CN114706972B - An automatic generation method of unsupervised scientific and technological information summaries based on multi-sentence compression - Google Patents

Abstract

The present invention relates to an unsupervised method for automatically generating a scientific and technological intelligence summary based on multi-sentence compression, and belongs to the technical field of natural language generation. For the generation of multi-document text in the field of scientific and technological intelligence, the source data is firstly obtained by a subject crawler based on the LDA subject similarity vocabulary expansion method. All text paragraphs are sorted by a text information value evaluation model based on three indicators: authority, timeliness, and content relevance of text information. The paragraphs with higher scores are selected as the original text for generating the final scientific and technological intelligence. Finally, an unsupervised multi-document summary method based on spectral clustering and multi-sentence compression is used to automatically generate a scientific and technological intelligence summary. This method effectively solves the problem that the generation of scientific and technological intelligence has high requirements for data timeliness and authority during data screening, and the problem that the traditional neural network-based multi-document generation method cannot be applied in the field of scientific and technological intelligence due to the lack of data sets.

Description

An automatic generation method of unsupervised scientific and technological information summaries based on multi-sentence compression

Technical Field

The invention relates to an unsupervised scientific and technological intelligence summary automatic generation method, in particular to an unsupervised scientific and technological intelligence summary automatic generation method based on multi-sentence compression, and belongs to the technical field of natural language generation.

Background technique

Scientific and technological intelligence work has played a key role in the formulation of national major scientific and technological strategies, the deployment of major scientific and technological plans, and economic and social development, and has contributed to the development of society, economy, and science and technology. It is a key component in the deployment of national scientific and technological plans and economic and social development.

In the field of scientific and technological intelligence, in the face of a big data environment, manually collecting, organizing, and screening valuable text data and manually writing intelligence reports consumes a lot of manpower and time costs. Therefore, people's current demand for intelligence is no longer satisfied with the orderly acquisition of information resources, and is no longer satisfied with the processing, organization, storage and access analysis with document units as the main feature. Instead, they have higher requirements for the depth of information analysis, including rapid evaluation and recommendation of data resources, extraction and analysis of knowledge units, multi-dimensional data fusion, fine-grained data analysis, and visualization and computerized data presentation and analysis, etc., striving to deduce and classify big data, remove the coarse and retain the fine, remove the false and retain the true, and realize basic automated intelligence summary generation.

However, in the era of information explosion, due to the complexity of the sources of scientific and technological intelligence information, how to quickly and accurately find the useful information you need from a large amount of information is a big challenge. To achieve basic automated intelligence generation, the first step is to efficiently collect effective information. In addition, since the timeliness and authority of intelligence are very important in intelligence research, they need to be considered when selecting literature. In addition, due to the different sources of information, the information structure is not unified, and it is also difficult to integrate and process multiple heterogeneous documents and generate a final report. In summary, in the process of realizing the automatic generation of scientific and technological intelligence summaries, the main problems that need to be solved are: comprehensive evaluation and recommendation of heterogeneous texts integrating factors such as time and multi-document summaries.

At present, the better methods for effective information collection include topic crawlers. Most researchers have adopted a combination of link-based and content-based crawling strategies, and have achieved good results. However, in the field of scientific and technological intelligence, the way to obtain information is usually authoritative think tanks at home and abroad, and there are fewer links in think tank web pages. Therefore, content-based crawling methods are more applicable in the intelligence field. In the research in the field of multi-document summarization, most of the latest results adopt the method of first sorting multiple documents, screening out the most important top N documents, and then using neural networks or neural networks combined with graph models. Some authors have also integrated pre-trained models such as pre-trained Bert into the model. The above methods have achieved good results in supervised multi-document summarization. However, in the field of scientific and technological intelligence, the lack of data sets is an issue that cannot be ignored, which makes supervised methods actually unusable in this field.

Summary of the invention

The purpose of the present invention is to solve the technical problems of the difficulty of manually collecting, screening and generating reports in the field of scientific and technological intelligence, and creatively propose an automated scientific and technological intelligence summary generation method that runs through data collection, data screening and intelligence generation. This method effectively solves the problem that in the process of data screening, scientific and technological intelligence generation has high requirements for data timeliness and authority, and the problem that the traditional neural network-based multi-document generation method cannot be applied in the field of scientific and technological intelligence due to the lack of data sets.

The innovation of the present invention is that: for the generation of multi-document text in the field of scientific and technological intelligence, firstly, the source data is obtained by the subject crawler based on the LDA (Latent Dirichlet Allocation, a document topic generation model, also known as a three-layer Bayesian probability model, including a three-layer structure of words, topics and documents) topic similarity vocabulary expansion method. All text paragraphs are sorted by a text information value evaluation model based on three indicators: authority, timeliness and content relevance of text information. The paragraphs with higher scores are selected as the original text for generating the final scientific and technological intelligence. Finally, an unsupervised multi-document summarization method based on spectral clustering and multi-sentence compression is used to automatically generate a summary of scientific and technological intelligence.

The present invention is achieved through the following technical solutions.

A method for automatically generating unsupervised scientific and technological intelligence summaries based on multi-sentence compression comprises the following steps:

Step 1: Use the topic crawler method based on the LDA topic similarity vocabulary expansion method to crawl text content and obtain source data.

Through the given initial keywords, when the topic description is insufficient, the topic crawler can collect topic-related resources by itself, continuously expand the corpus, cyclically train the model, and continuously improve, expand, and update the topic description, so as to obtain the desired content more comprehensively and accurately.

Step 2: Evaluate and rank the crawled texts based on the relevance of their content to keywords and the timeliness and authority of the source texts. Select the texts of paragraphs with scores ranking at least in the top 40 as the original texts for generating the final scientific and technological intelligence.

Step 3: Use the result text obtained in step 2 as the input of the model, and use the unsupervised multi-document summarization model based on spectral clustering and multi-sentence compression to obtain the summary result.

Beneficial Effects

Compared with the prior art, the method of the present invention has the following advantages:

1. This method proposes a paper patent text information evaluation model and a think tank article text information evaluation model. The model has strong versatility and can be applied to all paper patent texts and all think tank articles.

2. This method provides an automated scientific intelligence summary generation method from data acquisition to text generation. It uses a subject crawler to improve the relevance of data to subject keywords, reduce redundant data, and optimize the efficiency of data acquisition and cleaning stages. In the text generation stage, a combination of spectral clustering and multi-sentence compression is used to improve the effect of unsupervised multi-document summarization.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is an overall flow chart of the method of the present invention;

FIG2 is an architecture diagram of the subject crawler module of step 1 of the method of the present invention and embodiment 1;

FIG3 is a flow chart of step 2 of the method of the present invention and the text information value assessment process of Example 1;

FIG4 is a flowchart of step 3 of the method of the present invention and the multi-document summarization algorithm of embodiment 1;

FIG. 5 is a flow chart of the multi-sentence compression algorithm used in step 3.4 of the method of the present invention and the multi-document summarization process of Example 1.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

A method for automatically generating unsupervised scientific and technological intelligence summaries based on multi-sentence compression comprises the following steps:

Step 1: Use the topic crawler method based on the LDA topic similarity vocabulary expansion method to crawl text content and obtain source data.

Since only a small number of keywords are given, the content crawled by the crawler does not completely match the content that is actually desired to be crawled. Therefore, the use of thematic crawlers can improve the accuracy, expand the crawling scope, and maximize the efficiency of crawling.

Through the given initial keywords, when the topic description is insufficient, the topic crawler can collect topic-related resources by itself, continuously expand the corpus, cyclically train the model, and continuously improve, expand, and update the topic description in order to obtain the desired content more comprehensively and accurately.

Specifically, step 1 includes the following steps:

Step 1.1: Based on the given initial keywords, crawl the corresponding result web pages, extract summaries from these newly added web pages, and use them as new training corpus for LDA.

Step 1.2: Do word embedding for the training predictions. This can be done using the word2vec model.

Step 1.3: Combined with the original corpus, new topic documents are obtained through LDA training to cover and update the topic documents of the original topic crawler.

Step 2: Evaluate and sort the crawled text based on the relevance of its content to keywords and the timeliness and authority of the source text.

The evaluation of the value of text information usually starts from the analysis of the information's dissemination source, dissemination characteristics, content characteristics, etc. The information dissemination source reflects the characteristics of the information publisher, including the release channel, the authority of the publisher, etc. The information dissemination characteristics reflect the formal characteristics of the information dissemination process. Only information that has been widely, deeply and quickly disseminated can have its intrinsic value fully reflected, usually including the number of disseminators, the speed of dissemination, the depth of the dissemination chain, etc. In addition, information has obvious timeliness characteristics, and expired information often becomes worthless.

Therefore, a text information value assessment model is constructed by extracting the three characteristic dimensions of text information: authority, timeliness, and content relevance.

Specifically, step 2 includes the following steps:

Step 2.1: Segment all text into paragraphs. In subsequent calculations, the calculations are performed in paragraphs.

Among them, the methods for evaluating the value of papers, patents, and journals are as follows:

For papers, patents, and journal texts, the impact factor, the total number of publications and total downloads of the first author, the number of downloads of the text, and the number of citations are used as authority evaluation indicators. The publication time is used as a timeliness indicator. The similarity between the abstract and the subject vocabulary is used as a content relevance indicator. Corresponding parameters are set for each indicator, and a text information value assessment model is constructed to comprehensively calculate the value score of the text.

Furthermore, the present invention proposes a value scoring calculation method for papers, patents, and journal texts, comprising the following steps:

Step 1: Calculate the authority x ₁ .

For authority x ₁ , factors related to authority include the authority of the journal in which the text is published, the authority of the author in the field, and the evaluation of the text by other researchers in the field.

The authority of a journal x ₁₁ is expressed as the ratio of the journal's impact factor to the maximum impact factor of all literature, as shown in Formula 1:

The authority of papers and patents is determined by the number of articles published by the author as the first author in the field and the total number of downloads of the articles published by the author as the first author, as shown in Formula 2:

The value of the paper itself is determined by the number of downloads and citations of the paper, as shown in Formula 3:

Step 2: Calculate the timeliness x ₂ .

Assuming that the decay coefficient of the value of text information over time is μ, and the time interval between the time of information acquisition and the time of information release is Δt, the calculation of the change of information value over time is shown in Formula 4:

x ₂ = e ^{- μΔt} (4)

Among them, e is a natural constant.

Step 3: Calculate content relevance x ₃ .

Specifically, the BM25 algorithm can be used to calculate the relevance of text content. Each word in the subject word library obtained by the subject crawler is regarded as q _i . For the summary a of the text, the relevance score of each word q _i and a is calculated, and the relevance score of q _i and a is weighted summed to obtain the relevance score Score(Q,a) of the current text and the subject word library, as shown in Formula 5:

Where _Wi represents the weight of the i-th word _qi , calculated using the TF-IDF algorithm; n represents the total number of words in the vocabulary; R( _qi ,a) represents the correlation between words _qi and a, calculated using Equation 6 and Equation 7:

Among them, tf _ta is the word frequency of word t in a; _La is the length of a, _Lave is the average length of all texts, variable k is a positive parameter used to standardize the range of article word frequency; b is an adjustable parameter, 0<b<1, which means to use the document length to determine the range of information; K is the intermediate result of the calculation.

Step 2.3: Evaluate the value of think tank articles as follows:

For think tank articles, the number of followers and the number of posts of the article author are used as authority indicators, the publication time is used as the timeliness indicator, the similarity between the article abstract and the subject vocabulary is used as the content relevance indicator, and corresponding parameters are set for each indicator to construct a think tank article text information value assessment model.

The present invention proposes a value scoring calculation method for think tank article text, comprising the following steps:

Step 1: Calculate the authority x ₁ .

For think tank articles, since there is no data such as download volume and citation volume, and there is no quantitative measurement standard for the authority of think tanks, the number of fans and the number of posts of the author of the article are used as the measurement indicators of its authority. The specific calculation is based on formula 8 and formula 9:

Step 2: Calculate the timeliness x ₂ .

The calculation method is the same as that described in the second step of step 2.2.

Step 3: Calculate content relevance x ₃ .

The calculation method is the same as that described in the third step of step 2.2.

Step 2.4: Calculate the information value of the text.

The value of text information is defined as a linear combination of the new authority feature, timeliness feature, and content relevance feature. At the same time, considering the multiplier effect of timeliness, the information value is calculated as:

X＝[δ ₁ (α ₁ x ₁₁ +α ₂ x ₁₂ +α ₃ x ₁₃ )+δ ₂ (βx ₃ )]x ₂ (10)

Wherein, X represents the value of this text information, α ₁ , α ₂ , α ₃ , δ ₁ , δ ₂ represent the influence factors of different features on the text value, and their values are selected according to actual needs. In the present invention, α ₁ = α ₂ = 0.3, α ₃ = 0.4, δ ₁ = δ ₂ = 0.5 can be taken.

Step 2.5: Sort each paragraph according to its text information value score, and select the top 40 paragraphs in the sorting results as the text data for subsequent multi-document summarization.

Step 3: Use an unsupervised multi-document summarization model based on spectral clustering and multi-sentence compression to obtain the summary result.

Since there is no annotated dataset similar to the text of papers, patents or think tank articles in the field of multi-document summarization, this paper proposes an unsupervised machine learning method based on spectral clustering and multi-sentence compression for summary generation. This method converts the original document into a sentence graph, taking into account both language and depth representation, then applies spectral clustering to obtain multiple sentence clusters, and finally compresses each cluster to generate the final summary.

Specifically, step 3 includes the following steps:

Step 3.1: Process text data.

For the set of paragraphs P = {p ₁ , p ₂ , ... p _n } related to a topic finally obtained in step 2, the ultimate goal is to generate a summary S that encompasses the important information in the original document and has no redundant information.

Sentences are used as the smallest processing unit of text, and considering that sentence compression is required in the last step, all stop words are retained. The specific method is: generate a sentence list (which can be generated by calling SpaCy's NLP module. SpaCy is a leading industrial-grade processing library in the field of NLP tasks) and use it as the input of the sentence graph constructed later.

Step 3.2: Build a structured sentence graph, where the nodes correspond to the sentences generated in step 3.1, and draw edges based on the lexical relations and deep semantic relations between sentences.

The goal of this step is to identify paired sentence connections that can represent the discourse structure of the paragraph set P, and to construct a sentence graph based on the Approximate Discourse Graph (ADG) combined with deep embedding technology.

Specifically, a graph G = (V, E) is constructed, where the node _vi ∈ V of the graph represents a sentence, V represents a set of nodes, e _i,j ∈ E represents the edge between node _vi and node v _j , and E represents the set of edges. For any two different nodes _vi and node v _j , if the sentences they represent have the following relationship, they are connected to each other and there is an edge with a value of 1 between them, that is, e _i,j = 1.

The construction rules of graph G include:

De-verbalized noun association: According to English grammar, when an event or entity is mentioned in a verb phrase, it is usually represented as a subordinate noun or noun phrase of the modified verb in the following sentence. The noun form of the verb phrase is found through WordNet (an English dictionary based on cognitive linguistics, which arranges words in alphabetical order and forms a "network of words" according to the meaning of the words). If the noun form of the verb phrase in a sentence appears in the sentence after a certain sentence, the nodes represented by the two sentences are connected to each other.

Entity continuation: This takes into account the relevance of words. If sentence _vi and sentence _vj contain the same entity category (such as organization, person, product, etc.), the two nodes are connected to each other.

Discourse markers: If there is a semantic relationship between adjacent sentences, such as the presence of conjunctions such as however, meanwhile, furthermore, etc., the nodes represented by the two sentences are connected to each other.

Sentence similarity: The sentence is represented by averaging all word vectors of a sentence and calculating the similarity score of the sentence using the cosine similarity of the two sentence vectors. If the similarity score reaches the set threshold, the two nodes are considered to be connected to each other.

Step 3.3: Apply graph clustering to obtain the graph partitions.

At present, most graph clustering methods identify node groups in the graph based on the edges connecting the nodes. The present invention adopts a spectral clustering method, which is as follows:

Step 1: Get the Laplacian matrix of the sentence graph constructed in the above way (which can be obtained by subtracting the adjacency matrix from the degree matrix of the graph);

Step 2: Calculate the first m eigenvectors of the matrix to define the eigenvector of each sentence;

Step 3: Divide these sentences into m categories through k-means clustering.

Thus, we obtain m sentence categories representing different key information. Next, we perform multi-sentence compression operations on the sentence sets of m categories to obtain m summaries. The compression process is shown in step 3.4.

Step 3.4: Generate a summary from the extracted subgraph.

Multi-sentence Compression (MSC) is to generate a summary sentence from each cluster containing a group of semantically related sentences. Currently, the classic method is to build a word graph and select a sentence built by the shortest path as the summary.

The present invention proposes a new implementation method, which expands the classical method, as follows:

Step 1: Build a word graph.

For the sentence set S = {s ₁ ,s ₂ ,…,s _n }, firstly, each word appearing in the sentence is mapped to a node. Since polysemy is common in natural language, each node uses a bigram (token, tag) as its identifier, and each time a repeated word is considered, the word graph is adjusted according to the following rules:

For words that are not stop words, non-punctuation marks, and have no candidate nodes (there is no (token, tag) corresponding to the word in the current word graph), a new node is directly created.

For a word that is not a stop word, is not a punctuation mark, and has only one candidate node, the word is directly mapped to the candidate node.

For words that are not stop words, non-punctuation, and have multiple candidate nodes: map the word to the node closest to the context, but keep the word graph acyclic - that is, two identical words in the same sentence cannot be mapped to the same node. If there is no node that meets the conditions, create a new node.

For stop words and punctuation marks, if there is a node with the same context, it is mapped to that node, otherwise a new node is created.

For the weight of the edge between nodes, the co-occurrence probability between nodes is considered. The greater the co-occurrence probability of two nodes, the smaller the edge weight. When there is an edge between two nodes, if they have a multi-hop connection, the edge weight is enhanced. As the path length becomes longer, the enhancement effect of the multi-hop connection is weakened. Specifically, it is expressed by formula 11:

Where w(e _i,j ) represents the weight of the edge between node i and node j; freq(i) and freq(j) represent the number of words mapped to node i and node j respectively; diff(s,i,j) refers to the distance between the offset positions of the word mapped to node i and the word mapped to node j in sentence s;

Step 2: Recall phase: Find F shortest paths in the above word graph, and each sentence formed by the path is a candidate answer.

The essence of this step is to solve the limited F shortest path problem. In the present invention, Yen's algorithm is used to solve this problem. The algorithm is divided into two parts. The first shortest path P(1) is calculated, and then the other F-1 shortest paths are calculated in sequence on this basis. When calculating P(i+1), all nodes on P(i) except the terminal node are regarded as deviation nodes, and the shortest path from each deviation node to the terminal node is calculated, and then spliced with the path from the starting node to the deviation node on the previous P(i) to form a candidate path, and then the shortest deviation path is obtained. The top 100 paths are selected as candidate sentence paths.

Step 3: Reorder the candidate answers above and select the candidate answer with the highest ranking as the final answer.

Specifically, TextRank is used to extract key phrases and new scores are designed for re-ranking. First, each node updates its score using Equation 12 until convergence:

Where S( _ni ) represents the score of node n _i in the word graph. The damping coefficient d can be 0.85. adj( _ni ) represents the nodes adjacent to node n _i , and w(e _j,i ) represents the weight of the edge between node n _j and node n _i .

Then, the key phrase r is obtained based on the keyword combination, and its score score(r) is as follows:

Among them, TextRank(w) represents the score of word node w calculated by the TextRank algorithm. The denominator is the weighted length length(r) of the key phrase r. The score is normalized to tend to select longer phrases.

Finally, the paths are re-ranked by multiplying the weighted length of the total paths in the candidate sentences obtained in the second step by the sum of the key phrase scores they contain. Based on the key phrase scores, the final score of each sentence is calculated:

Among them, length(c) represents the weighted length of sentence c, and path(c) represents the complete path of sentence c.

The one with the smallest score is selected as the generated summary, and finally the summaries generated by m categories are connected to obtain the final complete summary.

Example

This embodiment describes a specific embodiment of the method of the present invention.

The implementation diagram is shown in the overall process of Figure 1. The present invention provides a complete process from text data acquisition to data processing and then to summary text generation in the process of generating scientific and technological intelligence summaries. When the present invention is implemented, first the subject crawler module starts working, and obtains the data required for analysis according to the keyword library provided by the user, then the text information value assessment module analyzes and sorts the obtained data, and finally the sorting result is used as the input of the summary generation module and brought into the model to obtain the final result.

First, according to the keywords provided by the user, the subject crawler module is used to obtain data in Google Scholar, DARPA, IARPA, and RAND Corporation. Figure 2 is a process for obtaining data in an unsupervised scientific and technological intelligence automatic generation method based on sentence graph clustering. According to step 1 introduced in the present invention, a certain number of web pages are crawled according to the given initial keywords, and then the abstracts of these newly added web pages are extracted as new training corpus for LDA, and then word2vec is used to perform word embedding on the training predictions. Finally, combined with the original corpus, a new subject document is obtained through LDA training to cover and update the subject document of the original subject crawler.

After obtaining the required text data, the text value is evaluated based on the text attribute data. The evaluation method is shown in Figure 3. First, all texts are divided into paragraphs, and then the authority, timeliness and content relevance of the text data are calculated based on the journal, author, download volume and other data of the text data. Then, the value of the text information is calculated by combining the authority, timeliness and content relevance. Finally, the text data is sorted according to the value of the text information, and the top 40 texts are selected as the text data for subsequent multi-document summarization.

Finally, the text generation stage. The specific process is shown in Figure 4. First, the text data is processed, and the paragraph data obtained in the previous step is divided into sentences. Then, the NLP module of the SpaCy library is called to generate a sentence list. Then, an undirected sentence graph is constructed according to the rules described in step 3.2. Then, the sentence graph is clustered according to the spectral clustering method described in step 3.3 to generate m classes. Finally, a summary is generated for the m classes using multi-sentence compression. The process of multi-sentence compression is shown in Figure 5. First, a word graph is constructed according to the rules described in the first step of step 3.4. Then, the 100 shortest paths in the graph are obtained using the Yen’s algorithm, and finally, the reordering is performed. The reordering process is: first, the key phrases are extracted using the TextRank algorithm, then the scores of the sentences are recalculated based on the key phrases, and finally, the scores of the 100 paths are sorted. The sentence connected by the words in the path with the smallest score is the summary result of the class. Finally, the summaries generated in the m classes are connected to generate the final summary.

The above is only a preferred embodiment of the present invention, and the present invention should not be limited to the contents disclosed in the embodiment and the drawings. Any equivalent or modification completed without departing from the spirit disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims (2)

1. An automatic generation method of an unsupervised scientific and technological information abstract based on multi-sentence compression is characterized by comprising the following steps:

Step 1: performing text content capturing by adopting a theme crawler mode based on an LDA theme similarity lexicon expansion method to acquire source data;

Step 2: evaluating and sorting the crawled texts according to the correlation between the content and the keywords and the timeliness and authority of the source texts; the method comprises the steps of constructing a text information value evaluation model by extracting three characteristic dimensions of authority, timeliness and content relativity of text information;

The method comprises the following steps:

step 2.1: segmenting all texts according to paragraphs; in the subsequent calculation, the calculation is performed in the unit of a paragraph;

The method for evaluating the value of papers, patents and journals is as follows:

Aiming at papers, patents and journals, taking an influence factor, the total text sending amount and total downloading amount of a first author, the text downloading amount and the quotation amount as authoritative judgment indexes, taking release time as timeliness indexes, taking similarity of abstracts and a subject word stock as content correlation indexes, setting corresponding parameters for each index, constructing a text information value evaluation model, and comprehensively calculating the value score of the text;

The value score calculating method for papers, patents and journals comprises the following steps:

the first step: computing authority x ₁;

For authority x ₁, the factors related to authority include the authoritative of the published journal of the text, the authoritative of the author in the art, and the evaluation of the text by other researchers in the art;

The authority x ₁₁ of the journal class is expressed by the ratio of the journal influence factor to the maximum value of all document influence factors, and the ratio is expressed by the following formula: The authoritativeness of papers and patents is determined by the number of articles published in the field by the author as the first author and the total amount of articles downloaded by the author as the first author, and the formula is as follows: /(I)

The value of the paper itself is determined by the download and citation of the paper, as shown in the following formula:

and a second step of: calculating timeliness x ₂;

Setting the attenuation coefficient of the text information value with time as mu, and setting the time interval between the information acquisition time and the information release time as deltat, and calculating the change of the information value with time as follows: x ₂＝e^-μΔt, wherein e is a natural constant;

and a third step of: calculating content correlation x ₃;

Each word in the topic word library acquired by the topic crawler is seen as q _i; for abstract a of the text, calculating the relevance Score of each word Q _i and a, weighting and summing the relevance scores of Q _i and a to obtain a relevance Score (Q, a) of the current text and the topic word stock, Wherein W _i represents the weight of the ith word q _i, calculated using the TF-IDF algorithm; n represents the total number of words in the word stock; r (q _i, a) represents the correlation of the word q _i with a, calculated by: /(I) Where tf _ta is the word frequency of word t in a; l _a is the length of a, L _ave is the average length of all texts, and the variable k is a positive parameter used for normalizing the range of article word frequencies; b is an adjustable parameter, 0< b <1, representing the range of information quantity represented by the decision to use the document length; k is an intermediate result in calculation;

the value score calculating method for the intelligent library article text comprises the following steps:

the first step: computing authority x ₁;

for the intelligent library articles, the number of vermicelli and the number of letters of the authors of the articles are used as the measurement index of authority of the intelligent library articles, and the following formula is adopted for calculation:

and a second step of: calculating timeliness x ₂;

The calculation method is the same as the second step of the value score calculation method of papers, patents and journal texts;

and a third step of: calculating content correlation x ₃;

the calculation method is the same as the third step of the value score calculation method of papers, patents and journal texts;

Step 2.3: evaluating the value of the intelligent library article; aiming at the intelligent library article text, using the number of vermicelli and the number of text of an article author as authority indexes, using the release time as timeliness indexes, using the similarity of an article abstract and a theme word stock as content correlation indexes, setting corresponding parameters for each index, and constructing an intelligent library article text information value evaluation model;

Step 2.4: calculating the information value of the text;

defining the text information value as a linear combination of new authoritative features, timeliness features and content relevance features; meanwhile, considering the time-lapse multiplier effect, obtaining a measured information value ：X＝[δ₁(α₁x₁₁+α₂x₁₂+α₃x₁₃)+δ₂(βx₃)]x₂,, wherein X represents the value of the text information, alpha ₁、α₂、α₃、δ₁、δ₂ represents the influence factors of different characteristics on the text value, and the value is selected according to actual needs;

Step 2.5: sorting each paragraph according to the text information value scores, and selecting a plurality of top 40 paragraphs of the sorting result as text data for carrying out multi-document summarization subsequently;

Step 3: taking the result text obtained in the step 2 as the input of a model, and adopting an unsupervised multi-document abstract model based on spectral clustering and multi-sentence compression to obtain an abstract result;

firstly, converting an original document into sentence patterns, simultaneously considering language and depth representation, then obtaining a plurality of sentence clusters by using spectral clustering, and finally compressing each cluster to generate a final abstract;

step 3.1: processing the text data;

For the paragraph set P= { P ₁,p₂,…p_n } which is finally obtained in the step 2 and is related to a theme, the final goal is to generate a summary S which contains important information in the original document and has no redundant information; taking sentences as the minimum processing unit of the text, considering that sentence compression is required in the last step, and reserving all stop words; the specific method comprises the following steps: generating a sentence list and taking the sentence list as the input of a sentence graph constructed later;

Step 3.2: establishing a structured sentence graph, wherein nodes correspond to sentences generated in the step 3.1, drawing edges according to vocabulary relations and deep semantic relations among the sentences, thereby identifying paired sentence connection capable of representing the speech structure of the paragraph set P, and constructing the sentence graph by adopting a technology based on approximate language graphs and combining depth embedding;

Constructing a graph G= (V, E), wherein a node V _i epsilon V of the graph represents a sentence, V represents a set of nodes, E _i,j epsilon E represents an edge between a node V _i and a node V _j, and E represents a set of edges; for any two different nodes v _i and v _j, if the sentences they represent have the following relationship, they are connected to each other and there is an edge with a value of 1 between them, i.e. e _i,j =1;

The graph G construction rule includes:

verbalized noun association: according to the english grammar, when an event or entity is mentioned in a verb phrase, the event or entity will typically be represented in the next sentence as a subordinate noun or noun phrase of the modified verb; finding noun forms of the verb phrase through WordNet; if noun forms of verb phrases in a sentence appear in the sentence after the sentence, nodes represented by the two sentences are connected with each other;

Entity continuation: this term considers the relevance on the word; if sentence v _i and sentence v _j and contain the same entity class, then the two nodes are connected to each other;

utterance markers: if there is a semantic relationship between adjacent sentences, the nodes represented by the two sentences are connected with each other;

Sentence similarity: calculating a similarity score of the sentence by averaging all word vectors of one sentence as sentence representations and using cosine similarity of the two sentence vectors; if the similarity score reaches a set threshold, judging that the two nodes are connected with each other;

step 3.3: and clustering the graph to obtain the intra-graph partition, wherein the method comprises the following steps of:

the first step: obtaining a Laplacian matrix of the constructed sentence graph;

and a second step of: calculating the first m eigenvectors of the matrix to define eigenvectors of each sentence;

And a third step of: dividing the sentences into m categories by a k-means clustering mode;

Obtaining m sentence categories representing different key information, and then respectively carrying out multi-sentence compression operation on the sentence sets of the m categories to obtain m abstracts;

step 3.4: generating a summary from the extracted subgraph;

The method for generating the abstract comprises the following steps:

The first step: constructing a word graph;

For the sentence set s= { S ₁,s₂,…,s_n }, first map to a node for each word that appears in the sentence; since the case of word ambiguity in natural language exists widely, each node uses a tuple (tag) as its identity, and each time a recurring word is considered, the word graph is adjusted according to the following rules:

For non-stop words and non-punctuation words without candidate nodes, directly establishing a new node;

For non-stop words and non-punctuation, only one word of the candidate node is used, and the word is directly mapped to the candidate node;

For non-stop words, non-punctuation, and words with multiple candidate nodes: mapping the word to the node closest to the context, but keeping the word graph loop-i.e. two identical words of the same sentence cannot be mapped to one node; if no node meeting the condition exists, a node is newly built;

For stop words and punctuations, if nodes with the same context exist, mapping the stop words and the punctuations into the nodes, otherwise, creating a node;

For the weight of the edge between the nodes, considering the co-occurrence probability between the nodes, the larger the co-occurrence probability of the two nodes is, the smaller the edge weight is, when the edge exists between the two nodes, if the edge exists between the two nodes, the edge weight is enhanced, and as the path length is longer, the multi-hop connection enhancement effect is weakened, and the following expression is specifically adopted:

Wherein w (e _i,j) represents the weight of the edge between node i and node j; freq (i) and freq (j) respectively represent the number of words mapped to the node i and the node j; diff (s, i, j) refers to the distance between the offset positions of the word mapped to node i and the word mapped to node j in sentence s;

and a second step of: a recall stage; f shortest paths are found in the word graph, and sentences formed by each path are candidate answers;

Solving a problem by adopting a Yen's algorithm; the algorithm is divided into two parts, the 1 st shortest path P (1) is calculated, and then other F-1 shortest paths are calculated in sequence on the basis; when solving P (i+1), regarding all nodes except for the termination node on P (i) as offset nodes, calculating the shortest path from each offset node to the termination node, and splicing the shortest path from the initial node to the offset node on the previous P (i) to form a candidate path, so as to obtain the shortest offset path; selecting the top-ranked 100 paths as candidate sentence paths;

And a third step of: re-ordering the candidate answers, and selecting the candidate answer with the highest ranking as the final answer; extracting key phrases by using TextRank, and designing new scores for reordering; first, each node updates its score using the following formula until convergence: wherein S (n _i) represents the score of node n _i in the word graph; d is a damping coefficient, and 0.85 is taken; adj (n _i) represents a node adjacent to node n _i, and w (e _j,i) represents the weight of the edge between node n _j and node n _i;

then, the key phrase r is obtained according to the key word combination, and the score of the key phrase r is obtained Wherein TextRank (w) represents the score of word node w calculated via TextRank algorithm; the denominator is the weighted length (r) of the key phrase r, and the normalization of the score is done to tend to select longer phrases;

Finally, re-ordering the paths by multiplying the weighted length of the total paths in the candidate sentences obtained in the second step by the sum of the scores of the key phrases contained in the paths; calculating a final score for each sentence based on the key phrase scores: Where length (c) represents the weighted length of sentence c and path (c) represents the complete path of sentence c;

and selecting the abstract with the minimum score as the generated abstract, and finally connecting the abstracts generated by the m categories to obtain the final complete abstract.

2. The automatic generation method of the unsupervised scientific and technological information abstract based on multi-sentence compression as claimed in claim 1, wherein the step 1 comprises the following steps:

step 1.1: crawling corresponding result webpages according to given initial keywords, and extracting abstracts from the newly added webpages to serve as LDA new training corpus;

step 1.2: word embedding is carried out on training expectation;

Step 1.3: and combining the original corpus, and obtaining a new topic document through LDA training, wherein the new topic document is used for overlaying and updating the topic document of the original topic crawler.