An english news element extraction method based on core word diffusion

Abstract

The application provides a WHO element extraction method in an English news scene, which comprises the following steps: cleaning and preprocessing data of network news, combining an article title, a lead and an article theme into one text data, and analyzing a tree structure of the text data; extracting all nouns in all noun phrases from a subtree of the text data tree structure, screening core words, recalling a core word set, calculating a characteristic value of each core word, and extracting a first core word in a sorting result as a final core word; expanding the core word into a complete element text, traversing leaf nodes corresponding to the core word, and finding simple sentence nodes in a parent node direction, connecting all texts from the simple sentence nodes to the leaf nodes and then to intermediate links, and finally generating the element text; and the application realizes extraction of WHO elements on multiple news samples, has high accuracy, can improve a recall rate of core words, and can expand a recall set.

Description

Technical Field

[0001] This invention relates to the fields of computer technology and natural language processing, specifically to a method for extracting English news elements based on core word diffusion. Background Technology

[0002] In today's information-saturated world, news reporting plays a crucial role, providing the public with a window into the world. However, a comprehensive understanding of a news event cannot be achieved solely through headlines or brief reports. To meet the public's demand for accurate and complete news reporting, journalists and news professionals have adopted a concise yet comprehensive method—the "5W1H" approach: What, Why, When, Where, Who, and How.

[0003] For researchers in the field of journalism, analyzing large amounts of news data on a particular event is essential, and extracting the 5W1H elements is a common method. Traditional manual extraction methods are inefficient and cannot meet the quantitative analysis requirements of researchers in the era of big data. Therefore, a computer-aided news element extraction method is needed to assist researchers in processing large amounts of news data.

[0004] News element extraction is also a type of question-answering system. The 5W process for news text typically involves three steps: preprocessing, candidate phrase extraction, and candidate scoring. The system's input data is usually text, such as headlines, leads, and body text. Other systems use Automatic Speech Recognition (ASR) to convert broadcast speech into text before inputting it into the system. This process yields five phrases, each corresponding to one of the 5Ws, which together represent the main events of the given news text. Preprocessing tasks include sentence segmentation, sentence tokenization, and word tagging using NLP, including part-of-speech tagging, coreference resolution, NER, or Semantic Role Tagging (SRL).

[0005] Existing technology processes each sentence from the article, with each sentence processed according to the following process:

[0006] Recall: Traverse the reference cluster set returned by coreNLP, extract the first noun phrase NP. If its right sibling branch is a verb phrase VP and its direct parent node is S, then the word corresponding to the node can be added to the recall candidate. If the node contains a subtree of verb phrase VP, then discard the node.

[0007] Sorting: Based on the characteristics of the recall candidates, the number of times the recall candidate word is mentioned in the reference resolution set, the position of the recall candidate word in the whole article, and whether the recall candidate word is a named entity are calculated respectively. Then, the numbers obtained from the three parts are multiplied by 0.095, 0.9, and 0.005 respectively and weighted and summed.

[0008] The recall of existing technologies relies on the set of terminology resolution returned by coreNLP. However, coreNLP is a model-based algorithm that does not return terminology resolution results for all news articles. This can lead to a lack of recall in existing technologies. In existing technologies, recall methods usually impose a lot of restrictions to avoid irrelevant candidate words from interfering with the ranking. This can result in a small recall set, making it impossible to extract the correct WHO elements from the articles.

[0009] Since identifying the subject of analysis is a very important issue for researchers in journalism, this invention focuses on answering the question of "who" in news. Summary of the Invention

[0010] In view of one or more technical defects in the prior art, the present invention proposes the following technical solution.

[0011] A method for extracting English news elements based on core word diffusion, characterized in that the method includes:

[0012] S1: The recall step is to select a subset that meets the core word requirements from the text data of online news. Specifically, it is to extract all nouns in all noun phrases corresponding to the WHO elements in the text data to obtain a core word candidate set.

[0013] S2: Sorting step, which is to form an initial matrix of candidate core words using the core word candidate set mentioned in S1, calculate the feature value of each candidate core word in the initial matrix of candidate core words and perform weighted summation, sort the candidate core words from largest to smallest according to the result of weighted summation, take the first core word of the sorting result as the last core word according to the parameter requirements and output the sorted candidate core word matrix.

[0014] S3: The text element generation step expands the last core word mentioned in S2 into a complete element text. Specifically, the core word text is expanded according to the tree structure of the text data mentioned in S1. After finding the leaf node corresponding to the last core word mentioned in S2 in the tree structure of the text data, a simple sentence node is found in the direction of the parent node of the leaf node. Then, the simple sentence node is connected to the leaf node, and all the texts passed through the intermediate links are connected to form the final WHO element text.

[0015] Furthermore, step S1 also includes the following steps prior to recall: cleaning the online news in S1, and combining the article title, lead, and article topic of the online news into the text data.

[0016] The technical advantage of this invention is that data cleaning before performing step S1 can further improve the accuracy of extracting WHO elements, ensure data integrity, and improve the reliability of subsequent steps.

[0017] Furthermore, the calculation of the feature value described in S2 includes calculating the positional bias value, candidate word frequency, word2vec value, and Jaccard coefficient of the article title and the sentence containing the candidate word. Finally, it is determined whether the candidate core word is a named entity. If the candidate word is a named entity word, the feature value is set to 1; otherwise, the feature value is set to 0.

[0018] Furthermore, the positional bias value is calculated based on the sentence position and the position of the candidate core word within the entire text, and the calculation method is as follows:

[0019] ;

[0020] in, It is the position offset value. These are weighting coefficients. The total number of words in the sentence containing the candidate core words. This is the total number of sentences. It is the index of the specific position of the core candidate word in the sentence. It is the subscript of the sentence containing the core candidate word.

[0021] The positional bias of core words can more intuitively describe the positional factors of candidate core words in the data. Generally speaking, the earlier a candidate core word appears, the more likely it is to become the final core word.

[0022] Furthermore, the candidate word frequency is obtained through the following steps: statistically analyzing the word frequency of the candidate core word in the candidate core word set and normalizing the statistical value of the word frequency.

[0023] Furthermore, the word2vec value is obtained through the following steps: loading a word2vec model to predict the word embedding sequence of the article title and the word embedding sequence of the sentence containing the candidate core words, and then calculating the cosine similarity between the two word embedding sequences.

[0024] Converting English words into vectors allows for more generalized analysis of words and sentences in English text, enabling further analysis of the relevance between candidate keywords and article titles.

[0025] Furthermore, the Jaccard coefficient is obtained through the following steps: comparing the similarity and difference between the set of nouns contained in the article title and the set of all candidate core words in the sentence containing the candidate core words.

[0026] By analyzing the similarity between candidate keywords and article title nouns using the Jaccard coefficient, we can more intuitively see which candidate keyword is more likely to become the final keyword.

[0027] The present invention also proposes a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the method as described in any of the preceding claims.

[0028] The technical advantages of this invention are: it solves the problems of small recall sets and low recall rates in traditional news element extraction methods, and it filters out the core words of the article from the recall set, expands the core words into phrases, realizes the extraction of WHO elements from English articles, and improves the accuracy of WHO element extraction. Attached Figure Description

[0029] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0030] Figure 1 This is the overall process of a WHO element extraction method in an English news scenario provided by an embodiment of the present invention.

[0031] Figure 2 This is a tree structure diagram derived from English sentence parsing, based on a specific embodiment of the present invention.

[0032] Figure 3 This is a flowchart of the recall steps of a WHO element extraction method in an English news scenario provided by an embodiment of the present invention.

[0033] Figure 4 This is a flowchart of the sorting steps for a WHO element extraction method in an English news scenario provided by an embodiment of the present invention.

[0034] Figure 5 This is a flowchart illustrating the steps of generating text elements in a WHO element extraction method for an English news scenario, provided by an embodiment of the present invention. Detailed Implementation

[0035] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0036] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0037] Figure 1 The present invention illustrates the overall flowchart of a WHO element extraction method for English news scenarios.

[0038] Including the following methods:

[0039] S1: The recall step is to select a subset that meets the core word requirements from the text data of online news. Specifically, it is to extract all nouns in all noun phrases corresponding to the WHO elements in the text data to obtain a core word candidate set.

[0040] S2: Sorting step, which is to form an initial matrix of candidate core words using the core word candidate set mentioned in S1, calculate the feature value of each candidate core word in the initial matrix of candidate core words and perform weighted summation, sort the candidate core words from largest to smallest according to the result of weighted summation, take the first core word of the sorting result as the last core word according to the parameter requirements and output the sorted candidate core word matrix.

[0041] S3: The text element generation step expands the last core word mentioned in S2 into a complete element text. Specifically, the core word text is expanded according to the tree structure of the text data mentioned in S1. After finding the leaf node corresponding to the last core word mentioned in S2 in the tree structure of the text data, a simple sentence node is found in the direction of the parent node of the leaf node. Then, the simple sentence node is connected to the leaf node, and all the texts passed through the intermediate links are connected to form the final WHO element text.

[0042] It should be noted that step S1 also includes the following steps prior to the recall: cleaning the online news in S1, and combining the article title, lead, and article theme of the online news into the text data.

[0043] It should be noted that the cleaning of the text data of the aforementioned online news includes, but is not limited to, filtering of dirty characters and removal of advertising text.

[0044] It should be noted that after the text data is cleaned, the tree structure of the text needs to be extracted using CoreNLP and the words need to be tagged with part-of-speech tags. The text data is then processed and merged.

[0045] It should be noted that coreNLP is an NLP toolkit from Stanford University, which integrates practical functions including word segmentation, part-of-speech tagging, and syntactic analysis.

[0046] In a specific exemplary embodiment, for example, the sentence "The Yankees went to great lengths to incorporate many features of their old stadium (bottom) into their new one (top)," is extracted and parsed using Core NLP to obtain its tree structure, as follows: Figure 2 As shown:

[0047] ROOT is the root node, S is a child node of ROOT, "(NP (DT The) (NNPS Yankees))" is a child node of S, and (VP (VBD went) is a child node of S and a sibling node of (NP (DT The) (NNPS Yankees)). Here, NP represents a noun phrase and VP represents a verb phrase.

[0048] The tree structure represents the hierarchical relationship of statements. For example, in the exemplary sentence, "The Yankees" is the subject, "went" is the predicate, "The Yankees" and "went" are at the same level, and "great lengths" modifies "went" and is a child node of "went".

[0049] It should be noted that each sentence contains tokens data corresponding to the number of words. Each token is an annotation of a word, including the position information of the index word, the original text information of the original text word, and the part-of-speech information of the pos word.

[0050] It should be noted that the initial matrix of candidate core words described in S2 is used to store various feature values ​​of candidate words.

[0051] It should be noted that the calculation of the feature value includes calculating the positional bias value, candidate word frequency, word2vec value, and Jaccard coefficient of the article title and the sentence containing the candidate word. Finally, it is determined whether the candidate core word is a named entity. If the candidate word is a named entity word, the feature value is set to 1; otherwise, the feature value is set to 0.

[0052] It should be noted that the positional bias value is calculated based on the candidate word's position within the sentence and its position in the entire text. The calculation method is as follows:

[0053] ;

[0054] in, It is the position offset value. These are weighting coefficients. The total number of words in the sentence containing the candidate core words. This is the total number of sentences. It is the index of the specific position of the core candidate word in the sentence. It is the subscript of the sentence containing the core candidate word.

[0055] It should be noted that the positional bias value is used to describe the positional factors of the candidate core words. The earlier the candidate core word appears, the higher the probability that it will be the final core word.

[0056] It should be noted that the word frequencies of the candidate words are obtained through the following steps: counting the word frequencies of the candidate core words in the candidate core word set and normalizing the statistical values ​​of the word frequencies.

[0057] It should be noted that the word2vec value is obtained through the following steps: loading the word2vec model to predict the word embedding sequence of the article title and the word embedding sequence of the sentence containing the candidate core words, and then calculating the cosine similarity between the two word embedding sequences.

[0058] It should be noted that word2vec is a relevant model for generating word vectors. It is mainly used to embed a high-dimensional space with the number of all words into a continuous vector space with a much lower dimension. Each word or phrase is mapped to a vector in the real number field to achieve word vectorization, which is used to measure the relevance between candidate core words and article titles.

[0059] It should be noted that the article title contains important information from the English news text. The higher the relevance between the sentence and the article title, the higher the probability that the sentence is a key sentence.

[0060] It should be noted that the Jaccard coefficient is obtained through the following steps: comparing the similarity and difference between the set of nouns contained in the article title and the set of all candidate core words in the sentence containing the candidate core words.

[0061] It should be noted that the Jaccard coefficient is the ratio of the size of the intersection between two sets to the size of the size of the union between two sets. It is used to compare the similarity and differences between finite sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.

[0062] Figure 3 A flowchart illustrating the recall steps of a WHO element extraction method in an English news context, according to an embodiment of the present invention, is shown. It includes:

[0063] S3-1: Parse the tree structure data of English news articles from their text data;

[0064] S3-2: Traverse all subtrees of the root node and determine whether the subtree is a noun phrase NP. If the subtree is a noun phrase NP, further determine whether the parent node of the subtree is a simple sentence S. If the subtree is not a noun phrase NP, continue traversing all subtrees of the root node.

[0065] S3-3: Determine whether the parent node of the noun phrase NP subtree described in S3-2 is a simple sentence S. If the parent node of the noun phrase NP subtree is a simple sentence S, then extract all nouns from the subtree of the noun phrase NP. Otherwise, continue traversing all subtrees of the root node until all subtrees of the root node have been traversed, and output all candidate nouns.

[0066] In one exemplary embodiment, the English text “The Yankees went to great lengths to incorporate many features of their old stadium (bottom) into their new one (top).” is parsed to extract the following: Figure 2 The tree structure data of the article shown is obtained by traversing all subtrees of the root node ROOT whose labels are noun phrases NP, finding the subtrees whose parent node is the noun phrase NP of the simple sentence S, and obtaining the candidate core word set {“TheYankees”}.

[0067] It should be noted that the goal of the recall step is to extract a subset of English news articles that meet the core word requirements.

[0068] It should be noted that WHO elements generally correspond to noun phrases, so it is necessary to extract all nouns from all noun phrases in the subtree.

[0069] It should be noted that the noun phrase must be a direct child node of a simple sentence. Specifically, in terms of data structure, the label of the subtree corresponding to the noun phrase is equal to NP, and the label of the direct parent node of the noun phrase is equal to S.

[0070] In an exemplary embodiment, the noun "The Yankees" is a direct child node of the simple sentence "S". The label of the subtree corresponding to the noun "The Yankees" is equal to NP, and the label of its direct parent node is equal to S. Therefore, the noun "Yankees" is recalled to the candidate core word set.

[0071] In specific embodiments, the recall step of the present invention solves the problem in the prior art that relies on the reference resolution set returned by corenlp and results in no recall. At the same time, it also further solves the problem in the prior art that the recall set is too small and the recall rate is too low due to the large number of restrictions imposed to avoid irrelevant candidate words from interfering with the ranking.

[0072] Figure 4 A flowchart illustrating the sorting steps of a WHO element extraction method in an English news context, according to an embodiment of the present invention, is shown. It includes:

[0073] S4-1: Input the candidate keywords and extract the article title;

[0074] S4-2: Generate the initial matrix of candidate core words;

[0075] S4-3: Traverse each candidate core word;

[0076] S4-4: Calculate the feature values ​​of the candidate core words mentioned in S4-3;

[0077] S4-5: Perform a weighted summation on the eigenvalues ​​described in S4-4;

[0078] S4-6: Sort the top N core words according to the total score and take them as the final core words, and output the final core word matrix.

[0079] It should be noted that in S4-5, N=1, meaning that the core word with the highest total score in the ranking results is taken as the final core word.

[0080] It should be noted that the initial matrix of candidate core words described in S4-2 is used to store various feature values ​​of candidate words.

[0081] It should be noted that the calculation of the feature value described in S4-3 includes calculating the positional bias value, candidate word frequency, word2vec value, and Jaccard coefficient of the article title and the sentence containing the candidate word. Finally, it is determined whether the candidate core word is a named entity. If the candidate word is a named entity word, the feature value is set to 1; otherwise, the feature value is set to 0.

[0082] It should be noted that the positional bias value is calculated based on the candidate word's position within the sentence and its position in the entire text. The calculation method is as follows:

[0083] ;

[0084] in, It is the position offset value. These are weighting coefficients. The total number of words in the sentence containing the candidate core words. This is the total number of sentences. It is the index of the specific position of the core candidate word in the sentence. It is the subscript of the sentence containing the core candidate word.

[0085] It should be noted that the positional bias value is used to describe the positional factors of the candidate core words. The earlier the candidate core word appears, the higher the probability that it will be the final core word.

[0086] It should be noted that the word frequencies of the candidate words are obtained through the following steps: counting the word frequencies of the candidate core words in the candidate core word set and normalizing the statistical values ​​of the word frequencies.

[0087] It should be noted that the word2vec value is obtained through the following steps: loading the word2vec model to predict the word embedding sequence of the article title and the word embedding sequence of the sentence containing the candidate core words, and then calculating the cosine similarity between the two word embedding sequences.

[0088] It should be noted that the word2vec model is a relevant model for generating word vectors. It is mainly used to embed a high-dimensional space with the number of all words into a continuous vector space with a much lower dimension. Each word or phrase is mapped to a vector in the real number field to achieve word vectorization, which is used to measure the relevance between candidate core words and article titles.

[0089] It should be noted that the sentence similarity calculation using word2vec involves first segmenting the sentence into words, then converting each word into a one-dimensional vector based on the word2vec model, then summing the vectors of all words to represent sentences of arbitrary length as vectors of fixed dimensions, and finally calculating the cosine similarity between the word vectors of two sentences for similarity comparison.

[0090] It should be noted that the article title contains important information from the English news text. The higher the relevance between the sentence and the article title, the higher the probability that the sentence is a key sentence.

[0091] It should be noted that the Jaccard coefficient is obtained through the following steps: comparing the similarity and difference between the set of nouns contained in the article title and the set of all candidate core words in the sentence containing the candidate core words.

[0092] It should be noted that the Jaccard coefficient is the ratio of the size of the intersection between two sets to the size of the size of the union between two sets. It is used to compare the similarity and differences between finite sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.

[0093] In one specific exemplary embodiment, a positional bias value is calculated for the core word "Yankees" recalled from the English text "The Yankees went to greatlengths to incorporate many features of their old stadium (bottom) into their new one (top)."

[0094] Calculation formula based on position offset value:

[0095] ;

[0096] Total number of words in the sentence containing the core word "Yankees" The index is 20, indicating the specific location of the keyword "Yankees" in the sentence. The weight is 2, and the weighting coefficients a and b are calculated at 0.5. Since there is only one sentence in this embodiment, the total number of sentences is... 1, the subscript of the sentence containing the core word "Yankees". The value is 1. Substituting this value into the position offset formula, we obtain the result. .

[0097] Similarly, in a specific exemplary embodiment, for the English text “An ice bridge holding avast Antarctic ice shelf in place has shattered and may herald a wider collapse caused by global warming, a scientist said Saturday. “It’s amazing how the ice has ruptured,” said David Vaughan, a glaciologist with the British Antarctic Survey. “Two days ago it was intact,” he said, referring to an asatellite image of the Wilkins ice shelf. The satellite picture, by the European Space Agency, showed that a strip of ice about 25 miles long that is believed to pin the ice shelf in place had snapped. The loss of the ice bridge could mean a wider breakup of the ice shelf, which is about the size of Connecticut.”, the core word to be recalled according to the method described in the recall steps is the second “ice” appearing in the sentence, that is, “ice” as a noun, and its positional bias value is calculated.

[0098] Calculation formula based on position offset value:

[0099] ;

[0100] Total number of words in the sentence containing the core word "ice" The index is 27, indicating the specific location of the word "ice" in the sentence. The weight is 8, with weight coefficients a and b calculated at 0.5, and the total number of sentences... The value is 5, the subscript of the sentence containing the core word "ice". The value is 1. Substituting this value into the position offset formula, we obtain the result. .

[0101] Figure 5 A flowchart illustrating the generation of element text using a WHO element extraction method in an English news context, according to an embodiment of the present invention, is shown. It includes:

[0102] S5-1: Traverse each candidate core word in the candidate core word matrix;

[0103] S5-2: Extract the subscript of the sentence containing the last core word;

[0104] S5-3: Traverse all subtrees of the sentence;

[0105] S5-4: Find the leaf node in the subtree to which the final core word belongs;

[0106] S5-5: Starting from the leaf node described in S5-4, find the node to which the simple sentence node S belongs;

[0107] S5-6: The branch paths connecting the simple sentence nodes to the leaf nodes of the final core words are used as WHO elements to form WHO element text.

[0108] In a specific embodiment, after performing the above-mentioned recall and sorting steps, the final core word "Yankees" is extracted. The index of the final core word "Yankees" is extracted, and the leaf node with label NP to which "Yankees" belongs is found from the leaf node where the final core word "Yankees" is located. The parent node of this node is the simple sentence node S. The branch path connecting the simple sentence node S to the node to which the final core word "Yankees" belongs is used as the WHO element. The result "The Yankees" is the generated WHO element text.

[0109] The technical advantages of this invention are as follows: the recall method aims to extract candidate words from news texts. The recall method directly determines whether WHO elements can be extracted. Compared with the prior art, the recall method of this invention traverses sentences, while the prior art traverses referential resolution sets. The recall method of this invention expands the recall results in candidate word extraction.

[0110] The sorting method aims to select the most representative words from the candidate words as core words, which directly determines the accuracy of WHO element extraction. Compared with the existing technology, the sorting method of the present invention is redesigned and expands the types of feature values ​​for the recall extraction method proposed in the present invention.

[0111] Both methods are indispensable. Compared with traditional methods, this invention can successfully and accurately extract WHO elements from some texts where WHO elements cannot be extracted.

[0112] Finally, it should be noted that the above embodiments are for illustration only and not for limiting the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting English news elements based on core word diffusion, characterized in that, The method includes: S1: The recall step is to select a subset that meets the core word requirements from the text data of online news. Specifically, it is to extract all nouns in all noun phrases corresponding to the WHO elements in the text data to obtain a core word candidate set. S2: Sorting step, which is to form an initial matrix of candidate core words using the core word candidate set mentioned in S1, calculate the feature value of each candidate core word in the initial matrix of candidate core words and perform weighted summation, sort the candidate core words from largest to smallest according to the result of weighted summation, take the first core word of the sorting result as the last core word according to the parameter requirements and output the sorted candidate core word matrix. S3: The text element generation step is to expand the last core word mentioned in S2 into a complete element text. Specifically, the core word text is expanded according to the tree structure of the text data mentioned in S1. After finding the leaf node corresponding to the last core word mentioned in S2 in the tree structure of the text data, a simple sentence node is found in the direction of the parent node of the leaf node. Then, the simple sentence node is connected to the leaf node, and all the texts passed through the intermediate links are connected to form the final WHO element text. The calculation of the feature value described in S2 includes calculating the position bias value, candidate word frequency, word2vec value, and Jaccard coefficient of the article title and the sentence containing the candidate word. Finally, it is determined whether the candidate core word is a named entity. If the candidate word is a named entity word, the feature value is set to 1; otherwise, the feature value is set to 0. The positional bias value is calculated based on the sentence position and the position of the candidate core word in the entire text. The calculation method is as follows: ； wherein x is a position bias value, a, b are weight coefficients, L token_length is the total number of words in the sentence where the candidate core word is located, L s_length is the total number of sentences, i token_i is the specific position index of the sentence where the core candidate word is located, i s_i is the index of the sentence where the core candidate word is located.

2. The method for extracting English news elements based on core word diffusion according to claim 1, characterized in that, Step S1 also includes the following steps prior to recall: cleaning the online news in S1 and combining the article title, lead, and article topic of the online news into the text data.

3. The method for extracting English news elements based on core word diffusion according to claim 1, characterized in that, The candidate word frequencies are obtained through the following steps: statistically analyzing the word frequencies of the candidate core words in the candidate core word set and normalizing the statistical values ​​of the word frequencies.

4. The method for extracting English news elements based on core word diffusion according to claim 1, characterized in that, The word2vec value is obtained through the following steps: loading a word2vec model to predict the word embedding sequence of the article title and the word embedding sequence of the sentence containing the candidate core words, and then calculating the cosine similarity between the two word embedding sequences.

5. The method for extracting English news elements based on core word diffusion according to claim 1, characterized in that, The Jaccard coefficient is obtained by comparing the similarity and difference between the set of nouns contained in the article title and the set of all candidate core words in the sentence containing the candidate core words.

6. A computer-readable storage medium storing computer program instructions thereon, characterized in that, The computer program instructions, when executed by a processor, implement the method as described in any one of claims 1-5.

Images