CN110362797B - Research report generation method and related equipment - Google Patents

Abstract

The invention discloses a research report generation method and related equipment. The invention constructs a research report dictionary by using multiple collected research reports, and then automatically outputs corresponding research based on event text, research report dictionary, outline generation model and report generation model. In the report, the outline generation model and the report generation model select words from the research report dictionary to form a word sequence as the report outline and research report based on the principle of probability optimization, overcoming the existing technology in the existing technology that manually writing reports consumes a lot of effort and labor costs. issues, and the research reports generated based on the report outline are of higher quality.

Description

Research report generation method and related equipment

Technical field

The present invention relates to the field of report generation, in particular to a research report generation method and related equipment.

Background technique

LSTM (Long Short-Term Memory) is a long short-term memory network, a time-cyclic neural network, suitable for processing and predicting important events with relatively long intervals and delays in time series.

Variational Auto-Encoder (VAE) is a deep generative model.

The financial field involves a large number of fixed-format report writing work, such as research reports, prospectuses, and investment letters of intent. Reports from different companies in different industries have different requirements. The writing of these reports often requires high timeliness and a large amount of data collection and analysis. In the existing technology, data is generally collected manually, analyzed, and reports written, so labor costs are high and a lot of energy is consumed.

Contents of the invention

The present invention aims to solve one of the technical problems in the related art, at least to a certain extent. To this end, one object of the present invention is to provide a research report generation method and related equipment for automatically generating a research report based on event text.

The technical solution adopted by the present invention is:

In a first aspect, the present invention provides a method for generating a research report, including:

Research report collection steps: Obtain multiple research reports from multiple information sources;

Dictionary acquisition step: perform data preprocessing and feature selection on multiple research reports to build a research report dictionary;

Outline acquisition step: Obtain the report outline corresponding to the event text according to the event text, the research report dictionary and the outline generation model. The outline generation model selects multiple words from the research report dictionary to form words based on the principle of probability optimization. The sequence serves as an outline for said report;

Report generation step: obtain a research report based on the event text, the report outline, the research report dictionary and a report generation model. The report generation model selects multiple words from the research report dictionary based on the principle of probability optimization. Sequences of words serve as the research report.

Further, the report generation model selects and outputs words one by one from the research report dictionary to generate the research report.

Further, the outline generation model updates the report outline according to the event text, the report outline, the research report dictionary, and the previous word output by the report generation step.

Further, the dictionary acquisition step further includes: adding a beginning tag and an ending tag to the research report dictionary.

Further, the outline generation model includes:

Perform vector representation on the event text and the opening tag according to the research report dictionary to obtain event vectors and opening tag vectors;

Obtain the hidden layer state of the event text and the hidden layer state of the starting mark according to the event vector, the beginning mark vector and the LSTM network;

The report outline is obtained according to the hidden state of the event text, the hidden state of the opening mark and the attention mechanism.

Further, the outline generation model also includes:

Perform a vector representation on the previous word output by the report generation step according to the research report dictionary to obtain a word vector;

Obtain the hidden layer state of the word according to the word vector and 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.

Further, the outline generation model includes:

Obtain the report outline according to the event text, the opening tag and the Transformer model;

The report outline is updated according to the previous word output by the report generation step, the event text and the Transformer model.

Further, the report generation model includes a VAE generation model.

Further, the research report generating method also includes:

Event entity recognition is performed according to the research report obtained in the research report collection step to obtain the corresponding event text. Multiple research reports and corresponding event texts constitute a training data set. The training data set is used to train the outline generation model and The report generation model.

In a second aspect, the present invention provides a research report generating device, including:

The research report collection module is used to obtain multiple research reports from multiple information sources;

The dictionary acquisition module is used to perform data preprocessing and feature selection on multiple research reports to build a research report dictionary;

An outline acquisition module, configured to obtain a report outline corresponding to the event text based on the event text, the research report dictionary, and an outline generation model. The outline generation model selects multiple words from the research report dictionary based on the principle of probability optimization. form a sequence of words to serve as an outline for said report;

A report generation module is configured to obtain a research report based on the event text, the report outline, the research report dictionary, and a report generation model. The report generation model selects multiple items from the research report dictionary based on the principle of probability optimization. Words form word sequences as the research report.

In a third aspect, the present invention provides a computer-readable storage medium that stores computer-executable instructions, and the computer-executable instructions are used to cause the computer to execute the research report generating method.

The beneficial effects of the present invention are:

The present invention constructs a research report dictionary by using multiple collected research reports, and then automatically outputs corresponding research reports based on the event text, research report dictionary, outline generation model, and report generation model. Among them, the outline generation model and the report generation model are based on the most probable You Principle selects words from the research report dictionary to form a word sequence as the report outline and research report, overcoming the technical problem in the existing technology that manual writing of reports consumes a lot of effort and labor costs, and the research report generated based on the report outline is of higher quality.

Description of the drawings

Figure 1 is a method flow chart of an embodiment of a research report generating method in the present invention;

Figure 2 is a method flow chart of a first specific embodiment of a research report generating method in the present invention;

Figure 3 is a schematic diagram of the training process in Figure 2;

Figure 4 is a schematic diagram of an example of generating a research report using the research report generation method in the present invention;

Figure 5 is a method flow chart of a second specific embodiment of a research report generating method in the present invention;

Figure 6 is a schematic diagram of the training process in Figure 5;

Figure 7 is a structural block diagram of an embodiment of a research report generating device in the present invention.

Detailed ways

It should be noted that, as long as there is no conflict, the embodiments and features in the embodiments of this application can be combined with each other.

Example 1

The main idea of the research report generation method in this embodiment is: just like an outline is drawn up before writing, the outline of a research report (hereinafter referred to as the research report) is first generated based on the event text, and then the final research report is generated based on the outline and event text. Report. In this article, event text refers to event narrative text, that is, a text collection that describes one or more events (or issues), such as event-based financial news (hereinafter referred to as news) in the financial field. People conduct macro analysis based on the news. In order to produce a macro research report (hereinafter referred to as a macro research report); another example is the argumentative writing in the college entrance examination Chinese language. The text of a certain issue or event is given in the title and is required to be written based on this text. The written composition is also It is equivalent to the research report in this article. Specifically, refer to Figure 1, which is a method flow chart of an embodiment of a research report generation method in the present invention; the research report generation method includes:

Research report collection steps: Obtain multiple research reports from multiple information sources. The information sources can be the Internet or paper data sources (such as libraries). Taking the financial field as an example, the information sources for obtaining research reports can be Sina Finance, Financial websites such as Oriental Fortune and Flush, and the information source for the composition can be the library or composition resources on the Internet;

Dictionary acquisition steps: Perform data preprocessing and feature selection on multiple research reports obtained to build a research report dictionary. Data preprocessing includes word segmentation of the research reports, and then counting the number of times words appear in the research reports for feature selection. Decide whether to put words into the research report dictionary based on the number of occurrences;

Outline acquisition steps: Obtain the report outline corresponding to the event text based on the event text, research report dictionary and outline generation model. The outline generation model selects multiple words from the research report dictionary to form a word sequence as the report outline based on the probability optimal principle. The method of selecting words according to the optimal principle can be to use a global search method for word output with optimal probability, that is, only one word is output at a time, and multiple words form a word sequence as a report outline; it can also be to use a beam search algorithm to search for a preset number of words. The method of word output with optimal probability is to output a preset number of words at one time, and the words output multiple times are then selected according to the principle of optimal probability to form a word sequence as a report outline;

Report generation steps: Obtain a research report based on the event text, report outline, research report dictionary and report generation model. The report generation model selects multiple words from the research report dictionary to form a word sequence as a research report based on the principle of probability optimization, which is similar to outline generation. , you can use global search for word output with optimal probability, or you can use beam search algorithm to select words for output, which will not be described again.

The present invention constructs a research report dictionary by using multiple collected research reports, and then automatically outputs corresponding research reports based on event text, research report dictionary, outline generation model and report generation model, overcoming the problem in the prior art that manual writing of reports consumes a lot of energy. and technical issues of labor costs, liberating manpower, improving the output efficiency of research reports, and reducing labor costs. And the research reports generated based on the report outline are of higher quality. Furthermore, the overall framework of the research report generation method is the Encoder-Decoder (encoding-decoding) structure (the so-called encoding is to convert the input sequence into a fixed-length vector; decoding is to convert the previously generated fixed vector into an output sequence. .), that is, input a sequence and output a sequence. The process of Encoder is to turn the event text sequence into a fixed-length vector expression, and the Decoder process is to turn this fixed-length vector expression into a variable-length text sequence, that is, a research report.

The following takes research reports in the financial field as an example to explain the process of collecting research reports and constructing a dictionary of research reports:

First, crawl the text of research reports in the macroeconomic section of websites such as Sina Finance, Oriental Fortune, and Flush (a total of 110,000 pieces of data). Then use regular expressions to extract Chinese characters and Chinese punctuation marks in the research report. Since it is difficult to directly generate high-accuracy results with numerical symbols, they are discarded. Then Jieba word segmentation is used to divide the text data into individual words to obtain a word set. In order to improve the efficiency of calculation, words that appear more than 5 times in the word set are selected and put into the research report dictionary. In addition, in the dictionary acquisition step, add opening tags and ending tags to the research report dictionary. Specifically, the first four keys (keys can be understood as elements in the dictionary) of the research report dictionary are mask mark (mask), unknown mark (unk), start mark (start), and end mark (end). Among them, the mask mark (mask) is used to indicate that unnecessary information is masked. The unknown tag (unk) is used to indicate words in the word set that are not in the dictionary. They are usually words that do not appear frequently but have meaning, such as organization names and personal names. The start tag (start) and the end tag (end) are used to be added to the beginning and end of each piece of text data to indicate the beginning and end of the text data. The key of the dictionary is the word (containing a single punctuation mark), and the value of the dictionary is the sequence number of the word.

In addition, research report generation methods also include:

Event entity recognition is performed based on the research report obtained in the research report collection step to obtain the corresponding event text. Multiple research reports and corresponding event texts form a training data set. The training data set is used to train the outline generation model and the report generation model. Take research reports in the financial field as an example. After crawling to obtain macro research reports, rule matching and part-of-speech matching are used to process the macro research reports to obtain news text, and extract text containing event-related content as news, thereby turning the data into a A collection of one-to-one news and macro research reports is used as a training data set. Specifically, the research report can be cut into segments and saved as an array, and then the text in the array can be matched according to the set event rules to obtain event paragraphs as news text.

In this embodiment, two outline generation models are provided. For the first outline generation model, refer to Figure 2. Figure 2 is a method flow chart of the first specific embodiment of the research report generation method in the present invention; news in the financial field is used as Taking event text as an example, the first outline generation model includes:

According to the research report dictionary, vector representation is performed on the event text and the start tag (<start>) to obtain the event vector and the start tag vector; in this embodiment, a text can be converted into a fixed-length text through word embedding (word embedding vector) vector representation.

Obtain the hidden layer state of the event text according to the event vector and the bidirectional LSTM network, and obtain the hidden layer state of the beginning mark according to the beginning mark vector and the one-way LSTM network. Among them, the one-way LSTM network requires less processing time, but the two-way The LSTM network can better capture bidirectional semantic dependencies;

Obtain the report outline based on the hidden state of the event text, the hidden state of the beginning tag and the attention mechanism;

Then use a vector representation of the previous word output from the report generation step according to the research report dictionary to obtain a word vector. Word embedding can also be used to convert words into word vectors;

Obtain the hidden layer state of the word based on the word vector and the one-way LSTM network;

Update the report outline based on the hidden state of the word, the hidden state of the event text, and the attention mechanism.

In this embodiment, the report generation model includes a VAE generation model. In the first outline generation model, the probability distribution sequence of words can be obtained through the attention mechanism. The outline generation model then selects the best value in the research report dictionary based on the probability distribution information of the words. Combining vocabulary with optimal probability and outputting the vocabulary as a report outline. In the same way, in the report generation model, another probability distribution information can be obtained after processing by the VAE generation model. The report generation model then selects the word combination with the best probability in the research report dictionary based on the probability distribution information to output the research report. Specifically, referring to Figure 2, the beginning of the research report is marked with a <start> tag, and the end is marked with an <end> tag. Use the first outline generation model and VAE generation model to predict the probability distribution of the next word in the research report starting from the <start> tag, and then select the word output from the research report dictionary based on the probability distribution to generate a research report. To put it simply, when the research report generation method is executed, first, input the news and a starting <start> tag, and then output the first word of the research report after processing by the outline generation model and report generation model, and the first output word is returned Go to the input end and replace the beginning tag, then generate the second output, and then return the second output to the input end to replace the first output, and so on, output words one by one until the output reaches the end tag <end>, then stop, and finally The research report is generated. That is, the currently output word needs to depend on the last output word, and the report outline is updated based on the last output word and the input news, which effectively improves the quality of the generated research report. Among them, when the above-mentioned word output is selected from the research report dictionary according to the probability distribution to generate a research report, the word with the highest probability can be selected in the research report dictionary through global search for the optimal solution (that is, only one word is output for each output), but when the research report When the reporting dictionary is large, the space efficiency of the global search for the optimal solution is very low. Therefore, the beam search algorithm can be used to improve search efficiency. Beam search uses the beam size parameter to limit the number of possible words retained at each step. It not only considers the probability of a single word, but also considers the words before and after being put together. The probability. Therefore, after VAE generation model processing, the beam search algorithm is used to obtain a preset number of possible words with the highest probability. In this embodiment, beam size=3 is set, that is, the preset number is 3, and three maximum possible words are retained in each step. the result of. Each prediction can get 3 outputs, and then return these 3 outputs to the input terminal for the next prediction. When the prediction is completed, based on all the output words used to generate the research report, according to the principle of probability optimization, every 3 Select one of the outputs as the final output to obtain the final research report.

The outline generation model and VAE generation model need to be trained before formal use. Refer to Figure 3. Figure 3 is a schematic diagram of the training process in Figure 2. Using news in the financial field as event text, the following is the training process of the first outline generation model. Be explained:

First define the input news as X, x is the word in the news, the expression of the news is as follows,

X＝(x ₁ ,…,x _m ) (1)

The decoding of latent vectors has two stages. The first stage generates the outline O of the research report, o is the word in the research report outline, and the second stage generates the final research report Y, y is the word in the research report, and defines the generated text. The length is L, and the expression is as follows:

O＝(o ₁ ,…,o _L ) (2)

Y＝(y ₁ ,…,y _L ) (3)

Then we extracted Chinese and Chinese characters from the news and macro research reports of the training data set through regular matching. Then, after jieba word segmentation, we made a preliminary length statistics of the text in units of words. The final length statistics are shown in Table 1. :

Table 1 Statistics of length of news and research reports

Then, for training needs, the news and research reports are "truncated and lengthened" to the same set length. For example, the length of news is 30 words and the length of research reports is 200 words. The input news text and research report text are converted into vector representations through word embedding to obtain news vectors and research report vectors. Here, a layer of Embedding network is directly used. Since the LSTM network cannot encode information from back to front when modeling sentences, the bidirectional LSTM network can better capture bidirectional semantic dependencies. Then the news vector after Embedding is input into the bidirectional LSTM network. Among them, after the news vector is input into the bidirectional LSTM network, the hidden layer state of the news can be obtained as H, h is the hidden layer sub-state, t is time, and the expression of h is as follows:

In order to generate high-quality research reports, the model encoding and decoding processes need to fully absorb the structure and content of the corresponding research reports. The Decoder process predicts the probability distribution of the next word in the research report in the vocabulary through the LSTM network. The hidden state of the macro research report is represented by S. The hidden state of each time step depends on the input of the previous time and the hidden state of the previous time. The expression is as follows:

The hidden layer state of the news and the hidden layer state of the macro research report calculate the attention score through the attention mechanism (such as equation (6)). Here, multiplicative attention is used to accumulate the attention score, and the softmax function is used to calculate the attention weight ( Such as formula (7)). Through the weighted average of the attention weight and the hidden layer state of the news (such as equation (8)), the context vector is obtained by multiplying the attention weight and the hidden layer state of the news. The expressions of formula (6), formula (7) and formula (8) are as follows:

Then, the context vector and the hidden layer state of the macro research report are concatenated to update the attention hidden layer state. The predicted output of an input word will be calculated by the attention hidden layer state. The expression is as follows:

Among them, Wc is the model parameter.

The first stage objective function of decoding is as follows:

Among them, P is the probability.

At this point, the probability distribution of words in the research report dictionary is obtained based on the input news, and the outline of the report can be generated based on the probability distribution.

In the second stage of decoding, in order to generate the final research report, this stage requires the input of news and an outline generated based on the news.

The decoding model uses a variational autoencoder model (VAE), which generates target variables by fusing dual input news X and outline O, and learns the posterior probability distribution of the latent variable z. P(z|X) can be rewritten as the following expression:

Assuming that the posterior distribution is a standard normal distribution, randomly sample from the distribution and decode to the original text. By measuring the reconstruction loss and regularization loss for training, 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)

The right side of the above inequality is ELBO, where the first term is to sample the attention from P(z|X,O), use the sampled attention as the input of the decoder to calculate the cross-entropy loss, and the second term is through KL Divergence measures the similarity of two probability distributions, ensuring that the posterior distribution is close to the prior distribution.

The second stage objective function of decoding is as follows:

At this point, the probability distribution information of the word can be obtained again based on the news and report outline, and the final research report can be obtained based on the probability distribution information.

The global objective function is obtained by the sum of the loss functions of the two decoding stages, and the expression is as follows:

After inputting the one-to-one news and macro research report training data sets obtained above into the model in Figure 3 and training for 50 rounds, the final model parameters can be adjusted and determined. Specifically, taking the global optimal search as an example, only one word is output at a time when the research report is generated. Each macro research report is input into the model sequentially from the beginning tag - the macro research report - the end tag, and each word is input for training, for example By inputting the opening tag at the beginning, you can get an output at the model output. This output is the first word of the predicted research report. Compare the output word with the first word of the real research report, and modify the model based on the comparison results. parameters; then input the first word of the macro research report into the model and obtain an output. This output is the second word of the predicted research report, and compare the second word with the second word of the real macro research report. Compare the words, and then adjust the model parameters based on the comparison results to continuously reduce the error between the model output and the real words. After training the model with multiple news-macro research report data, the structure and parameters of the model will be saved after training. According to the model with finally determined model parameters, the newly input news can be predicted using the model in Figure 2. After the news is input, the research report in Figure 4 can be obtained. Figure 4 is a research report generated using the research report generation method in the present invention. Schematic diagram of the reported example.

Referring to Figure 5, Figure 5 is a method flow chart of a second specific embodiment of a research report generation method in the present invention; in Figure 5, the event text takes news in the financial field as an example, and the second outline generation model includes:

Get the report outline based on event text, opening tag and Transformer model;

Update the report outline based on the last word, event text, and Transformer model output by the report generation step. Among them, the report generation model also uses the VAE generation model. The Transformer model is equivalent to replacing the LSTM network and attention mechanism. The attention score of the Transformer model is solved by equation (15),

Among them, Q, K, V are the three matrix vectors that the input X is transformed into. Similar to Figure 2, the execution process of the model in Figure 5 is to input news and an opening <start> tag, and then output the first word of the research report after being processed by the outline generation model and the report generation model. This first output The word is returned to the input terminal to replace the beginning tag, and so on. The model in Figure 5 will output the reported words one by one. In addition, similarly, the outline generation model and report generation model can choose the method of global search for optimal solutions to select words, and the beam search algorithm can also be applied to the model in Figure 5 to improve search efficiency.

Refer to Figure 6, which is a schematic diagram of the training process in Figure 5; the training data set is input into the model in sequence for training. Similar to Figure 4, the macro research report is input into the model in the form of beginning mark - macro research report - end mark. Compare the model output with real research words to adjust model parameters.

Example 2

Embodiment 2 is provided based on Embodiment 1. Embodiment 2 provides a research report generating device. Refer to Figure 7. Figure 7 is a structural block diagram of an embodiment of the research report generating device in the present invention. The research report generating device includes:

The research report collection module is used to obtain multiple research reports from multiple information sources;

The dictionary acquisition module is used to perform data preprocessing and feature selection on multiple research reports to build a research report dictionary;

The outline acquisition module is used to obtain the report outline corresponding to the event text based on the event text, the research report dictionary and the outline generation model. The outline generation model selects multiple words from the research report dictionary to form a word sequence as the report outline based on the principle of probability optimization;

The report generation module is used to obtain research reports based on event text, report outline, research report dictionary and report generation model. The report generation model selects multiple words from the research report dictionary to form a word sequence as a research report based on the principle of probability optimization.

The specific working process description of the research report generating device may refer to the description in Embodiment 1, and will not be described again. The research report generation device can be used to automatically generate research reports, freeing up manpower and improving the efficiency of research report output.

Example 3

Embodiment 3 is provided based on Embodiment 1. Embodiment 3 provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions. The computer-executable instructions are used to cause the computer to execute the implementation. The research report generation method described in Example 1. For a specific description of the method for generating a research report, please refer to the description in Embodiment 1 and will not be described again.

The above is a detailed description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can also make various equivalent modifications or substitutions without violating the spirit of the present invention. , these equivalent modifications or substitutions are included in the scope defined by the claims of this application.

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.