A text key phrase extraction method, storage medium and device integrating inter-sentence correlation relationship

Abstract

The application belongs to the field of natural language processing, and particularly relates to a text key phrase extraction method, a storage medium and a device that integrate inter-sentence correlation, comprising: extracting a nominal phrase in a part-of-speech combination mode; filtering the nominal phrase by combining two Trie trees to obtain a candidate phrase set; calculating a global semantic similarity score of the candidate phrase; clustering each sentence to obtain a sentence cluster containing different semantic information; and sorting the candidate phrases in the sentence cluster according to the global semantic similarity score of the candidate phrase to obtain a key phrase set; the application integrates the inter-sentence relationship in the unsupervised key phrase extraction model based on embedding, improves the accuracy of the model, and constructs different Trie trees to calculate mutual information and left and right information entropy as a candidate phrase extraction method in the key phrase extraction model, thereby reducing the probability of incomplete semantic information appearing in the extracted candidate phrase set.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing, and specifically relates to a method, storage medium, and device for extracting key phrases from text that incorporates inter-sentence relationships. Background Technology

[0002] Listed companies publish numerous periodic reports annually on various financial websites, such as Tonghuashun, Shenzhen Stock Exchange, Shanghai Stock Exchange, and Eastmoney, greatly facilitating access to company information. Investors can use this information for investment, stock trading, and other economic activities. Financial companies can utilize this information for data analysis, compiling various bar charts and line graphs for presentation to a wider audience. Because periodic reports are lengthy, contain numerous financial terms and a large amount of unstructured data, the sheer volume of information increases the reading difficulty. Key phrases, concise summaries of textual information, represent the theme and core viewpoints of the text. Compared to long texts, they are concise, efficient, highlight key points, and are easy to understand. High-quality key phrases also help readers quickly comprehend the content. Therefore, this invention employs key phrase extraction technology to extract the main information from the MD&A section of periodic reports. This can assist users in enhancing their understanding of the reports and may even replace the original reports, improving user work efficiency.

[0003] Automatic key phrase extraction (APTE) is a technique that can automatically extract important information and keywords from articles. Current unsupervised APTE models can be categorized into statistical, graph-based, and embedding-based models. Statistical models primarily determine whether a candidate phrase is a key phrase by statistically analyzing its frequency, part-of-speech tagging, and other features. Graph-based models use adjacency relationships between words for identification. Embedding-based models use pre-trained language models to incorporate external and semantic information, obtaining word vector encodings to ultimately extract key phrases. Currently, embedding-based unsupervised APTE methods offer higher accuracy.

[0004] While current embedding-based unsupervised key phrase extraction methods outperform traditional methods, they only calculate the semantic relevance between candidate phrases and documents, failing to capture the connections between different types of context and candidate phrases. Therefore, a document might contain sentences with different semantic information, and these sentences, while expressing different information, may also contain key phrases, but these phrases are not extracted by the model. Furthermore, current embedding-based unsupervised key phrase extraction models use grammatical templates for candidate phrase extraction. In periodic reports, this can result in multiple consecutive noun phrases, each containing words with independent semantic information. Existing methods that use part-of-speech combination, mutual information, and left / right information entropy to obtain noun phrases avoid this problem. These methods fall into two main categories: first, addressing the issue of incorrect word segmentation by word segmentation tools, these methods first identify new words in the document and add them to the word segmentation library to prevent incorrect segmentation; second, after obtaining candidate phrases using part-of-speech combination, the values ​​of mutual information and left / right information entropy are calculated as part of the subsequent key phrase score. The first approach is not suitable for embedding-based unsupervised key phrase extraction models because the model needs to obtain a candidate phrase set before extracting key phrases. The second approach, on the other hand, requires optimization of the process of calculating mutual information and left and right information entropy to make its extraction more accurate. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method for extracting key phrases from text that incorporates inter-sentence relationships. This method includes:

[0006] S1. Obtain the text data to be processed, and extract noun phrases from the text using part-of-speech combination;

[0007] S2. Construct two different Trie trees;

[0008] S3. Filter noun phrases based on two different Trie trees, remove phrases with incomplete semantic information, and obtain a candidate phrase set;

[0009] S4. The word vector representations of candidate phrases and documents were obtained by referring to the SIF model, and then the semantic similarity score was calculated and used as the global semantic score of the candidate phrase.

[0010] S5. Using the WRD algorithm, calculate the similarity between sentences, and use the AP clustering algorithm to cluster each sentence to obtain sentence clusters containing different semantic information;

[0011] S6. Based on the global semantic score of the candidate phrases, sort the candidate phrases in the sentence clusters, extract the key phrases from different sentence clusters, and collect all the key phrases to obtain the key phrase set corresponding to the document.

[0012] To achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the above-described text key phrase extraction methods.

[0013] To achieve the above objectives, the present invention also provides a text key phrase extraction device incorporating inter-sentence relationships, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the text key phrase extraction device incorporating inter-sentence relationships performs any of the above-described text key phrase extraction methods.

[0014] The beneficial effects of this invention are:

[0015] 1. This invention, based on the existing embedding-based unsupervised key phrase extraction model, incorporates inter-sentence relationships to realize a text key phrase extraction method that integrates inter-sentence relationships. This includes obtaining candidate phrase word vectors and document word vector representations; using cosine similarity calculation to obtain the semantic similarity between the two, which is then used as the global semantic score of the candidate phrase; obtaining sentence vector encoding using a pre-trained model and calculating the similarity between sentences using WRD to obtain a similarity matrix; clustering sentences with different semantic information based on the similarity matrix; ranking the clustered sentences and extracting key phrases to obtain the final key phrase set, thus improving the accuracy of current embedding-based unsupervised key phrase extraction models.

[0016] 2. To make the model suitable for periodic reporting, this invention utilizes an external financial dictionary and modifies the word vector values ​​when generating word vectors using a pre-trained language model.

[0017] 3. This invention provides a novel candidate phrase extraction method. Based on existing methods for obtaining noun phrases through part-of-speech combination, two different Trie trees are constructed to calculate mutual information and left and right information entropy, respectively, to filter the noun phrases and obtain the final candidate phrase set. Unlike previous methods that use mutual information and left and right information entropy to calculate the internal cohesion and edge freedom of words, this invention constructs two different Trie trees, one for calculating mutual information and the other for calculating left and right information entropy. Calculating mutual information requires calculating the co-occurrence frequency of the phrase and the occurrence frequency of individual words that make up the phrase. The occurrence frequency of individual words can be calculated using an external dictionary, reflecting the importance of the word from external information. The error generated by the information obtained in the above way is smaller than that generated by using only a single document. To address the problem that when calculating left and right information entropy, external dictionaries only involve the frequency of occurrence of individual words, making it impossible to obtain the richness of the word sets to the left and right of that word. At the same time, the reference data for calculating left and right information entropy using only a single text is too limited, this invention collects second-order words and their corresponding third-order words from all documents as initial data to construct a Trie tree, thereby increasing the richness of the words that can be paired with second-order words. Compared to the previous method of constructing only a Trie to calculate these two values, this method is also more accurate. Attached Figure Description

[0018] Figure 1 This is a flowchart of a text key phrase extraction method incorporating inter-sentence relationships, according to the present invention.

[0019] Figure 2 This is a flowchart of candidate phrase extraction in steps 101-103 of Embodiment 1 of the present invention;

[0020] Figure 3 This is a flowchart illustrating the construction process of TrieA in step 103 of Embodiment 1 of the present invention;

[0021] Figure 4 This is a flowchart of the construction process of TrieB in step 103 of embodiment 1 of the present invention;

[0022] Figure 5 This is a structural diagram of the key phrase extraction model in steps 104-106 of Embodiment 1 of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] A method for extracting key phrases from text that incorporates inter-sentence relationships, such as... Figure 1 As shown, the method includes:

[0025] S1. Obtain the text data to be processed, and extract noun phrases from the text using part-of-speech combination;

[0026] S2. Construct two different Trie trees;

[0027] S3. Filter noun phrases based on two different Trie trees, remove phrases with incomplete semantic information, and obtain a candidate phrase set;

[0028] S4. Calculate the global semantic similarity score for each candidate phrase in the candidate phrase set;

[0029] S5. Using the WRD algorithm, calculate the similarity between sentences, and then use the AP clustering algorithm to cluster each sentence to obtain sentence clusters containing different semantic information.

[0030] S6. Based on the global semantic score of the candidate phrases, sort the candidate phrases in the sentence clusters, extract the key phrases from different sentence clusters, and collect all the key phrases to obtain the key phrase set corresponding to the document.

[0031] Example 1

[0032] This embodiment describes a method for extracting key phrases from text that incorporates inter-sentence relationships, such as... Figure 2 As shown, the method includes:

[0033] Step 101: Extract noun phrases from the document using part-of-speech combination. Specifically:

[0034] The input text was segmented using the HanLp word segmentation tool, resulting in multiple segmentation results. Stop words in the segmentation results were then marked, and their corresponding parts of speech were uniformly changed to [IN]. The stop word list used was the Harbin Institute of Technology stop word list. Next, based on the parts of speech corresponding to the segmented words, part-of-speech combinations were performed to obtain the noun phrases contained in the document. Noun phrases consist of nouns, verbs, and adjectives. The part-of-speech combination method is as follows: if it is a single word, it consists of a word with the part of speech of noun; if it is two words, one of the two words must be a noun. For example, the word segmentation result for "Based on ecological perspective mobile library service environment design." is "('based','in'),('ecology','n'),('perspective','n'),('of','in'),('mobile','vn'),('library','nis'),('service','vn'),('environment','n'),('design','vn'),('。','w')", where "based on", "of", and "." are stop words, and their corresponding parts of speech are adjusted to IN. After combining the parts of speech, the resulting noun phrases are: "'ecological perspective'", "mobile library", "library service", "service environment", and "environment design".

[0035] Step 102: Construct two different Trie trees for subsequent calculation of mutual information and left and right information entropy. Details are as follows:

[0036] In this step, the present invention constructs two Trie trees (TrieA and TrieB) to calculate mutual information and left and right information entropy, respectively. (See reference here.) Figure 3 The process of constructing the TrieA tree first involves introducing an external dictionary, treating each line of the dictionary as a node in the tree, thus initializing the tree. When calculating a document, the corresponding second-order words in the document are added to the initialized TrieA. In the TrieA tree, each node contains four attributes: word_finish, count, child, and isback, representing whether it is the last word in a phrase, the count of words in the phrase, child nodes, and whether it is a left-neighboring third-order word, respectively.

[0037] Constructing a TrieB tree, refer to the following: Figure 4The process of constructing the TrieB tree involves using the candidate phrase list extracted in step 101 to extract second-order words. The words to the left of each second-order word are merged with the second-order word to form left third-order words, and the words to the right of each second-order word are merged with the second-order word to form right third-order words. For example, the second-order word "service environment" corresponds to the third-order words "library service environment" and "service environment design." This process continues, extracting the third-order words corresponding to all second-order words in all datasets and merging them into a single array. At this point, in the TrieB tree, the third-order words corresponding to the second-order word "service environment" may not only be "library service environment" and "service environment design," but may also include other third-order words if they appear in other datasets.

[0038] Step 103: Using the two constructed Trie trees, filter out phrases with incomplete semantic information from the noun phrases. Specifically:

[0039] First, using the TrieA tree constructed in step 102, calculate the occurrence probability of words that make up second-order words, and the occurrence probability of each second-order word in the document. Substitute these two values ​​into the following formula to calculate the mutual information (PMI) value.

[0040]

[0041] In the formula, x and y represent two different words contained in a second-order term; p′ x p′ represents the probability of x derived from an external dictionary. y p represents the probability of y derived from an external dictionary. xy This represents the probability of second-order words appearing. For example, the mutual information values ​​for "mobile library," "library service," "service environment," and "environmental design" are 27.54, 24.99, 20.91, and 19.36, respectively. A higher value for "mobile library" indicates a stronger correlation between the words "mobile" and "library," while a lower value for "environmental design" indicates a weaker correlation between "environment" and "design."

[0042] Using the constructed TrieB tree, calculate the left and right information entropy. The calculation process is as follows: First, calculate the total number of left and right adjacent words of the phrase in the TrieB tree, and then substitute them into the following formula to obtain the left and right information entropy values ​​corresponding to the second-order words.

[0043]

[0044]

[0045] Where W represents the corresponding second-order candidate phrase; p(a) represents the probability of the left third-order word 'a' appearing in the left third-order words of W; p(b) represents the probability of the right third-order word 'b' appearing in the right third-order words of W; count l Represents the set of left-adjacent words of the second-order word W; count r Let W represent the set of right-neighboring words of the second-order word W. For example, the left-neighbor information entropy values ​​for "mobile library," "library service," "service environment," and "environmental design" are 4.57, 5.35, 2.52, and 1.58, respectively, and the right-neighbor information entropy values ​​are 1.0, 3.98, 5.87, and 1.92, respectively. Based on the calculation results, "library service" has a tighter internal cohesion, while "environmental design" has a slightly lower internal cohesion. The final score is calculated using the following formula based on word frequency, mutual information, and the corresponding left and right information entropy values.

[0046]

[0047] Next, the candidate phrase list obtained in step 101 is filtered based on the final score. Words with scores below 1 are removed, and if all scores are above 1, the bottom 20% of words are deleted. The final list is the candidate phrase list. For example, for the phrase "Design of mobile library service environment based on ecological perspective," the table below lists the frequency, mutual information, left and right information entropy values, calculation results, and whether it can be used as a candidate phrase. Filtering out the last few words yields the final second-order words, which are then added to the candidate phrase set.

[0048]

[0049]

[0050] The structural diagrams for steps 104 to 106 are as follows: Figure 5 As shown.

[0051] Step 104: Calculate the global semantic similarity score of the candidate phrases. Details are as follows:

[0052] First, the word vector representation E corresponding to the word segmentation result is obtained using the pre-trained model ELMo. n 'n' represents the number of word segmentation results in the document. For example, the sentence "based on, ecology, perspective, of, mobile, library, service, environment, design" corresponds to the word vector representation E. n={E1,E2,E3,E4,E5,E6,E7,E8,E9}. Stop words need to be removed from the sentence; the stop word list used here is the one published by Harbin Institute of Technology. Then, to make the extracted key phrases more consistent with the language characteristics of the financial field, this invention utilizes the dictionary used in step 102 to change the word vector weights in the sentence vector. Finally, using the calculation method in the SIF algorithm, the candidate phrase word vector E is obtained by calculating all context embeddings in the segmented document. C The word vector E corresponding to the document D The similarity score is calculated using the formula below, and the score is used as the global semantic similarity score of the candidate phrases.

[0053]

[0054] For example, the global semantic score for "Mobile-Library" is 0.7342; the global semantic score for "Library-Service" is 0.4232; the global semantic score for "Ecology-Perspective" is 0.6453; and the global semantic score for "Service-Environment" is 0.5772.

[0055] Step 105: Calculate the similarity between sentences and use a clustering algorithm to obtain sentence clusters that do not contain different semantic information. Specifically:

[0056] First, the word vector representation of the sentence is obtained according to step 104. Then, the similarity score between sentences is calculated using the WRD algorithm, resulting in the similarity matrix Sim. m×m , where m represents the number of sentences in the document. Then, the sentence clustering unit (CLU) is obtained using the similarity matrix. The Aptitude Test (AP) algorithm included in Google's scikit-learn is called to obtain a set of category labels (kind) corresponding to different sentences, as shown in the formula below.

[0057] kind = AP(Sim) m×m )

[0058] Each sentence will have a corresponding kind tag. For example, in the sentence "Design of a mobile library service environment based on an ecological perspective," the kind tag is 1. This sentence will be placed in the same sentence cluster as other sentences in the document that are also tagged with 1. Ultimately, this will result in R distinct sentence clusters. Each sentence cluster contains the sentence's order value in the document, where 's' represents the sentence number within the sentence cluster.

[0059] Step 106: Sort candidate phrases in sentence clusters, extract key phrases from different sentence clusters, and finally obtain the key phrase set corresponding to the document. Specifically:

[0060] First, after obtaining the similarity between candidate phrases and documents, the candidate phrases are placed into different sentence clusters based on their positional information. Then, the candidate phrases within each sentence cluster are ranked according to their global semantic information scores. Here, C refers to the number of candidate phrases. For example, in the sentence "Design of a mobile library service environment based on an ecological perspective," the candidate phrases are ranked as follows: "mobile-library," "ecological-perspective," "service-environment," and "library-service." Secondly, different semantic information has different levels of importance in a document, which is reflected in the varying number of sentences contained in different sentence clusters. Some sentence clusters may contain five sentences, while others may contain only one. Clearly, sentence clusters with five sentences account for a larger proportion of the text and are more important. The importance of a sentence cluster in the document needs to be determined by the ratio p of the number of sentences (count) in the cluster to the total number of sentences (sum) in the document, as shown in the formula below:

[0061]

[0062] In the formula, count represents the number of sentences in different sentence clusters; sum represents the number of sentences in the document.

[0063] Finally, based on the importance of sentence clusters in the document, determine how many key phrases each sentence cluster should contain. Extract key phrases based on this determination, and then summarize them to obtain the final set of key phrases, as shown in the formula below:

[0064]

[0065] In the formula, CCL represents the candidate phrase cluster obtained from the sentence cluster; i represents the j-th sentence cluster; and R represents the number of sentence clusters. For example, sentence cluster number 1 contains 5 sentences with an importance of 0.6. If 10 candidate phrases are to be extracted, 6 candidate phrases will be extracted from the first sentence cluster. The first 6 candidate phrases from sentence cluster number 1 are then selected and summarized to obtain the final set of key phrases corresponding to the document.

[0066] Example 2

[0067] In one embodiment of the present invention, the present invention further includes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described text key phrase extraction methods.

[0068] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented using computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0069] A text key phrase extraction device incorporating inter-sentence relationships includes a processor and a memory; the memory stores a computer program; the processor is connected to the memory and executes the computer program stored in the memory to enable the text key phrase extraction device incorporating inter-sentence relationships to perform any of the aforementioned text key phrase extraction methods.

[0070] Specifically, the memory includes various media capable of storing program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card, or optical disk.

[0071] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0072] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting key phrases from a text incorporating inter-sentence relationship, characterized in that, The method comprises the following steps: S1, obtaining text data to be processed, and extracting a nominal phrase in the text by using a part-of-speech combination mode; S2, constructing two different Trie trees; S3, filtering the nominal phrase according to the two different Trie trees, removing phrases with incomplete semantic information expression in the nominal phrase, and obtaining a candidate phrase set; S4, calculating the global semantic similarity score of each candidate phrase in the candidate phrase set; S5, calculating the similarity between sentences by using a WRD algorithm, and clustering the sentences by using an AP clustering algorithm to obtain a sentence cluster containing different semantic information; S6, sorting the candidate phrases in the sentence cluster according to the global semantic similarity score of the candidate phrases, extracting key phrases from different sentence clusters, and collecting all the key phrases to obtain a key phrase set corresponding to the document.

2. The method of claim 1, wherein the method further comprises: The method for extracting the nominal phrase in the text by using the part-of-speech combination mode comprises the following steps: performing word segmentation on the input text by using a word segmentation tool to obtain a plurality of word segmentation results; marking the stop words in the word segmentation results; and combining the word segmentation results by using a part-of-speech combination mode to obtain the nominal phrase contained in the document; wherein the part-of-speech combination mode comprises: if it is a single word, the word is composed of a word with a part of speech of a noun, and if it is two words, one of the two words must be a noun.

3. The method of claim 1, wherein the method further comprises: The two different Trie trees are constructed by the following steps: the first Trie tree is used to calculate the mutual information of the nominal phrase, and the second Trie tree is used to calculate the left and right information entropy of the nominal phrase; the first Trie tree is constructed by the following steps: obtaining an external dictionary, and taking the words in the dictionary as the nodes of the TrieA tree; the second Trie tree is constructed by the following steps: obtaining all the two-order candidate phrases of the text, obtaining the left and right three-order candidate phrases corresponding to the two-order candidate phrases according to the documents where the two-order candidate phrases are located, and finally merging all the three-order candidate phrases into an array for initializing the second Trie tree.

4. The method of claim 1, wherein the method further comprises: The filtering of the nominal phrase according to the two different Trie trees comprises the following steps: The mutual information value of the nominal phrase is calculated according to the two probability values, and the expression is: ； wherein, and denotes the probability that two different words are contained in a second order word; denotes the probability that a word is contained in a second order word; the probability based on an external dictionary; denotes the probability that a word is contained in a second order word; the probability based on an external dictionary; denotes the probability that a second order word occurs; The total number of left and right adjacent words of the phrase in the second Trie tree is calculated, and the value of the left and right information entropy corresponding to the two-order word is calculated according to the total number of left and right adjacent words of the phrase, and the calculation formula is: ； ； wherein, represents the corresponding second order candidate phrase; represents a left third order word in the probability of occurrence of the left third order word represents a right third order word in the probability of occurrence of the right third order word represents a second order word the left adjacent word set of the second order word represents the right adjacent word set of the second order word ​ The score is calculated according to the values corresponding to the word frequency, mutual information and left and right information entropy, the nominal phrase is screened according to the score, and a candidate phrase set is obtained.

5. The method of claim 4, wherein the method further comprises: The formula for calculating the score is: ； wherein, represents a candidate phrase final score; represents a candidate phrase frequency of occurrence in the text; represents a candidate phrase left information entropy of represents a candidate phrase right information entropy of 6. The method of claim 1, wherein the method further comprises: The process of calculating the global semantic similarity score of each candidate phrase in the candidate phrase set includes: using the pre-trained model ELMo to obtain the word vector representations corresponding to the word segmentation results. The word weights in the document are calculated using a financial dictionary, and the word vectors are adjusted based on the calculated weights. The value; using the SIF algorithm to calculate word vector representations Contextual embeddings are used to obtain candidate phrase word vectors. Word vectors corresponding to documents The global semantic similarity score of candidate phrases is calculated using a similarity calculation formula.

7. The method of claim 1, wherein the method further comprises: The process of obtaining the sentence cluster containing different semantic information includes: calculating the similarity score between the sentence word vector and the sentence word vector by using the WRD algorithm to obtain a similarity matrix , wherein m represents that there are m sentences in the document; clustering the similarity matrix by using the AP clustering algorithm to obtain a group of category labels corresponding to different sentences , collecting the sentences with the same label to obtain R different sentence clusters Each sentence cluster contains the position information of the sentence in the document, wherein the represents the position number in the sentence cluster.

8. The method of claim 1, wherein the method further comprises: The process of obtaining the key phrase set corresponding to a document includes: placing candidate phrases into different sentence clusters based on their positional information, and sorting the candidate phrases in the sentence clusters according to their global semantic similarity scores. Where C refers to the number of candidate phrases; based on the number of sentences contained in the sentence cluster. The total number of sentences in the document ratio The importance of the sentence cluster in the document is determined; based on the importance of the sentence cluster in the document, it is determined how many key phrases the sentence cluster should contain, and key phrases are extracted based on this, and the final set of key phrases is obtained by summarizing them.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the text key phrase extraction method of any one of claims 1 to 8.

10. A text key phrase extraction device incorporating inter-sentence relationships, characterized in that, The device comprises a processor and a memory; the memory is used to store a computer program; the processor is connected with the memory and is used to execute the computer program stored in the memory, so that the device for extracting key phrases in a text by integrating the inter-sentence association relationship executes the text key phrase extraction method of any one of claims 1 to 8.

Images