CN113127632B - Heterogeneous graph-based text summarization method and device, storage medium and terminal - Google Patents

Abstract

The invention discloses a text summarization method and device based on heterogeneous graphs, a storage medium and a terminal, wherein the method comprises the following steps: performing knowledge fusion on a preset knowledge base and a target text, acquiring word characteristics and sentence characteristics of the target text, and constructing a text heterogeneous graph of the target text based on the word characteristics and the sentence characteristics; updating the text heterogeneous graph through a graph attention network based on the edge weight and the attention weight to obtain an updated version of the text heterogeneous graph; calculating multi-class abstract indexes of sentence vectors in the updated version text heterogeneous graph, and calculating classification weights of corresponding sentence vectors according to the multi-class abstract indexes corresponding to each sentence vector; and respectively weighting sentence features in the updated version text heterogeneous image based on the classification weight of the sentence vector, acquiring corresponding sentence labels based on the weighted sentence features, and generating a text abstract according to the acquired sentence labels. The invention has a more direct mode that sentences and words are respectively used as two types of nodes to construct a heterogeneous graph, and word nodes are used as intermediaries of the sentences, thereby enriching the association among the sentences and indirectly transmitting information.

Description

Heterogeneous graph-based text summarization method and device, storage medium and terminal

technical field

The present invention relates to the technical field of text generation, and in particular, to a method and device for text summarization based on heterogeneous graphs, a storage medium and a terminal.

Background technique

Automatically generating text summaries is an important task in the field of natural language processing, which aims to compress the original text and generate a short description containing the main content of the original text. There are two main research areas: generative and extractive. Generative methods focus on generating summaries verbatim after encoding the entire document, while extractive methods directly select sentences from documents to combine into summaries. Extractive summarization methods are more efficient and generate more readable summaries than generative ones.

The key step in the extractive text summarization task is to establish the relationship between each sentence and the article. At present, most of the existing methods are based on the recurrent neural network (RNN) to obtain the sentence association, but this kind of method usually cannot capture the sentence. long-range dependencies. Using graph structure to represent text is an effective way to solve the above problems, but how to model text as graph reasonably remains to be studied. Recently, Graph Neural Networks (GNNs) have demonstrated powerful feature extraction capabilities for graph data. Subsequently, text summarization methods based on Graph Neural Networks have also been proposed. Some works use Rhetorical Structure Theory (RST) to decompose sentences into basic semantic units ( EDU) and construct the rhetorical structure theory structure tree, and then use the graph convolutional network (GCN) to complete the graph information aggregation update. Although the method based on the basic semantic unit achieves good results, the process of generating the basic semantic unit is complicated, and only one kind of node is used to construct the graph. In addition, the strength of the association between sentences in the extractive summary is more important, but the current work on heterogeneous graphs only adds edges between different types of nodes, so that there is no direct association between sentences.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is that the process of generating the basic semantic unit in the existing text summary generation method based on graph neural network is relatively complicated, only one kind of node is used to construct the graph, and the correlation between sentences is weak, which is not conducive to Generation of extractive summaries.

In order to solve the above technical problems, the present invention provides a method for text summarization based on heterogeneous graphs, including:

Perform knowledge fusion with the preset knowledge base and the target text, obtain word features and sentence features of the target text, and construct a text heterogeneity graph of the target text based on the word features and sentence features;

Based on the edge weight and the attention weight, the text heterogeneous graph is updated through the graph attention network to obtain an updated version of the text heterogeneous graph;

Calculate the multi-category summary index of the sentence vector in the updated version of the text heterogeneity graph, and calculate the classification weight of the corresponding sentence vector according to the multi-category summary index corresponding to each sentence vector;

The classification weights based on the sentence vectors respectively weight the sentence features in the updated text heterogeneity graph, obtain corresponding sentence labels based on the weighted sentence features, and generate text summaries according to the obtained sentence labels.

Preferably, knowledge fusion is performed between the preset knowledge base and the target text, and the acquisition of word features and sentence features of the target text includes:

Respectively vectorize the knowledge in the preset knowledge base and the content coding in the target text to obtain the knowledge vector in the preset knowledge base and the word vector in the target text;

Respectively calculate the attention weight of each word vector in the target text and the knowledge vector in the preset knowledge base, and obtain the attention weight of each word vector in the target text;

Taking the attention weight of the word vector in the target text as the weight in turn, weighting and merging the knowledge vectors in the preset knowledge base respectively to obtain the knowledge weight of each word vector in the target text;

Obtain the word feature of the corresponding word vector based on the knowledge weight of each word vector in the target text;

Perform local feature capture and global feature capture on the word features of the word vectors contained in each sentence vector in the target text, respectively, to obtain the local features and global features of each sentence vector, and then according to the local features of each sentence vector and The global feature obtains the sentence features of the corresponding sentence vectors respectively.

Preferably, constructing the text heterogeneity graph of the target text based on the word features and sentence features includes:

Based on the sentence features of the sentence vectors in the target text, calculate the homogeneous edge weights between all sentence vectors in the target text by a cosine similarity calculation method;

Based on the word feature of the word vector in the target text and the sentence feature of the sentence vector to which the word vector belongs, calculate the heterogeneous edge weights between all the word vectors in the target text and the sentence vector to which the word vector belongs by the TF-IDF algorithm;

Taking the word vector in the target text as the word node, the sentence vector in the target text as the sentence node, based on the word feature of the word vector, the sentence feature of the sentence vector, the homogeneous edge weight between the sentence vectors and the word The heterogeneous edge weights between the vectors and the sentence vectors to which they belong to construct the text heterogeneity graph of the target text.

Preferably, based on edge weights and attention weights, the text heterogeneous graph is updated through a graph attention network, and the step of obtaining an updated version of the text heterogeneous graph includes:

Calculate the attention weights between all sentence vectors in the target text, and calculate the attention weights between all word vectors in the target text and the sentence vectors to which the word vectors belong, and obtain all the word vectors in the target text. Homogeneous edge weights and all heterogeneous edge weights;

Based on the attention weights between all sentence vectors in the target text, the attention weights between all word vectors and the sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights, through the graph attention The force network updates all word nodes and all sentence nodes in the text heterogeneity graph to obtain an updated version of the text heterogeneity graph.

Preferably, based on the attention weights between all sentence vectors in the target text, the attention weights between all word vectors and the sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights, Updating all word nodes and all sentence nodes in the text heterogeneous graph through the graph attention network includes:

Taking the word node as the central node, the product of the attention weight between the sentence nodes connected to the central node and the heterogeneous edge weights between the sentence nodes connected to the central node is used as the weight. The sentence features of the sentence nodes connected by the nodes are weighted and aggregated to realize the update of the word nodes;

Taking the sentence node as the central node, and taking the product of the attention weight between the word nodes connected to the central node and the heterogeneous edge weights between the word nodes connected to the central node as the weight, the The word features of the word nodes connected by the nodes are weighted and aggregated to realize the update of the sentence nodes;

Take the sentence node as the central node, and take the product of the attention weight between the word nodes connected to the central node and the weight of the homogeneous edge between the sentence nodes connected to the central node as the weight. The sentence features of the sentence nodes connected by the nodes are weighted and aggregated to realize the update of the sentence nodes.

Preferably, the steps of calculating the multi-category summary indicators of the sentence vectors in the updated version of the text heterogeneity graph, and respectively calculating the classification weights of the corresponding sentence vectors according to the multi-category summary indicators corresponding to each sentence vector include:

Calculate the relevance score, redundancy score, new information score, and metric score for recall evaluation of each sentence vector in the updated version of the text heterogeneity graph;

Based on the relevance score, redundancy score, new information score, and metric score for recall evaluation of each sentence vector, respectively, the classification weight of the corresponding sentence vector is calculated by the Sigmoid function.

Preferably, the steps of calculating the relevance score, redundancy score, new information score and recall-oriented metric score of a single sentence vector in the updated text heterogeneity graph include:

Based on the text feature of the updated version of the text heterogeneity graph and the sentence feature of the sentence vector, the correlation score of the sentence vector is calculated by a bilinear function;

Based on the sentence feature of the sentence vector in the updated version of the text heterogeneity graph, the redundancy score of the sentence vector is calculated by the bilinear function;

Based on the sentence feature of the sentence vector in the updated version of the text heterogeneity map and the knowledge vector in the preset knowledge base, the new information content score of the sentence vector is calculated by a bilinear function;

Based on the unencoded vectorized target text and the textual content of the sentence vector in it, a recall-oriented metric score for the sentence vector is calculated by the recall-oriented metric function.

In order to solve the above technical problems, the present invention also provides a text summarization device based on heterogeneous graph, including:

The text heterogeneity graph building module is used to fuse the preset knowledge base with the target text, obtain the word features and sentence features of the target text, and construct the text heterogeneity of the target text based on the word features and sentence features picture;

an update module, configured to update the text heterogeneous graph through the graph attention network based on the edge weight and the attention weight to obtain an updated version of the text heterogeneous graph;

A classification weight acquisition module, configured to calculate the multi-category summary index of the sentence vector in the updated version of the text heterogeneity graph, and calculate the classification weight of the corresponding sentence vector according to the multi-category summary index corresponding to each sentence vector;

The summary generation module is used to weight the sentence features in the updated text heterogeneity map based on the classification weights of the sentence vectors, obtain corresponding sentence labels based on the weighted sentence features, and generate text summaries according to the obtained sentence labels. .

In order to solve the above-mentioned technical problem, the present invention also provides a computer-readable storage medium on which a computer program is stored, which implements the heterogeneous graph-based text summarization method when the program is executed by a processor.

In order to solve the above technical problems, the present invention also provides a terminal, including: a processor and a memory;

The memory is used for storing a computer program, and the processor is used for executing the computer program stored in the memory, so that the terminal executes the heterogeneous graph-based text summarization method.

Compared with the prior art, one or more embodiments of the above solutions may have the following advantages or beneficial effects:

Using the heterogeneous graph-based text summarization method provided by the embodiment of the present invention, the text is constructed into a text heterogeneous graph according to semantic and syntactic relations, and the node features of words and sentences are updated in combination with the graph attention network, and multiple summaries are designed. The related metrics weighted the sentences for the final summary extraction, not only considering the information transfer between words and sentences, but also the interaction between sentences. The further added external knowledge base can better help the model understand the text, and the multi-angle indicators designed for the summary task can add corresponding weights to sentences before classification, which effectively improves the model's ability to utilize text features, and then generates more accurate and reliable data. Readable abstract.

Other features and advantages of the present invention will be set forth in the description which follows, and in part will become apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the description, claims and drawings.

Description of drawings

The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and together with the embodiments of the present invention, are used to explain the present invention, and do not constitute a limitation to the present invention. In the attached image:

1 shows a flowchart of a method for text summarization based on a heterogeneous graph according to an embodiment of the present invention;

FIG. 2 shows a schematic process diagram of a text summarization method based on a heterogeneous graph according to Embodiment 1 of the present invention;

3 shows a schematic diagram of the construction process of a text heterogeneous graph in a text summarization method based on a heterogeneous graph according to Embodiment 1 of the present invention;

4 shows a schematic diagram of a single-layer update process of a text heterogeneous graph in a text summarization method based on a heterogeneous graph according to Embodiment 1 of the present invention;

5 shows a schematic diagram of an ablation learning experiment result of a text summarization method based on a heterogeneous graph according to Embodiment 1 of the present invention;

6 shows a schematic diagram of the experimental results based on the CNN&DailyMail data set and the comparative experimental results performed with other abstracting methods in the first embodiment of the present invention;

7 shows a schematic diagram of the influence of multi-angle indicators on the abstract in the text abstract method based on the heterogeneous graph according to Embodiment 1 of the present invention;

FIG. 8 shows a multi-angle index quantization example according to an embodiment of the present invention;

FIG. 9 shows a schematic structural diagram of a text summarization device based on a heterogeneous graph according to Embodiment 2 of the present invention;

FIG. 10 shows a schematic structural diagram of a terminal according to Embodiment 4 of the present invention.

Detailed ways

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples, so as to fully understand and implement the implementation process of how the present invention applies technical means to solve technical problems and achieve technical effects. It should be noted that, as long as there is no conflict, each embodiment of the present invention and each feature of each embodiment can be combined with each other, and the formed technical solutions all fall within the protection scope of the present invention.

Automatically generating text summaries is an important task in the field of natural language processing. There are two main ways to generate text summaries, generative and extractive. Generative methods focus on generating summaries verbatim after encoding the entire document, while extractive methods directly select sentences from documents to combine into summaries. Compared with generative methods, extractive summarization methods have the advantages of higher efficiency and better readability. The key step in the extractive text summarization task is to establish the relationship between each sentence and the article, and the existing methods are usually unable to capture the long-distance dependencies of sentences. Recently, graph neural networks (GNNs) have shown powerful feature extraction capabilities for graph data, and subsequently, text summarization methods based on graph neural networks have also been proposed. Although the summary generation method based on graph neural network has achieved good results, the process of generating basic semantic units is relatively complicated. Only one type of node is used to construct the graph, and only edges are added between different types of nodes, which makes the correlation between sentences weak. .

Example 1

In order to solve the above technical problems existing in the prior art, an embodiment of the present invention provides a text summarization method based on a heterogeneous graph.

Fig. 1 shows a flow chart of a method for text summarization based on a heterogeneous graph according to an embodiment of the present invention; Fig. 2 shows a schematic process diagram of a method for text summarization based on a heterogeneous graph according to an embodiment of the present invention. Referring to FIG. 1 and FIG. 2 , the method for extracting text summarization based on a heterogeneous graph according to an embodiment of the present invention includes the following preparation steps.

Further, in order to more clearly illustrate the specific implementation method of the method for extracting text summarization based on heterogeneous graphs of the present invention, the following definitions are made in advance:

是句子i的单词集。Definition of sentence set and word set: Given a target text d containing m sentences and n words, S={s ₁ , s ₂ ... s _m } is the sentence set of d,

is the set of words in sentence i.

The definition of text graph is: G={V, E} represents the graph, V represents the node set, and E represents the edge set. Since the heterogeneous graph used in the present invention contains two types of nodes, they can be divided into word node sets and sentence node sets. Therefore, the text graph TG = {V _TG , E _TG } is designed as a heterogeneous graph, where:

(1) V _TG =W∪S contains two types of nodes, W={W ₁ , W ₂ , . . . , W _m } represents a set of word sets. S={s ₁ , s ₂ . . . s _m } represents a set of sentences.

代表异质边，E_homo＝{(s_i，s_j)|s_i，s_j∈S}代表同质边。(2)E _TG =E _heter ∪E _homo , where

represents a heterogeneous edge, and E _homo = {(s _i , s _j )|s _i , s _j ∈ S} represents a homogenous edge.

(3) e _ij represents the weight of the heterogeneous edge (wi _ij , s _i ), and e′ _ij represents the weight of the homogeneous edge (s _i , s _j ).

Step S101 , perform knowledge fusion on the preset knowledge base and the target text, obtain word features and sentence features of the target text, and construct a text heterogeneity graph of the target text based on the word features and sentence features.

First of all, it is necessary to realize knowledge fusion, that is, to use the external knowledge base to complete the knowledge enrichment of word features, so that the feature representation of the text has both semantic perception and knowledge perception. Specifically, a preset knowledge base is selected, and the language type of the preset knowledge base must be the same as the language type of the target text. That is, when the target text is Chinese, the default knowledge base needs to select the Chinese knowledge base; when the target text is English, the default knowledge base needs to select the English knowledge base. Secondly, in order to integrate the knowledge of the preset knowledge base into the word features of the target text, it is necessary to first encode and vectorize the knowledge in the selected preset knowledge base and the content of the target text to obtain all knowledge vectors of the preset knowledge base and All word vectors of the target text. In order to simplify the description, directly use d to represent the target text after encoding and vectorization, W to represent the set of word vector sets of the target text after encoding and vectorization, and w _i to represent the word vectors in W; and then suppose that K represents the preset knowledge of encoding vectorization library, k represents the knowledge vector in K.

Then, the word features of each word vector in the target text are obtained. Among them, the method of obtaining the word feature of each word vector is: calculating the attention weight of the word vector _wi and all the knowledge vectors k in the preset knowledge base through bilinear operation, so as to obtain the attention of the word vector _wi force weight, the calculation method of the attention weight β _i of the word vector w _i is:

β _i =BiLinear(K,W _KB ,w) (1)

Among them, W _KB is a trainable weight parameter.

After calculating the attention weight of each word vector in the target text, it is necessary to take the attention weight of the word vector in the target text as the weight in turn, and weight and merge the knowledge vectors in the preset knowledge base respectively to obtain each word in the target text. A vector of knowledge weights. The way to obtain the knowledge weight knowledge of a single word vector is:

knowledge=β _i K (2)

At this time, the knowledge weight knowledge of the word vector already contains word-related knowledge.

Then, based on the knowledge weight of each word vector in the target text, the word features of the corresponding word vector are obtained. That is to connect the knowledge weight of each word vector with the corresponding word vector w ^k = [w, knowledge], that is, to obtain the word feature of the corresponding word vector, which will have both semantic perception and knowledge perception.

After the word features of all word vectors in the target text are obtained, the sentence features of all sentence vectors in the target text can be obtained. Specifically, the word features of the word vectors contained in each sentence vector in the target text after encoding vectorization are subjected to local feature capture and global feature capture respectively, so as to obtain the local features and global features of each sentence vector. Local features and global features obtain sentence features for each sentence vector, respectively. Among them, the capture of local features is extracted by convolutional neural network (CNN), and the capture of global features is extracted by BiLSTM. At the same time, after obtaining the sentence features of all sentence vectors in the target text after encoding vectorization, the text features of the target text after encoding vectorization can also be obtained. The specific sentence features and text features are calculated as follows:

D=BiLSZM([s ₁ , . . . , s _m ]) (4).

After obtaining the word features and sentence features in the target text after encoding vectorization, the text heterogeneity graph of the target text can be constructed through semantic grammar. Quality edge weights. The construction process of the text heterogeneity graph is shown in FIG. 3. Referring to FIG. 3, the embodiment of the present invention regards the text summary as a classification problem, and regards the sentence as the smallest unit to be classified. Therefore, the association between the sentences when generating the summary is particularly Importantly, the specific response method of this embodiment is to add homogeneous edges between sentences to construct a sentence-sentence fully connected undirected graph, and the homogeneous edges represent the semantic relevance between sentences. The request method for the amount of homogeneous edge representation is: based on the sentence features of the sentence vectors in the target text after coding vectorization, calculate the homogeneous edge weights between all sentence vectors in the target text after coding vectorization through the cosine similarity calculation method. . Heterogeneous edges between words and sentences represent the relationship between the two. In order to add more text-related information, this embodiment is based on the word features of the word vectors in the target text after encoding and vectorization and the sentence features of the sentence vectors to which the word vectors belong. , and calculate the heterogeneous edge weights between all word vectors in the target text and the sentence vectors to which the word vectors belong through the TF-IDF algorithm.

Take the word vector in the target text as the word node, and the sentence vector in the target text as the sentence node, based on the word feature of the word vector, the sentence feature of the sentence vector, the homogeneous edge weight between the sentence vectors, and the relationship between the word vector and the sentence vector to which it belongs. The heterogeneous edge weights of , construct a text-heterogeneous graph of the target text.

Step S102 , based on the edge weight and the attention weight, update the text heterogeneous graph through the graph attention network to obtain an updated version of the text heterogeneous graph.

In order to further illustrate the update process of text heterogeneous graph using graph attention network, we express the word feature and sentence feature of the word vector in the target text after encoding vectorization as the hidden state _Hw of the word node and the sentence node's Hide the state H _s and express the text features as _HD .

FIG. 4 shows a schematic diagram of a single-layer update process of a text heterogeneous graph in a text summarization method based on a heterogeneous graph according to Embodiment 1 of the present invention. Referring to Figure 4, calculate the attention weights between all sentence vectors in the target text, and calculate the attention weights between all word vectors in the target text and the sentences to which the word vectors belong, and obtain all the same sentence vectors in the target text. Primitive edge weights and all heterogeneous edge weights. The calculation method of the attention weight between the sentence vectors is as follows, and the calculation method of the attention weight between the word vector and the sentence to which the word vector belongs can also be calculated by referring to the following formula:

The attention weight between two sentence vectors is calculated as:

Among them, hi and _{h j represent} the hidden states of two sentence nodes, W _a , W _q , W _k , and W _v are trainable parameters, and α _ij is the attention weight between hi and _{h j .}

The update increment _ui is calculated by the attention weight, and the update increment can be calculated by formula (6):

Among them, μ _i can represent the increment of word nodes, and can also represent the increment of sentence nodes, and N _i represents the set of neighbor nodes of node i.

In order to make the semantic related information also participate in the update, the heterogeneous edge weight e _ij and the homogeneous edge weight e' _ij are introduced to control the update degree of nodes from the perspective of semantics and attention model. Homogeneous edge weights and heterogeneous edge weights are calculated as follows:

或

or

At this time, equation (6) can be modified as:

The above process is how the graph attention network calculates the update increment by using the attention weights between sentence vectors, the attention weights between word vectors and the sentences to which the word vectors belong, the weights of homogeneous edges and the weights of heterogeneous edges.

The process of updating the text heterogeneous graph through the graph attention network is actually the update of the word node and the sentence node in the text heterogeneous graph. Further, each update iteration process of the graph attention network actually includes three processes, namely, the update of sentence nodes to word nodes, the update of word nodes to sentence nodes, and the mutual update of sentence nodes.

The updating of the word node by the sentence node includes: using the word node as the central node, using the attention weight between the sentence nodes connected to the central node and the heterogeneous edges between the sentence nodes connected to the central node The product of the weights is used as the weight, and the sentence features of the sentence nodes connected to the central node are weighted and aggregated to realize the update of the word nodes. The update of the sentence node by the word node includes: taking the sentence node as the central node, and taking the attention weight between the word nodes connected to the central node and the heterogeneous edge weight between the word nodes connected to the central node. The product is used as a weight, and the word features of the word nodes connected to the central node are weighted and aggregated to realize the update of the sentence nodes. The mutual update between sentence nodes includes: taking the sentence node as the central node, taking the attention weight between the sentence nodes connected to the central node and the homogeneous edge weight between the sentence nodes connected to the central node. The product is used as a weight, and the sentence features of the sentence nodes connected to the central node are weighted and aggregated to realize the update of the sentence nodes. At the sentence node level, LSTM can also be applied to update the text feature H _d . The above aggregation method is: using the corresponding attention weight and the weighted aggregation of the edge full value. To illustrate the above, the following presents the t-th update process of the graph attention network:

表示单词对句子传递的消息，通过多层感知机(MLP)对该层隐藏状态进行更新。优选地，多层感知机为包含两个线性隐藏层的多层感知机。where GAT(G, H _s , H _w ) represents the graph attention update layer, G is the text graph, H _s is the sentence feature, as the request matrix in the attention mechanism, H _w is the word feature, as the attention mechanism in the Key and value matrix.

Represents the message conveyed by word-to-sentence, and the hidden state of this layer is updated through a multi-layer perceptron (MLP). Preferably, the multilayer perceptron is a multilayer perceptron comprising two linear hidden layers.

The text feature H _d is updated after each iteration update:

Through the update iteration of the homogeneous and heterogeneous graph structure based on the graph attention network GAT, the sentence will obtain more cross-sentence information through the indirect connection of the words, and the homogeneous edge between the sentence vectors enables it to obtain long-distance association, which is the abstract Extract provides more information.

Step S103: Calculate the multi-category summary index of the sentence vector in the updated version of the text heterogeneity graph, and calculate the classification weight of the corresponding sentence vector according to the multi-category summary index corresponding to each sentence vector.

Specifically, calculate the relevance score, redundancy score, new information score, and metric score for recall evaluation of each sentence vector in the updated text heterogeneity graph; The redundancy score, the new information score, and the metric score for recall evaluation are used to calculate the classification weight of the corresponding sentence vector through the Sigmoid function. That is, in order to extract suitable sentences as abstracts, the present invention sets multi-angle sentence evaluation indicators, which are evaluated from four perspectives: relevance (Rel), redundancy (Red), new information (Info) and Rouge-F1 score. The sentences are evaluated and scored, the scores are used to weight the sentence features, and the optimal N items are selected as the extraction results.

Relevance is a very intuitive metric, which indicates the degree of relevance between a sentence and the full text. The higher the value, the better the sentence represents the topic of the article; the relevancy score of a single sentence in the updated version of the text heterogeneity graph is calculated as follows: based on the updated version The text features of the text heterogeneity graph and the sentence features of the sentence vector are used to calculate the relevance score of the sentence vector through a bilinear function. Redundancy is a concept relative to relevance. A good abstract should not only match the theme of the original text, but also keep it as concise as possible, that is, the redundancy of the abstract itself should be as low as possible; a single sentence vector in the updated text heterogeneity graph The redundancy score is calculated as follows: based on the sentence features of the sentence vector in the updated version of the text heterogeneity graph, the redundancy score of the sentence vector is calculated by a bilinear function. Relevance is a criterion that ignores background knowledge and other sources of information, while the amount of new information will be combined with background knowledge to evaluate the content of the text. For readers, reading the abstract is more hopeful to know the knowledge that they did not know before. These New knowledge is new information content; the calculation method of the new information content score of a single sentence vector in the updated text heterogeneity map is: based on the sentence features of the sentence vector in the updated text heterogeneity map and the preset knowledge base. The knowledge vector, the new informativeness score of the sentence is calculated by a bilinear function. Rouge is a commonly used machine scoring method in text summarization. Using it as an evaluation index can further improve the accuracy of summarization; the metric score calculation method for recall evaluation of a single sentence vector in the updated text heterogeneity graph is as follows: based on For the unencoded vectorized target text and the textual content of that sentence vector in it, a recall-oriented metric score for the sentence vector is calculated by the recall-oriented metric function.

The sentence s is scored based on the four summary indicators, and the specific calculation formula is as follows:

Rel=h _s W _rel H _D (15)

Red=h _s W _red A _s (16)

Info=h _s W _info H _k (17)

Rouge=R(s, ref) (18)

Score=Sigmoid(Rel-Red+Info+Rouge) (19)

where h _s is the feature vector of the sentence s, h _d is the feature vector of the text to which it belongs, H _k is the knowledge base code, s is the sentence itself, ref is the reference summary, W _rel , W _red , and W _info are learnable parameters, R For the Rouge calculation function, the calculation results of each indicator are processed by the Sigmoid function as the classification weight of the current sentence.

Step S104 , weight the corresponding updated sentence features based on the classification weight of each sentence vector, obtain corresponding sentence labels based on the weighted sentence features, and generate text summaries according to the obtained sentence labels.

Specifically, after the classification weight of each sentence vector is obtained in the above steps, the updated sentence features are weighted respectively to obtain the weighted sentence features, and then the perceptron classifier is used to obtain the corresponding sentence features based on the weighted sentence features. sentence labels, and then generate text summaries based on the obtained sentence labels. In the process of using the perceptron classifier to select the summary sentences, the best N sentences are selected as the summary, and the triple blocking strategy is used to reduce the summary redundancy.

In order to verify the effectiveness of the present invention, the text summarization method based on the heterogeneous graph of the present invention is compared with other methods, and the Rouge evaluation method is used to evaluate the effect of each abstraction method. The score of the Rouge evaluation is shown in Figure 5. . Among them, KHHGS represents the invention of a heterogeneous graph-based text summarization method.

In the comparative experiment, the present invention uses the CNN&Daily Mail data as the data set, and compares it with several other summarization methods. The results are shown in Figure 5. From the comparative experimental results, it can be found that compared with the RNN model BiGRU, the KHHGS model (the method of this application) has obvious advantages, and is also ahead of the Transformer model. From the data in the table, it can be found that the Transformer score is similar to that of the homogeneous graph method, indicating that Transformer can indeed be regarded as a fully connected graph at the sentence level. The effect of KHHGS is also better than the previous heterogeneous graph abstract model HSG, and the three indicators of Rouge are improved by 0.14/0.46/0.97 respectively. It can be seen that the relevant strategy proposed by the present invention can effectively improve the heterogeneous graph model for text abstraction. effect of the task.

In order to better explain the effectiveness of the relevant strategy proposed by the present invention on the text summarization task, the model is ablated and learned, and the three modules of the external knowledge base, the homogeneous graph and the summarization index calculation are removed from the model for experimental analysis. . As shown in Figure 6, each piece of data in the table represents the experimental result after removing the relevant modules. After the knowledge base is added, Rouge-2 and Rouge-L are improved to a certain extent, but Rouge-1 is not significantly improved. It may be because the knowledge base used in the present invention is a general knowledge base, and the effect on news text corpus is weak. , in view of the fact that there is currently no recognized mature news knowledge base; the addition of homogenous graphs can significantly increase various Rouge indicators, especially the Rouge-L value, which may be due to the addition of homogenous graphs to enhance the connection between sentences. The relationship can be better used by the model, which affects the number of longest overlapping substrings in the final abstract extraction result; in addition, the multi-angle standard for abstracts can effectively improve the effect, and further experiments are conducted on the impact of different indicators on abstracts. As shown in Figure 7, the horizontal axis represents the range of a certain index score, and the vertical axis represents the probability that the sentence obtained within this range belongs to the reference abstract. It can be seen from the figure that the relevance and Rouge score have a greater impact on the abstract results. The sentences with higher scores of these two items are highly likely to belong to the reference abstract; the new information volume also has a certain degree of influence on the abstract, but it is not significant. Since the present invention uses the general knowledge base as the background knowledge for calculating the new information volume, here In the item content, it can be considered that the general knowledge with low recognition degree is filtered, similar to the operation of removing stop words during data processing, so that the model can focus on other key sentences. Redundancy has no obvious impact. It may be because the present invention proposes a sentence-level summarization method, and the Rouge score used for model evaluation belongs to the summary-level evaluation. The evaluation of single sentence redundancy cannot effectively improve the summarization method. the final effect.

In order to more intuitively express the role of the indicators proposed by the present invention, a test sample is selected to quantify the indicators, as shown in Figure 8, each row in the table is a sentence of the original text, and the corresponding indicators on the right are calculated by formulas. Normalized scores and total scores, through the above analysis, redundancy does not play a key role in this model, so it is not added to the quantification list. The data in the table shows that there is a certain relationship between the sentence length and the relevance index. The longer the sentence, the higher the relevance to the original text and the more information it contains. Therefore, the longest sentence 2 in the table gets the highest relevance score, and at the same time It is not difficult to judge from the content of sentence 2. Its description is very similar to the reference abstract, so it has obtained a high Rouge score, and finally sentence 2 has the highest comprehensive score.

To sum up, the text summarization method based on heterogeneous graph proposed by the present invention has great advantages in capturing long-distance dependencies compared with the existing commonly used summarization methods. Abstract's multi-angle indicators can make full use of text data features, thereby improving the quality of text summaries.

The heterogeneous graph-based text summarization method provided by the embodiment of the present invention constructs a text heterogeneous graph based on semantic and syntactic relationship connections, updates two types of node features of words and sentences in combination with a graph attention network, and designs multiple abstract related The metric for weighted evaluation of sentences for final summary extraction takes into account not only the information transfer between words and sentences, but also the inter-sentence interactions. The further added external knowledge base can better help the model understand the text, and the multi-angle indicators designed for the summary task can add corresponding weights to sentences before classification, which effectively improves the model's ability to utilize text features, and then generates more accurate and reliable data. Readable abstract. A more direct way in the embodiment of the present invention is to construct a heterogeneous graph by using sentences and words as two types of nodes respectively, and using word nodes as the intermediary of sentences to enrich the association between sentences and transmit information indirectly.

Embodiment 2

In order to solve the above technical problems existing in the prior art, an embodiment of the present invention provides a text summarization device based on a heterogeneous graph.

FIG. 9 shows a schematic structural diagram of a text summarization device based on a heterogeneous graph according to Embodiment 2 of the present invention; with reference to FIG. 9 , the text summarization device based on a heterogeneous graph of the present invention includes sequentially connected text heterogeneous graph building modules, updating module, classification weight acquisition module and summary generation module.

Among them, the text heterogeneity graph building module is used to fuse the preset knowledge base with the target text, obtain the word features and sentence features of the target text, and build the text heterogeneity graph of the target text based on the word features and sentence features.

The update module is used to update the text heterogeneity graph through the graph attention network based on the edge weight and attention weight, and obtain the updated version of the text heterogeneity graph.

The classification weight acquisition module is used to calculate the multi-class summary index of the sentence vector in the updated version of the text heterogeneity graph, and calculate the classification weight of the corresponding sentence vector according to the multi-class summary index corresponding to each sentence vector.

The summary generation module is used to weight the sentence features in the updated text heterogeneity map based on the classification weights of the sentence vectors, obtain the corresponding sentence labels based on the weighted sentence features, and generate text summaries according to the obtained sentence labels.

The heterogeneous graph-based text summarization device provided by the embodiment of the present invention constructs a text heterogeneous graph by connecting the text according to semantic and syntactic relations, updates the node features of words and sentences in combination with the graph attention network, and designs multiple abstract related The metric for weighted evaluation of sentences for final summary extraction takes into account not only the information transfer between words and sentences, but also the inter-sentence interactions. The further added external knowledge base can better help the model understand the text, and the multi-angle indicators designed for the summary task can add corresponding weights to sentences before classification, which effectively improves the model's ability to utilize text features, and then generates more accurate and reliable data. Readable abstract. A more direct way in the embodiment of the present invention is to construct a heterogeneous graph by using sentences and words as two types of nodes respectively, and using word nodes as the intermediaries of sentences to enrich the association between sentences and transmit information indirectly.

Embodiment 3

In order to solve the above-mentioned technical problems existing in the prior art, an embodiment of the present invention further provides a storage medium, which stores a computer program, and when the computer program is executed by a processor, can realize the text based on the heterogeneous graph in the first embodiment. Summary of all steps in the method.

The specific steps of the heterogeneous graph-based text summarization method and the beneficial effects obtained by applying the readable storage medium provided by the embodiment of the present invention are the same as those of the first embodiment, and will not be repeated here.

It should be noted that the storage medium includes various media that can store program codes, such as ROM, RAM, magnetic disk or optical disk.

Embodiment 4

In order to solve the above technical problems existing in the prior art, an embodiment of the present invention further provides a terminal.

FIG. 10 shows a schematic structural diagram of a terminal according to Embodiment 4 of the present invention. Referring to FIG. 10 , the terminal in this embodiment includes a processor and a memory that are connected to each other; the memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory. , so that all steps in the method for text summarization based on a heterogeneous graph in Embodiment 1 can be implemented when the terminal executes it.

The specific steps of the method for text summarization based on the heterogeneous graph and the beneficial effects obtained by the terminal provided by the application of the embodiment of the present invention are the same as those of the first embodiment, which will not be repeated here.

It should be noted that the memory may include random access memory (Random Access Memory, RAM for short), and may also include non-volatile memory (non-volatile memory), such as at least one disk memory. Similarly, the processor can also be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it can also be a digital signal processor (Digital Signal Processing, DSP for short) , Application Specific Integrated Circuit (ASIC for short), Field Programmable Gate Array (FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

Although the disclosed embodiments of the present invention are as above, the content described is only an embodiment adopted to facilitate understanding of the present invention, and is not intended to limit the present invention. Any person skilled in the art to which the present invention belongs, without departing from the spirit and scope disclosed by the present invention, can make any modifications and changes in the form and details of the implementation, but the protection scope of the present invention is still The scope as defined by the appended claims shall prevail.

Claims (7)

1. A text summarization method based on heterogeneous graphs comprises the following steps:

performing knowledge fusion on a preset knowledge base and a target text, acquiring word features and sentence features of the target text, and constructing a text heterogeneous graph of the target text based on the word features and the sentence features;

updating the text heterogeneous graph through a graph attention network based on the edge weight and the attention weight to obtain an updated version of the text heterogeneous graph;

calculating multi-class abstract indexes of sentence vectors in the updated version text heterogeneous image, and calculating classification weights of corresponding sentence vectors according to the multi-class abstract indexes corresponding to each sentence vector;

respectively weighting sentence features in the updated version text heterogeneous image based on the classification weight of the sentence vector, acquiring corresponding sentence labels based on the weighted sentence features, and generating a text abstract according to the acquired sentence labels;

wherein constructing a text heterogeneous graph of the target text based on the word features and sentence features comprises:

based on sentence features of sentence vectors in the target text, calculating the homogeneous edge weight between every two sentence vectors of all the sentence vectors in the target text in a cosine similarity calculation mode;

calculating heterogeneous edge weights among all word vectors and sentence vectors to which the word vectors belong in the target text through a TF-IDF algorithm based on the word features of the word vectors in the target text and the sentence features of the sentence vectors to which the word vectors belong;

taking word vectors in the target text as word nodes and sentence vectors in the target text as sentence nodes, constructing a text heterogeneous graph of the target text based on word features of the word vectors, sentence features of the sentence vectors, homogeneous edge weights among the sentence vectors and heterogeneous edge weights among the word vectors and the sentence vectors to which the word vectors belong,

and the multiclass abstract indexes of the sentence vector comprise a relevance score, a redundancy score, a new information score and a recall-oriented evaluation metric score,

updating the text heterogeneous graph through a graph attention network based on the edge weight and the attention weight, wherein the step of acquiring the updated text heterogeneous graph comprises the following steps:

calculating attention weights between every two sentence vectors of all sentence vectors in the target text, calculating attention weights between all word vectors in the target text and the sentence vectors to which the word vectors belong, and acquiring all homogeneous edge weights and all heterogeneous edge weights in the target text;

updating all word nodes and all sentence nodes in the text heterogeneous graph through a graph attention network based on the attention weight between every two sentence vectors in the target text, the attention weight between all word vectors and the sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights to obtain an updated version text heterogeneous graph,

wherein updating all word nodes and all sentence nodes in the text heterogeneous graph through a graph attention network based on attention weights between every two sentence vectors of all sentence vectors in the target text, attention weights between all word vectors and sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights comprises:

taking word nodes as central nodes, taking the product of attention weight between sentence nodes connected with the central nodes and heterogeneous edge weight between sentence nodes connected with the central nodes as weight, and carrying out weighted aggregation on sentence characteristics of the sentence nodes connected with the central nodes to realize the updating of the word nodes;

taking sentence nodes as central nodes, taking the product of attention weight between word nodes connected with the central nodes and heterogeneous edge weight between the word nodes connected with the central nodes as weight, and carrying out weighted aggregation on word characteristics of the word nodes connected with the central nodes to realize the updating of the sentence nodes;

and taking sentence nodes as central nodes, taking the product of attention weight between the sentence nodes connected with the central nodes and homogeneous edge weight between the sentence nodes connected with the central nodes as weight, and carrying out weighted aggregation on the sentence characteristics of the sentence nodes connected with the central nodes to realize the update of the sentence nodes.

2. The method of claim 1, wherein performing knowledge fusion on a preset knowledge base and a target text, and obtaining word features and sentence features of the target text comprises:

respectively encoding and vectorizing knowledge in a preset knowledge base and content in a target text to acquire a knowledge vector in the preset knowledge base and a word vector in the target text;

respectively calculating the attention weight of each word vector in the target text and the attention weight of the knowledge vector in the preset knowledge base to obtain the attention weight of each word vector in the target text;

sequentially taking the attention weight of the word vector in the target text as a weight, and respectively weighting and combining the knowledge vectors in the preset knowledge base to obtain the knowledge weight of each word vector in the target text;

acquiring word features of corresponding word vectors based on knowledge weight of each word vector in the target text;

and respectively performing local feature capture and global feature capture on the word features of the word vectors contained in each sentence vector in the target text to obtain the local features and the global features of each sentence vector, and respectively obtaining the sentence features of the corresponding sentences according to the local features and the global features of each sentence vector.

3. The method of claim 1, wherein calculating multiple types of summarization indexes of sentence vectors in the updated text heterogeneous map and calculating classification weights of corresponding sentence vectors according to the multiple types of summarization indexes corresponding to each sentence vector respectively comprises:

calculating a relevance score, a redundancy score, a new information score and a recall rate evaluation-oriented metric score of each sentence vector in the updated text heterogeneous graph;

and calculating the classification weight of the corresponding sentence vector through a Sigmoid function based on the relevance score, the redundancy score, the new information score and the recall rate evaluation-oriented metric score of each sentence vector.

4. The method of claim 3, wherein the step of calculating relevance scores, redundancy scores, new information scores, and recall-assessment-oriented metric scores for individual sentence vectors in the updated textual heterogeneous graph comprises:

calculating the relevance score of the sentence vector through a bilinear function based on the text characteristics of the updated version text heterogeneous image and the sentence characteristics of the sentence vector;

calculating a redundancy score of the sentence vector through a bilinear function based on the sentence characteristics of the sentence vector in the updated text heterogeneous image;

calculating a new information quantity score of the sentence vector through a bilinear function based on the sentence characteristics of the sentence vector in the updated text heterogeneous graph and the knowledge vector in the preset knowledge base;

and calculating the recall-rate evaluation-oriented metric score of the sentence vector through the recall-rate evaluation-oriented metric function based on the target text which is not coded and vectorized and the text content of the sentence vector.

5. A text summarization device based on heterogeneous graphs is characterized by comprising:

the text heterogeneous graph building module is used for carrying out knowledge fusion on a preset knowledge base and a target text, acquiring word features and sentence features of the target text, and building a text heterogeneous graph of the target text based on the word features and the sentence features;

the updating module is used for updating the text heterogeneous graph through a graph attention network based on the edge weight and the attention weight to obtain an updated version of the text heterogeneous graph;

the classification weight acquisition module is used for calculating multi-class abstract indexes of sentence vectors in the updated version text heterogeneous image and calculating classification weights of corresponding sentence vectors according to the multi-class abstract indexes corresponding to each sentence vector;

the abstract generating module is used for weighting sentence characteristics in the updated version text heterogeneous graph respectively based on the classification weight of the sentence vector, acquiring corresponding sentence labels based on the weighted sentence characteristics and generating a text abstract according to the acquired sentence labels;

wherein constructing a text heterogeneous graph of the target text based on the word features and sentence features comprises:

based on sentence characteristics of sentence vectors in the target text, calculating the homogeneous edge weight between every two sentence vectors of all sentence vectors in the target text in a cosine similarity calculation mode;

calculating heterogeneous edge weights among all word vectors and sentence vectors to which the word vectors belong in the target text through a TF-IDF algorithm based on the word features of the word vectors in the target text and the sentence features of the sentence vectors to which the word vectors belong;

taking word vectors in the target text as word nodes and sentence vectors in the target text as sentence nodes, constructing a text heterogeneous graph of the target text based on word features of the word vectors, sentence features of the sentence vectors, homogeneous edge weights among the sentence vectors and heterogeneous edge weights among the word vectors and the sentence vectors to which the word vectors belong,

and the multiclass abstract indexes of the sentence vector comprise a relevance score, a redundancy score, a new information score and a recall-oriented evaluation metric score,

updating the text heterogeneous graph through a graph attention network based on the edge weight and the attention weight, and acquiring an updated text heterogeneous graph, wherein the step of acquiring the updated text heterogeneous graph comprises the following steps:

calculating attention weights between every two sentence vectors of all sentence vectors in the target text, calculating attention weights between all word vectors in the target text and the sentence vectors to which the word vectors belong, and acquiring all homogeneous edge weights and all heterogeneous edge weights in the target text;

updating all word nodes and all sentence nodes in the text heterogeneous graph through a graph attention network based on the attention weight between every two sentence vectors in the target text, the attention weight between all word vectors and the sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights to obtain an updated version text heterogeneous graph,

wherein, based on the attention weights between every two sentence vectors of all sentence vectors in the target text, the attention weights between all word vectors and the sentence vectors to which the word vectors belong, all homogeneous edge weights and all heterogeneous edge weights, updating all word nodes and all sentence nodes in the text heterogeneous graph through a graph attention network comprises:

taking word nodes as central nodes, taking the product of attention weight between sentence nodes connected with the central nodes and heterogeneous edge weight between the sentence nodes connected with the central nodes as weight, and performing weighted aggregation on sentence characteristics of the sentence nodes connected with the central nodes to realize the updating of the word nodes;

taking sentence nodes as central nodes, taking the product of attention weight between word nodes connected with the central nodes and heterogeneous edge weight between the word nodes connected with the central nodes as weight, and performing weighted aggregation on the word characteristics of the word nodes connected with the central nodes to realize the updating of the sentence nodes;

and taking sentence nodes as central nodes, taking the product of attention weight between the sentence nodes connected with the central nodes and homogeneous edge weight between the sentence nodes connected with the central nodes as weight, and carrying out weighted aggregation on the sentence characteristics of the sentence nodes connected with the central nodes to realize the update of the sentence nodes.

6. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements the method for text summarization based on heterogeneous maps according to any one of claims 1 to 4.

7. A terminal, comprising: a processor and a memory;

the memory is used for storing computer programs, and the processor is used for executing the computer programs stored by the memory so as to enable the terminal to execute the text summarization method based on heterogeneous graphs according to any one of claims 1 to 4.

Images