, embodiment 3 providing a computer-readable storage medium storing computer-executable instructions for causing the computer to perform the study report generating method as described in embodiment 1. A specific description of the study report generating method may refer to the description of embodiment 1, and will not be repeated.
While the preferred embodiment of the present application has been described in detail, the present application is not limited to the embodiments, and those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present application, and the equivalent modifications or substitutions are included in the scope of the present application as defined in the appended claims.
Claims (6)
1. A research report generation method, comprising:
study report acquisition step: acquiring a plurality of research reports from a plurality of information sources;
dictionary acquisition step: feature selecting a plurality of the study reports to construct a study report dictionary, adding a beginning marker and an ending marker to the study report dictionary;
outline acquisition: acquiring a report outline corresponding to the event text according to the event text, the research report dictionary and an outline generation model, wherein the outline generation model selects a word sequence formed by a plurality of words from the research report dictionary as the report outline according to a probability optimal principle; wherein the outline generation model comprises: vector representation is carried out on the event text and the beginning marks according to the research report dictionary so as to obtain event vectors and beginning mark vectors; acquiring the hidden layer state of the event text and the hidden layer state of the beginning mark according to the event vector, the beginning mark vector and an LSTM network; acquiring the report outline according to the hidden layer state of the event text, the hidden layer state of the beginning mark and the attention mechanism; performing vector representation on the last word output by the report generating step according to the research report dictionary to obtain a word vector; acquiring the hidden layer state of the word according to the word vector and the LSTM network; updating the report outline according to the hidden layer state of the word, the hidden layer state of the event text and the attention mechanism; acquiring the report outline according to the event text, the beginning mark and a transducer model; updating the report outline according to the last word, the event text and the transducer model which are output in the report generating step;
report generation: acquiring a research report according to the event text, the report outline, the research report dictionary and a report generation model, wherein the report generation model selects a plurality of word forming word sequences from the research report dictionary as the research report according to a probability optimal principle; wherein the report generation model comprises a VAE generation model that selects word outputs from a study report dictionary according to a probability distribution to generate the study report.
2. The study report generating method of claim 1, wherein the report generating model selects and outputs words one by one from the study report dictionary to generate the study report.
3. The study report generating method of claim 2, wherein the outline generating model updates the report outline based on the event text, the report outline, the study report dictionary, a last word output by the report generating step.
4. A research report generating method according to any one of claims 1 to 3, wherein said research report generating method further comprises:
and carrying out event entity recognition according to the research reports obtained in the research report acquisition step to obtain corresponding event texts, wherein a plurality of research reports and the corresponding event texts form a training data set, and the training data set is used for training the outline generation model and the report generation model.
5. A study report generating apparatus, comprising:
a study report acquisition module for acquiring a plurality of study reports from a plurality of information sources;
a dictionary acquisition module for performing feature selection on a plurality of the study reports to construct a study report dictionary, and adding a beginning mark and an ending mark to the study report dictionary;
the outline acquisition module is used for acquiring a report outline corresponding to the event text according to the event text, the research report dictionary and an outline generation model, wherein the outline generation model selects a plurality of word forming word sequences from the research report dictionary as the report outline according to a probability optimal principle; wherein the outline generation model comprises: vector representation is carried out on the event text and the beginning marks according to the research report dictionary so as to obtain event vectors and beginning mark vectors; acquiring the hidden layer state of the event text and the hidden layer state of the beginning mark according to the event vector, the beginning mark vector and an LSTM network; acquiring the report outline according to the hidden layer state of the event text, the hidden layer state of the beginning mark and the attention mechanism; performing vector representation on the last word output by the report generating step according to the research report dictionary to obtain a word vector; acquiring the hidden layer state of the word according to the word vector and the LSTM network; updating the report outline according to the hidden layer state of the word, the hidden layer state of the event text and the attention mechanism; acquiring the report outline according to the event text, the beginning mark and a transducer model; updating the report outline according to the last word, the event text and the transducer model which are output in the report generating step;
the report generation module is used for acquiring a research report according to the event text, the report outline, the research report dictionary and a report generation model, and the report generation model selects a plurality of word forming word sequences from the research report dictionary as the research report according to a probability optimal principle; wherein the report generation model comprises a VAE generation model that selects word outputs from a study report dictionary according to a probability distribution to generate the study report.
6. A computer-readable storage medium storing computer-executable instructions for causing the computer to perform the study report generation method of any one of claims 1 to 4.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910513763.9A CN110362797B (en) | 2019-06-14 | 2019-06-14 | Research report generation method and related equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910513763.9A CN110362797B (en) | 2019-06-14 | 2019-06-14 | Research report generation method and related equipment |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110362797A CN110362797A (en) | 2019-10-22 |
| CN110362797B true CN110362797B (en) | 2023-10-13 |
Family
ID=68216086
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910513763.9A Active CN110362797B (en) | 2019-06-14 | 2019-06-14 | Research report generation method and related equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110362797B (en) |
Families Citing this family (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111046629B (en) * | 2019-12-16 | 2022-03-01 | 北大方正集团有限公司 | Outline display method, device and equipment |
| CN112242185A (en) * | 2020-09-09 | 2021-01-19 | 山东大学 | Medical image report automatic generation method and system based on deep learning |
| CN112417846B (en) * | 2020-11-25 | 2024-06-28 | 中译语通科技股份有限公司 | Text automatic generation method and device, electronic equipment and storage medium |
| CN113160963A (en) * | 2020-12-18 | 2021-07-23 | 中电云脑(天津)科技有限公司 | Event determination method and device, electronic equipment and storage medium |
| CN114490778A (en) * | 2022-02-15 | 2022-05-13 | 北京固加数字科技有限公司 | Financial research and report automatic generation system and method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140142939A1 (en) * | 2012-11-21 | 2014-05-22 | Algotes Systems Ltd. | Method and system for voice to text reporting for medical image software |
| CN109635302A (en) * | 2018-12-17 | 2019-04-16 | 北京百度网讯科技有限公司 | A kind of method and apparatus of training text summarization generation model |
-
2019
- 2019-06-14 CN CN201910513763.9A patent/CN110362797B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140142939A1 (en) * | 2012-11-21 | 2014-05-22 | Algotes Systems Ltd. | Method and system for voice to text reporting for medical image software |
| CN109635302A (en) * | 2018-12-17 | 2019-04-16 | 北京百度网讯科技有限公司 | A kind of method and apparatus of training text summarization generation model |
Non-Patent Citations (2)
| Title |
|---|
| A Self-Report Study that Gauges Perceived and Induced Emotion with Music;Darryl Griffiths et al.;《2015 Internet Technologies and Applications (ITA)》;20150908;第239-244页 * |
| 核电厂运行经验报告的定量评估方法;吴杰 等;《发电设备》;20160930;第30卷(第5期);第314-318页，第341页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110362797A (en) | 2019-10-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110362797B (en) | Research report generation method and related equipment | |
| CN106202010B (en) | Method and apparatus based on deep neural network building Law Text syntax tree | |
| CN110222188B (en) | Company notice processing method for multi-task learning and server | |
| CN114020862B (en) | Search type intelligent question-answering system and method for coal mine safety regulations | |
| CN113672708B (en) | Language model training method, question-answer pair generation method, device and equipment | |
| CN111858932A (en) | Multiple-feature Chinese and English emotion classification method and system based on Transformer | |
| CN111966812B (en) | Automatic question answering method based on dynamic word vector and storage medium | |
| CN110688857B (en) | Article generation method and device | |
| CN113779994B (en) | Element extraction method, element extraction device, computer equipment and storage medium | |
| CN108829823A (en) | A kind of file classification method | |
| CN113821635A (en) | Text abstract generation method and system for financial field | |
| CN118093834B (en) | AIGC large model-based language processing question-answering system and method | |
| CN111783423A (en) | Training method and device of problem solving model and problem solving method and device | |
| CN113127604B (en) | Comment text-based fine-grained item recommendation method and system | |
| CN116151235A (en) | Article generating method, article generating model training method and related equipment | |
| CN117094325B (en) | Named entity identification method in rice pest field | |
| CN111581386A (en) | Construction method, device, equipment and medium of multi-output text classification model | |
| CN114417824B (en) | Chapter-level relation extraction method and system based on dependency syntax pre-training model | |
| CN113157866B (en) | Data analysis method, device, computer equipment and storage medium | |
| CN110347813A (en) | A kind of corpus processing method, device, storage medium and electronic equipment | |
| CN116266266B (en) | Multi-tone word disambiguation method, device, equipment and storage medium | |
| CN113836893B (en) | Extraction type machine reading and understanding method integrating multiple paragraph information | |
| CN117786092B (en) | Commodity comment key phrase extraction method and system | |
| CN110287355B (en) | Remote sensing image description method based on retrieval topic memory network | |
| CN111158640B (en) | One-to-many demand analysis and identification method based on 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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |