A dialogue text scenario division method based on a neural network architecture search technology

Abstract

The application discloses a dialogue text scene division method based on a neural network architecture search technology, and comprises the following steps: dialogue text sentence vector conversion, constructing a sentence vector model to convert dialogue text sentences into sentence vector features of the same dimension; dividing dialogue text into several paragraphs according to semantic features, constructing a segmentation model, taking dialogue text segmentation as a sequence labeling task, and using three segmentation symbols to identify a text paragraph; using CRF to constrain the sequence of segmentation symbols in the last output in the model, and searching a recurrent neural network (RNN) structure based on the neural network architecture search technology; dividing discrete text paragraphs according to dialogue scene division and dialogue text clustering, and adjusting the final clustering effect to form a clustering cluster. The application solves the problem of training sample shortage, can provide better support for some downstream tasks in the dialogue text field, has stronger dialogue semantic capturing capacity, and can better divide dialogue text according to scenes.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method for segmenting dialogue text scenarios based on neural network architecture search technology. Background Technology

[0002] With the development of computer networks, a type of text data primarily consisting of dialogues is gradually increasing, such as social chat, customer service conversations, medical diagnosis, and AI question answering. By modeling dialogue texts and segmenting them according to context, we can better understand the semantic information within them, providing excellent support for many downstream tasks. Dialogue text context segmentation has a very broad application prospect and has become a research hotspot in the field of natural language processing.

[0003] As a subfield of automated machine learning, neural network architecture search is a technique designed for specific tasks to automatically search for network model architectures. Neural network architecture search technology has broad application value and research significance, and has received widespread attention from researchers in recent years, achieving certain research progress. Search space, search strategy, and evaluation are the three most important aspects of evaluating neural network architecture search technology. The search space defines the basic unit or operation of the network architecture, while the search strategy defines the specific implementation method for searching a complete sub-architecture from the search space. Finally, the evaluation strategy is responsible for scoring the final performance of the searched sub-architecture and returning the score to the search strategy to guide the next step of the search process.

[0004] Reinforcement learning is a commonly used search strategy in neural network architecture search techniques. In reinforcement learning-based neural network architecture search methods, a recurrent neural network (RNN) is typically used to sample a sub-network structure. This sub-network structure is then trained on a task training set until convergence. A score is calculated based on the final performance of the sub-network structure, and this score is used as a reward in reinforcement learning to update the RNN model. The RNN model searching for the sub-network architecture is also called the controller, which searches for specified network architecture parameters at different time steps according to a set configuration. The weights of the controller and the searched network architectures are updated alternately. When the weights of the searched network architectures are updated, the controller weights remain unchanged; while during controller updates, previously searched network architecture samples are evaluated.

[0005] Current dialogue text context segmentation methods have poor dialogue semantic capture capabilities. They cannot extract sentence vectors containing semantic information by constructing positive and negative sample pairs. Furthermore, there is a shortage of training samples during the dialogue semantic capture process, which fails to provide better support for downstream tasks related to dialogue text. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a dialogue text scenario segmentation method based on neural network architecture search. The method uses neural network architecture search technology to complete the search for key models in the dialogue text, solves the problem of training sample shortage, has a stronger dialogue semantic capture capability, and, combined with scenario paragraph clustering algorithm, can better segment the dialogue text according to scenario.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A dialogue text context segmentation method based on neural network architecture search technology includes the following steps:

[0009] Step 1: Dialogue text sentence vector conversion. Train a sentence vector model to capture the semantic information of the dialogue. After the sentence vector model is built, use the model to convert the dialogue text sentences into sentence vector features of the same dimension.

[0010] Step 2: Divide the dialogue text into several paragraphs according to semantic features, construct a segmentation model, and treat the dialogue text segmentation as a sequence labeling task, using three segmentation symbols to identify a text paragraph; use Conditional Random Field (CRF) to constrain the final output of the order between segmentation symbols in the segmentation model, and search for a recurrent neural network (RNN) structure to capture dialogue semantic information based on neural network architecture search technology.

[0011] Step 3: Divide the discrete text segments according to the dialogue context. Based on the DBSCAN algorithm, a dialogue segment context clustering algorithm is proposed to cluster the dialogue text. The final clustering effect is adjusted by controlling the density radius and sample threshold. The final cluster can be regarded as the divided dialogue context.

[0012] Specifically, step one includes the following sub-steps:

[0013] Step 11: Construct the training sample dataset for the dialogue text sentence vector model DSimCSE;

[0014] Step 12: Use the objective function of positive sample pairs as the training objective of the sentence vector model, and train the sentence vector model. The objective function of positive sample pairs is shown in the following formula:

[0015]

[0016] Where S i S represents the original statement in a positive sample pair. + i represents the new statement in the positive sample pair, m represents the number of dialogue statements, t represents the temperature hyperparameter, and sim represents the cosine similarity function;

[0017] For negative sample pairs, the training objective of the positive sample pairs is reversed before training the sentence vector model;

[0018] Step 13: Input the dialogue statements into the dialogue text sentence vector model DSimCSE for sentence vector feature transformation to obtain the sentence vector features corresponding to the dialogue statements.

[0019] Specifically, step 11 includes the following steps:

[0020] Step 111: Use different statements from the same dialogue text as negative sample pairs;

[0021] Step 112: When constructing positive sample pairs, for each dialogue statement, firstly, the part-of-speech discriminator identifies words in the dialogue statement that are nouns or verbs. By repeating these words, new statements are obtained, and the new statements and the original statements are used as a pair of positive sample pairs.

[0022] Specifically, step two includes the following sub-steps:

[0023] Step 21: Search for a novel recurrent neural network architecture (DRNN) using a reinforcement learning-based neural network architecture search method to capture semantic information in dialogue sentence vectors;

[0024] Step 22: After the output feature matrix of the DRNN, add a layer using a Conditional Random Field (CRF) to constrain the order between segment symbols, where... <s>Indicates the sentence that begins a paragraph. <m>This indicates a sentence in the middle of a paragraph. <e>This indicates the end of a paragraph; given the input dialogue sequence Dialogue={s1,s2,...,s...} m The segmentation markers output by CRF are T = {t1, t2, ..., t}. m }, then the score for T can be expressed as follows:

[0025]

[0026] Where Score(S,T) represents the score of the segmentation symbol T, A represents the transition matrix of the CRF, and P represents the output probability matrix;

[0027] After training is complete, the transition matrix parameters of the CRF are determined; the scores of all sequences are output, and the sequence with the highest score is taken as the final semantic output of the dialogue sentence vector, as shown in the following formula:

[0028]

[0029] Among them, T S This represents all the segmented marker sequences that can be output when given an input dialogue sequence S;

[0030] Step 23: Optimize the parameters of the segmented model on the training set;

[0031] Step 24: Optimize the controller parameters on the validation set;

[0032] Step 25: Obtain the segmentation results of the dialogue text.

[0033] Specifically, step 21 includes the following sub-steps:

[0034] Step 211: Define the search space of DRNN as a directed acyclic graph with N nodes. Each node in the graph represents an activation function calculation node, and the directed edges between nodes represent the direction of information flow. The starting node of the edge represents the input feature, and the ending node of the edge represents the output feature. The calculation from the input feature to the output feature is determined by the activation function on the ending node.

[0035] Step 212: Control the DRNN search process through the controller, and search for the computation node information on the DRNN structure in two time steps; when searching for node i, the first time step determines the predecessor node of the current node to ensure that the predecessor node has been searched before node i; the second time step determines the activation function required on node i.

[0036] Specifically, step 23 includes the following sub-steps:

[0037] Step 231: Fix controller parameters;

[0038] Step 232: The controller searches for a DRNN architecture;

[0039] Step 233: Capture dialogue semantic information using the DRNN architecture and output the final CRF result. Calculate the segmentation loss as shown in the following formula:

[0040]

[0041] Step 234: Optimize the parameters of the segmented model.

[0042] Specifically, step 24 includes the following sub-steps:

[0043] Step 241: Fix the parameters on the feature nodes in the DRNN;

[0044] Step 242: The controller searches for multiple DRNN architectures;

[0045] Step 243: Calculate the controller loss using the best-performing DRNN;

[0046] Step 244: Optimize the controller parameters.

[0047] Specifically, step three includes the following sub-steps:

[0048] Step 31: For the final set of paragraph sentence vectors S = {s1, s2, ..., s...} m }, s i ∈R n The entire paragraph vector is represented by the average of the paragraph's sentences, as shown in the following formula:

[0049]

[0050] Step 32: Traverse the collection of paragraphs. For two paragraphs S... i and S j The distance between the two can be calculated using the following formula:

[0051]

[0052] Step 33: Obtain the core paragraph set by adjusting the density radius and sample threshold, and cluster the density-connected core paragraphs into one class as a different scenario;

[0053] Step 34: Repeat step 33 to obtain all the divided dialogue scenarios.

[0054] Specifically, step 33 includes the following sub-steps:

[0055] Step 331: Initialize the core paragraph set to empty, iterate through each paragraph in turn, calculate the distance between the core paragraph and other paragraphs, count the adjacent paragraphs of each paragraph within the density radius eps, and add them to the core paragraph set if the number of adjacent paragraphs is greater than or equal to the sample threshold MinPts.

[0056] Step 332: Iterate through the core paragraph set, group paragraphs within the core paragraph density radius eps into one category, and remove these paragraphs from the core paragraph set.

[0057] The beneficial effects of this invention are:

[0058] 1. An unsupervised dialogue text sentence vector model is proposed, which solves the problem of training sample shortage and can provide better support for some downstream tasks in the field of dialogue text.

[0059] 2. The dialogue text segmentation model based on neural network architecture search technology has a stronger ability to capture dialogue semantics, and combined with the contextual paragraph clustering algorithm, it can better divide the dialogue text according to the context. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0061] Figure 2 This is a schematic diagram illustrating the construction process of positive and negative samples in the DSimCSE sentence vector model.

[0062] Figure 3 This is a schematic diagram of the DRNN search process.

[0063] Figure 4 This is a diagram of the dialogue text segmentation model structure.

[0064] Figure 5 This is a diagram of the training algorithm for a dialogue text segmentation model.

[0065] Figure 6 This is a diagram of the scenario paragraph clustering algorithm. Detailed Implementation

[0066] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the following detailed description of the technical solution is provided. Obviously, the described embodiments are only a portion of the embodiments of this invention, not all of them, and should not be construed as limiting the scope of implementation of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the protection scope of this invention.

[0067] Example 1:

[0068] In this embodiment, as Figure 1 As shown, a dialogue text context segmentation method based on neural network architecture search technology includes the following steps:

[0069] Step 1: Dialogue text sentence vector conversion. Train a sentence vector model to capture the semantic information of the dialogue. First, solve the problem of constructing positive and negative sample pairs based on the policy. After the sentence vector model is constructed, it can be used to convert the dialogue sentences into sentence vector features of the same dimension.

[0070] Step 2: Divide the dialogue text into several paragraphs according to semantic features. First, a segmentation model is proposed. This segmentation model treats the segmentation of dialogue text as a sequence labeling task and uses three segmentation symbols to identify a text paragraph. In this segmentation model, a Conditional Random Field (CRF) is used to constrain the final output order between the segmentation symbols. Then, an RNN structure for capturing dialogue semantic information is searched based on neural network architecture search technology.

[0071] Step 3: Divide the discrete paragraphs according to the dialogue context. Based on the DBSCAN algorithm, a dialogue paragraph context clustering algorithm is proposed. The final clustering effect is adjusted by controlling the density radius and sample threshold. The final cluster can be regarded as the divided dialogue context.

[0072] Furthermore, step 1 includes the following sub-steps:

[0073] Step 11: As Figure 2 As shown, the training sample dataset for constructing the dialogue text sentence vector model DSimCSE is presented.

[0074] Furthermore, step 11 includes the following steps:

[0075] Step 111: Use different statements from the same dialogue text as negative sample pairs;

[0076] Step 112: When constructing positive sample pairs, for a dialogue statement, firstly, the part-of-speech discriminator identifies words in the statement that are nouns or verbs. By repeating these words, new statements are obtained, and the new statements and the original statements are treated as a pair of positive sample pairs.

[0077] Step 12: Train the sentence vector model using the following formula as the training objective, where S i S represents the original statement in a positive sample pair. + i Let represent the new statement in the positive sample pair, m represent the number of dialogue statements, t represent the temperature hyperparameter, and sim represent the cosine similarity function. This formula is the objective function for positive sample pairs, while for negative sample pairs, the training objective is the opposite.

[0078]

[0079] Step 13: Input the dialogue statement into DSimCSE to obtain the corresponding sentence vector features;

[0080] Furthermore, step 2 includes the following sub-steps:

[0081] Step 21: Search for a novel RNN architecture, DRNN, using a reinforcement learning-based neural network architecture search method to capture semantic information in dialogue sentence vectors.

[0082] Furthermore, step 21 includes the following sub-steps:

[0083] Step 211: Design the search space of the DRNN, defining it as a directed acyclic graph with N nodes.

[0084] Each node in the graph represents an activation function computation node, and the directed edges between nodes represent the direction of information flow. The starting node of the edge represents the input feature, and the ending node of the edge represents the output feature. The computation from the input feature to the output feature is determined by the activation function on the ending node.

[0085] Step 212: Control the DRNN search process through the controller, such as... Figure 3 As shown, when searching the DRNN structure, the controller searches for the information of a computation node in the DRNN in two time steps. When searching for node i, the first time step determines the predecessor node of the current node, that is, which node the input features of the computation originate from. It should be ensured that the predecessor node has been searched before node i, i.e., its number should be between [1, i-1]. This connection method between nodes also utilizes the idea of ​​skip connections, which makes the searched DRNN units have stronger representational capabilities. The second time step mainly determines the activation function that should be applied to node i. These activation functions include common ones such as ReLU, tanh, and sigmoid.

[0086] Step 22: As Figure 4 As shown, a CRF layer is added after the output feature matrix of the DRNN to constrain the order between segment symbols, where <s>Indicates the sentence that begins a paragraph. <m>This indicates a sentence in the middle of a paragraph. <e>This indicates the end of a paragraph. Given the input dialogue sequence Dialogue={s1,s2,...,s...} m The segmentation markers output by CRF are T = {t1, t2, ..., t}. m }, then the score of T can be defined as follows, where A represents the transition matrix of the CRF and P represents the probability matrix of the output.

[0087]

[0088] Once training is complete, the transition matrix parameters of the CRF are fixed. At this point, it is only necessary to output the scores of all sequences and take the sequence with the highest score as the final output. The specific process is shown in the following formula, where TS represents all segmented labeled sequences that can be output when the input dialogue sequence S is given, including those that are logically impossible.

[0089]

[0090] Step 23: Optimize the parameters of the segmented model on the training set.

[0091] Furthermore, step 23 includes the following sub-steps:

[0092] Step 231: Fix controller parameters;

[0093] Step 232: The controller searches for a DRNN architecture;

[0094] Step 233: Capture dialogue semantic information through DRNN and output the final CRF result. Calculate the segmentation loss as shown in the following formula:

[0095]

[0096] Step 234: Optimize the parameters of the segmented model;

[0097] Step 24: Optimize the controller parameters on the validation set.

[0098] Furthermore, step 24 includes the following sub-steps:

[0099] Step 231: Fix the parameters on the feature nodes in the DRNN;

[0100] Step 232: The controller searches for multiple DRNN architectures;

[0101] Step 233: Calculate the controller loss using the best-performing DRNN;

[0102] Step 234: Optimize controller parameters;

[0103] Step 25: Obtain the segmented results of the dialogue text. The training method for the overall model is as follows: Figure 5 As shown.

[0104] Furthermore, step 3 includes the following sub-steps:

[0105] Step 31: For the final set of paragraph sentence vectors S = {s1, s2, ..., s...} m }, s i ∈R n The entire paragraph vector is represented by the average of the paragraph's sentences, as shown in the following formula:

[0106]

[0107] Step 32: Traverse the collection of paragraphs. For two paragraphs S... i and S j The distance between the two can be calculated using the following formula:

[0108]

[0109] Step 33: Obtain the core paragraph set by adjusting the density radius and sample threshold, and cluster the density-connected core paragraphs into one class as different scenarios. The specific process is as follows: Figure 6 As shown.

[0110] Furthermore, step 33 includes the following sub-steps:

[0111] Step 331: Initialize the core paragraph set to empty, iterate through each paragraph in turn, calculate the distance between each paragraph and other paragraphs, count the adjacent paragraphs of each paragraph within the density radius eps, and add the adjacent paragraphs to the core paragraph set if the number of adjacent paragraphs is greater than or equal to the sample threshold MinPts.

[0112] Step 332: Iterate through the core paragraph set, group paragraphs within the core paragraph density radius eps into one category, and remove these paragraphs from the core paragraph set.

[0113] Step 34: Repeat step 33 to obtain all the divided dialogue scenarios.

[0114] This embodiment trains a sentence vector model to capture dialogue semantic information, converting dialogue sentences into sentence vector features of the same dimension. It addresses the problem of constructing positive and negative sample pairs based on a strategy. Simultaneously, the dialogue text is divided into several paragraphs according to semantic features. First, a segmentation model is proposed, treating dialogue text segmentation as a sequence labeling task, using three segmentation symbols to identify a text paragraph. Conditional Random Fields (CRFs) are used in this segmentation model to constrain the final output order of segmentation symbols. A recurrent neural network (RNN) structure for capturing dialogue semantic information is searched using neural network architecture search technology. Discrete paragraphs are then divided according to dialogue context. A dialogue paragraph context clustering algorithm based on the DBSCAN algorithm is proposed. The final clustering effect is adjusted by controlling the density radius and sample threshold; the resulting clusters can be considered as segmented dialogue contexts. This invention proposes an unsupervised dialogue text sentence vector model, solving the problem of training sample shortage and providing better support for downstream tasks related to dialogue text. The dialogue text segmentation model implemented based on neural network architecture search technology has stronger dialogue semantic capture capabilities, and combined with the context paragraph clustering algorithm, it can better segment dialogue text according to context.

[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.< / e> < / m> < / s> < / e> < / m> < / s>

Claims

1. A method for dialogue text scenario partitioning based on neural network architecture search technology, characterized in that, The method comprises the following steps: Step one: converting the dialogue text sentence vector, training a sentence vector model for capturing dialogue semantic information, and converting the dialogue text sentence into a sentence vector feature of the same dimension by using the model after the sentence vector model is constructed; Step two: dividing the dialogue text into several paragraphs according to the semantic features, constructing a segmentation model, taking the dialogue text segmentation as a sequence labeling task, and using three segmentation symbols to identify a text paragraph; in the segmentation model, the conditional random field (CRF) is used to constrain the sequence of segmentation symbols, and a recurrent neural network (RNN) structure for capturing dialogue semantic information is searched based on a neural network architecture search technology; Specifically, the following sub-steps are included: Step 21: searching a new recurrent neural network architecture (DRNN) for capturing semantic information in the dialogue sentence vector by using a neural network architecture search method based on reinforcement learning; Step 22: After the output feature matrix of the DRNN, a layer using conditional random field (CRF) is attached to constrain the order between segment symbols, where <s>indicates a paragraph start statement, <m>to indicate a sentence in the middle of a paragraph, <e>represents the end-of-sentence statement; for the known input dialogue sequence Dialogue={ , ,.., }, the segment marker symbol output by the CRF is T={ , ,.., }, and the score of T can be shown as follows:< / e> < / m> < / s> <s> wherein, Score S T denotes the score of the segment label symbol T, A denotes the transition matrix of the CRF, and P denotes the probability matrix of the output.​​ After the training is completed, the transition matrix parameters of the CRF are determined; all sequence scores are output, and the sequence with the maximum sequence score is taken as the final semantic output of the dialogue sentence vector, which is specifically shown in the following formula: wherein, represents all the segmented token sequences that can be output when the input dialogue sequence S is input. Step 23: optimizing the parameters of the segmentation model on the training set; Step 24: optimizing the parameters of the controller on the validation set; Step 25: obtaining the segmentation result of the dialogue text; Step three: dividing the discrete text paragraphs according to the dialogue scenarios, clustering the dialogue text by using a dialogue paragraph scenario clustering algorithm based on the DBSCAN algorithm, adjusting the final clustering effect by adjusting the density radius and sample threshold, and finally forming a clustering cluster which can be regarded as a divided dialogue scenario.

2. The dialogue text scenario partitioning method based on neural network architecture search technology according to claim 1, characterized in that, The step one specifically includes the following sub-steps: Step 11: constructing a training sample data set of a dialogue text sentence vector model DSimCSE; Step 12: taking the objective function of the positive sample pair as the training target of the sentence vector model, training the sentence vector model, and the objective function of the positive sample pair is shown in the following formula: wherein denotes the original sentence in the positive sample pair, denotes the new sentence in the positive sample pair, m denotes the number of dialogue sentences, t denotes the temperature hyperparameter, and sim denotes the cosine similarity function. For the negative sample pair, the training target of the positive sample pair is taken in reverse to train the sentence vector model; Step 13: inputting the dialogue sentence into the dialogue text sentence vector model DSimCSE to convert the sentence vector feature, and obtaining the sentence vector feature corresponding to the dialogue sentence.

3. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 2, characterized in that, The step 11 includes the following steps: Step 111: taking different sentences in the same dialogue text as negative sample pairs; Step 112: when constructing the positive sample pair, for each dialogue sentence, first, the part-of-speech identifier is used to identify the words with the part-of-speech of noun or verb in the dialogue sentence, the new sentence is obtained by repeating the words, and the new sentence and the original sentence are taken as a group of positive sample pairs.

4. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 1, characterized in that, The step 21 includes the following sub-steps: Step 211: defining the search space of the DRNN as a directed acyclic graph with N nodes, each node in the graph representing an activation function calculation node, and the directed edges between the nodes representing the flow direction of information; the starting node of the edge represents the input feature, and the terminal node of the edge represents the output feature, and the calculation from the input feature to the output feature is determined by the activation function on the terminal node; Step 212: search the computing node information on the DRNN structure by the controller controlling the search process of the DRNN, search in two time steps; when searching node i, the first time step determines the predecessor node of the current node, ensuring that the predecessor node has been searched before node i; the second time step determines the activation function required on node i.

5. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 1, characterized in that, The step 23 comprises the following sub-steps: Step 231: fix the controller parameters; Step 232: the controller searches a DRNN architecture; Step 233: capture the dialogue semantic information through the DRNN architecture, and output the final CRF result, calculate the segmentation loss, as shown in the following formula: Step 234: optimize the parameters of the segmentation model.

6. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 1, characterized in that, The step 24 comprises the following sub-steps: Step 241: fix the parameters on the feature nodes in the DRNN; Step 242: the controller searches multiple DRNN architectures; Step 243: calculate the controller loss through the best-performing DRNN; Step 244: optimize the parameters of the controller.

7. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 1, characterized in that, The step three comprises the following sub-steps: Step 31: For the final obtained paragraph sentence vector set S = { , ,.., }, , use the paragraph sentence mean to represent the entire paragraph vector, as shown in the following formula:​ Step 32: Traverse the set of passages, and for each pair of passages and calculate the distance between them by the formula: Step 33: obtain the core paragraph set by adjusting the density radius and sample threshold, and regard the core paragraphs connected in density as a different scene; Step 34: repeat step 33 to obtain all divided dialogue scenes.

8. The dialog text scenario partitioning method based on neural network architecture search technology according to claim 7, characterized in that, The step 33 comprises the following sub-steps: Step 331: initialize the core paragraph set to be empty, traverse the paragraphs one by one, calculate the distance between the core paragraph and other paragraphs, and count the adjacent paragraphs of each paragraph within the density radius eps, if the number of adjacent paragraphs is greater than or equal to the sample threshold MinPts, add it to the core paragraph set; Step 332: traverse the core paragraph set one by one, cluster the paragraphs within the core paragraph density radius eps into a class, and delete these paragraphs from the core paragraph set. < / s>

Images