A paraphrase sentence recognition method and system based on semantic primitive knowledge and abstract semantic representation

Abstract

The application belongs to the field of natural language processing, and particularly relates to a method and system for paraphrase recognition based on semantic primitive knowledge and abstract semantic representation, which comprises the following steps: performing word segmentation on a sentence, and performing word-level vector representation and semantic primitive knowledge representation; performing mean value processing on the semantic primitive knowledge representation result, and extracting interactive attention feature information of the mean value processing result by using global semantic information to obtain global semantic primitive representation; performing abstract semantic analysis on a to-be-recognized paraphrase sentence from a sentence structure to obtain a single-root directed acyclic graph, and performing global semantic primitive representation and word-level vector representation; extracting global and local feature information in the order of the directed acyclic graph, and performing distance feature measurement on the information; inputting the distance feature measurement result into a neural network to obtain a recognition result; the application introduces external semantic primitive knowledge to perform semantic representation, the accuracy of the semantic primitive knowledge representation is assisted by global semantic information, and the abstract semantics of a Chinese paraphrase sentence is analyzed to obtain semantic relations, so that the accuracy of paraphrase recognition is improved.

Description

Technical Field

[0001] The invention belongs to the field of natural language processing, and particularly relates to a paraphrase sentence recognition method and system based on sememe knowledge and abstract semantic representation. Background Art

[0002] Paraphrase recognition belongs to a subtask related to semantic matching in the field of natural language processing. Its purpose is to identify whether the given sentence pairs are paraphrases / paraphrases (that is, express the same meaning). Tasks associated with paraphrase recognition include question answering, information extraction, similarity calculation, etc. Paraphrase recognition involves semantic representation and matching of sentence pairs. Therefore, technical methods of semantic representation and matching are often applied to paraphrase recognition. Given the characteristics of paraphrase sentence pairs such as synonym replacement or syntactic structure changes, current common paraphrase recognition methods start from synonym replacement or syntactic structure changes to extract the information differences between sentence pairs for recognition. In the semantic representation stage of sentence pairs, large-scale pre-trained distributed word vectors are used for semantic representation fusion. Different model pre-trained word vectors are used for word representation to solve the synonym replacement problem. Different interaction methods are used in the model structure of neural networks for feature extraction to solve syntactic structure changes. However, using pre-trained word vectors in the semantic representation stage cannot well solve the synonym replacement problem. That is, the word vectors between synonyms cannot fully represent their similarity, and it will bring errors in subsequent semantic interaction calculations, resulting in unreasonable paraphrase recognition results.

[0003] Sememe knowledge is a language and common sense knowledge base used to describe the concepts represented by words themselves, and a sememe is the smallest unit for describing a concept; words expressing the same semantics or similar semantics have overlapping sememe parts, and these overlapping sememes determine the similar parts between words. For example: "moisture" and "boast". Among them, "moisture" includes the semantic item "exaggeration", and "exaggeration" includes sememes: "information" and "exaggerate". Among them, "boast" includes the semantic item "boast", and "boast" includes sememes: "exaggerate". The words "moisture" and "boast" both contain "exaggerate", indicating their similarity. Therefore, for synonyms, using sememe knowledge to represent them can theoretically obtain a higher semantic representation effect.

[0004] Attention neural networks are neural networks that incorporate attention mechanisms. Originally used in image processing, attention mechanisms differ from convolutional neural networks and long short-term memory networks in that they primarily simulate the different focuses of human vision when browsing images. Inspired by the applications of attention in image processing, some researchers have applied attention mechanisms to natural language processing to extract "important information" from sentence structure. This process can be categorized into self-attention mechanisms and cross-attention mechanisms based on whether or not it interacts with external information. The difference lies in whether or not external information is "attended." Research indicates that attention mechanisms can extract more textual information features. Furthermore, considering the semantic primitive knowledge interaction between sentence pairs, cross-attention mechanisms are suitable for extracting feature information from semantic primitive knowledge.

[0005] Traditional paraphrase recognition suffers from poor accuracy due to the semantic complexity of certain words in a sentence. Furthermore, it only recognizes words of the same level, resulting in low accuracy and hindering sentence processing. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes a paraphrasing recognition method based on semantic primitives knowledge and abstract semantic representation. This method includes:

[0007] S1: Obtain the paraphrased statement to be identified, perform word segmentation on the paraphrased statement, and obtain the semantic knowledge of each word segment;

[0008] S2: Represent the word segmentation and semantic primitive knowledge into vector representations to obtain lexical-level vector representations and semantic primitive vector representations;

[0009] S3: Obtain the global semantic information of the paraphrased statement to be identified, perform mean processing on the lexical-level semantic primitive vector representation and fuse the global semantic information to obtain the global semantic primitive vector representation;

[0010] S4: Perform abstract semantic parsing on the paraphrased statement to be identified to obtain a single-rooted directed acyclic graph. Use global semantic vectors and lexical-level vectors to represent the single-rooted directed acyclic graph, respectively.

[0011] S5: Based on the order of the single-rooted directed acyclic graph, a self-attention mechanism is used to extract word-level local feature information from the word nodes in the single-rooted directed acyclic graph; a bidirectional gated recurrent neural network is used to extract sentence-level global feature information from the word nodes in the single-rooted directed acyclic graph.

[0012] S6: Fuse all extracted local feature information, fuse all global feature information, and then concatenate the fused word-level local feature information and the fused sentence-level global feature information to obtain the recognition result.

[0013] Preferably, the process of representing word segmentation and semantic primitive knowledge into vectors includes: representing word segmentation using pre-trained word vectors to obtain lexical-level vector representations; parsing semantic primitive knowledge to obtain semantic primitive parsing results for each character or word; the semantic primitive parsing results are a three-dimensional matrix, where the first dimension represents the sentence segmentation length, the second dimension represents the number of semantic primitives for each word, and the third dimension represents the vector representation of each semantic primitive; the word segmentation representation is a two-dimensional matrix, where the first dimension represents the sentence segmentation length and the second dimension represents the vector representation of each word.

[0014] Preferably, the process of parsing semantic primitives includes: initializing an instance object hownet, and obtaining all semantic primitives parsing results of the target word according to the attribute methods carried by the instance object hownet; wherein the semantic primitives parsing results of a word contain multiple senses, each sense contains multiple semantic primitives, and all semantic primitives of the word are used as the semantic primitives set of the word.

[0015] Preferably, the process of averaging and fusing global semantic information on lexical-level semantic primitive vector representations includes:

[0016] Step 1: Use a bidirectional gated recurrent unit neural network to extract global semantic information from the paraphrased statement to be recognized;

[0017] Step 2: Calculate the mean of the lexical-level vector representations using the mean method to obtain the semantic primitive vector representation of a single word;

[0018] Step 3: Use an interactive attention mechanism to fuse global semantic information and the semantic primitive vector representation of individual words to obtain the global semantic primitive vector representation.

[0019] Preferably, the process of abstracting and semantically parsing the feature information of global semantic primitive vector representation and lexical-level vector representation includes:

[0020] Step 1: Use natural language processing tools to perform abstract semantic analysis on the sentence to obtain a single-rooted directed acyclic graph of the sentence;

[0021] Step 2: Use the feature information represented by the global semantic primitive vector to perform semantic representation on the single-rooted directed acyclic graph, and obtain the single-rooted directed acyclic graph represented by the global semantic primitive vector; use the feature information represented by the lexical vector to perform semantic representation on the single-rooted directed acyclic graph, and obtain the single-rooted directed acyclic graph represented by the lexical vector.

[0022] Preferably, the process of obtaining lexical-level local feature information and sentence-level global feature information includes: extracting local and global feature information from both the single-rooted directed acyclic graph (DAG) represented by the global semantic primitive vector and the single-rooted DAG represented by the lexical-level vector. A self-attention mechanism is used to calculate the query vector Q, value vector K, and attribute value vector V of the word order in the acyclic graph to obtain lexical-level local feature information; the words in the acyclic graph are then input into a bidirectional gated recurrent neural network in the specified order to obtain sentence-level global feature information. The final results include: global feature information represented by the global semantic primitive vector, local feature information represented by the global semantic primitive vector, global feature information represented by the lexical-level vector, and local feature information represented by the lexical-level vector.

[0023] Preferably, the process of obtaining the concatenated semantic features includes:

[0024] Step 1: Use the Manhatan and Euclidean functions to fuse all word-level local feature information to obtain fused word-level local feature information;

[0025] Step 2: Use cosine, difference, and multiplication to fuse all sentence-level global feature information to obtain fused sentence-level global feature information;

[0026] Step 3: Use a feedforward neural network to concatenate and fuse the fused word-level local feature information and the fused sentence-level global feature information to obtain the recognition result.

[0027] 10. A paraphrasing recognition system based on semantic primitive knowledge and abstract semantic representation, the system comprising: a text acquisition unit, a text segmentation unit, a text representation unit, a semantic primitive knowledge fusion unit, an abstract semantic representation unit, a local feature extraction unit, a global feature extraction unit, a feature fusion unit, and a feedforward neural network unit;

[0028] The text acquisition unit is used to acquire the paraphrased statement to be detected and send it to the text segmentation unit;

[0029] The text segmentation unit is used to segment the paraphrased statement to obtain the words in the paraphrased statement;

[0030] The text representation unit is used to represent the words after word segmentation, wherein word-level pre-trained word vectors and semantic primitive-level pre-trained vectors are used for representation respectively.

[0031] The semantic primitive knowledge fusion unit performs global semantic primitive vector representation on the lexical-level vector representation through external semantic primitive knowledge.

[0032] The abstract semantic representation unit is used to perform abstract semantic parsing on the vector results of the fused semantic primitive knowledge representation, and is represented by global semantic primitive vectors and lexical vectors respectively.

[0033] The local feature extraction unit is used to extract local semantic features based on the single-rooted directed acyclic graph obtained by the abstract semantic representation unit.

[0034] The global feature extraction unit is used to extract global feature information based on the single-rooted directed acyclic graph obtained by the abstract semantic representation unit.

[0035] The feature fusion unit is used to concatenate local and global features, and uses a multi-angle distance metric formula to measure the spatial characteristics of the concatenated semantic features.

[0036] The feedforward neural network unit is used to perform spatial paraphrasing and recognition of semantic features to obtain recognition results.

[0037] Preferably, the semantic primitive knowledge fusion unit performs global semantic primitive vector representation on the lexical vector representation in two parts: the first part is to perform mean processing on the semantic primitive vector representation of the lexical to obtain the semantic primitive representation of a single lexical; the second part is to use the global semantic information of the current sentence to perform fusion to obtain the global semantic primitive vector representation.

[0038] The beneficial effects of this invention are:

[0039] This invention employs semantic primitive knowledge and abstract semantic representation to identify paraphrased sentences, improving recognition accuracy. By introducing external semantic primitive knowledge to represent words, this invention performs "semantic representation" of words from a semantic concept perspective, minimizing the problem of polysemy. Furthermore, this invention introduces external semantic primitive knowledge to represent words and provides two methods for processing semantic primitive knowledge representation (semantic primitive mean method and global context semantic primitive attention method), enabling the autonomous selection of appropriate semantic primitive knowledge based on global semantic information to the greatest extent possible, providing better semantic representation for words. Finally, this invention extracts feature information from paraphrased sentence pairs at different dimensions, including word and sentence levels, enabling the capture of information from different dimensions to achieve higher paraphrasing recognition accuracy. Attached Figure Description

[0040] Figure 1 This is a flowchart of a method for paraphrasing recognition based on semantic primitive knowledge and abstract semantic representation in this invention;

[0041] Figure 2 This is a schematic diagram of a paraphrase recognition system based on semantic primitives and abstract semantic representation in this invention;

[0042] Figure 3This is a schematic diagram of a method for representing vocabulary using the semantic primitive mean method according to the present invention;

[0043] Figure 4 This is a schematic diagram of a method for representing vocabulary through global information semantic primitive attention according to the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] A paraphrasing recognition method based on semantic primitives knowledge and abstract semantic representation, such as Figure 1 As shown, the method includes:

[0046] S1: Obtain the paraphrased statement to be identified, perform word segmentation on the paraphrased statement, and obtain the semantic knowledge of each word segment;

[0047] S2: Represent the word segmentation and semantic primitive knowledge into vector representations to obtain lexical-level vector representations and semantic primitive vector representations;

[0048] S3: Obtain the global semantic information of the paraphrased statement to be identified, perform mean processing on the lexical-level semantic primitive vector representation and fuse the global semantic information to obtain the global semantic primitive vector representation;

[0049] S4: Perform abstract semantic parsing on the paraphrased statement to be identified to obtain a single-rooted directed acyclic graph, and use global semantic vectors and lexical-level vectors to represent the single-rooted directed acyclic graph respectively;

[0050] S5: Based on the order of the single-rooted directed acyclic graph, a self-attention mechanism is used to extract word-level local feature information from the word nodes in the single-rooted directed acyclic graph; a bidirectional gated recurrent neural network is used to extract sentence-level global feature information from the word nodes in the single-rooted directed acyclic graph.

[0051] S6: Fuse all extracted local feature information, fuse all global feature information, and then concatenate the fused word-level local feature information and the fused sentence-level global feature information to obtain the recognition result.

[0052] This embodiment will combine the LCQMC dataset (including but not limited to this dataset) to... Figure 1 The steps described are explained in detail. The LCQMC dataset is used to identify whether a given pair of sentences expresses the same meaning; "1" indicates that the pair of sentences expresses the same meaning, and "0" indicates that the pair of sentences does not express the same meaning.

[0053] The word segmentation process for paraphrased sentences includes: first, using word segmentation tools (including but not limited to jieba segmentation) to segment the sentence pairs, ultimately forming word segmentation results for individual characters or words. The segmented results are first represented using pre-trained word vectors. Next, the semantic primitives parsing interface is called to obtain the semantic primitives parsing results for each segmented character or word. The semantic primitives parsing results for each character or word ignore semantic features and only consider the set of all semantic primitives for that word. Finally, each data sentence is represented by a three-dimensional matrix [10, 15, 200], where the first dimension represents the sentence segmentation length, the second dimension represents the number of semantic primitives for each word, and the third dimension represents the vector representation of each semantic primitive. The initial lexical representation is a two-dimensional matrix [10, 200], where the first dimension represents the sentence segmentation length, and the second dimension represents the vector representation of each word.

[0054] Semantic representation calculation is performed using lexical-level vector representations and semantic primitive knowledge vector representations. This content aims to eliminate the discrepancy between the three-dimensional results of semantic primitive representations and the two-dimensional results of lexical representations, considering that a word in a semantic primitive representation is obtained by concatenating multiple semantic primitive vectors, and the concatenation result lacks an understanding of the importance of semantic content. The semantic primitive mean method is adopted (…). Figure 3 ) and semantic primitive attention methods that rely on global information ( Figure 4 The semantic primitive features are extracted, and the final result is a two-dimensional matrix [10,200].

[0055] The representation result obtained after mean processing is feature information that encompasses the semantic representation of semantic primitives. The formula for calculating the mean of semantic primitives is as follows:

[0056] Sentence = [w1, w2, w3, ... w L1 ]

[0057]

[0058] Here, Sentence represents the word segmentation sequence obtained after word segmentation of a single sentence in the paraphrase recognition statement. The sequence length of each Sentence is extended to a specified length L1. The Sememes corresponding to a word w are summed (where L2 represents the L2 Sememes corresponding to each word), and the average of the summation results is used as the semantic primitive representation result of the current word w.

[0059] When global semantic vectors are used to assist the semantic primitive representation results, the semantic primitive knowledge of the word, guided by global semantic information, will select the appropriate semantic primitive knowledge representation under the current semantic context using the cross-attention method. A simplified diagram of this method is shown below. Figure 4As shown, the representation result obtained after cross-attention is semantic primitive knowledge feature information that covers global semantics. The formula for calculating semantic primitive attention under global information is as follows:

[0060] Sentence = [w1, w2, w3, ... w L1 ]

[0061] x i =W e w i (i∈[1,L1])

[0062] H i =B i -GRU(x i (i∈[1,L1])

[0063]

[0064] Here, Sentence represents the word segmentation sequence obtained after word segmentation of a single sentence in the paraphrase recognition statement. e Let x represent pre-trained distributed lexical vectors. i Indicates the current word w i In the pre-trained word vector W e The corresponding word vectors. H represents the global information of the sentence obtained by passing the word-level word vectors through a gated recurrent neural network Bi-GRU. The semantic primitive representation vectors are used as Q and V in the attention formula, and the semantic global information H is used as K(d) in the attention formula. k (This represents the dimension of the word vector). The final result obtained from attention, Sememe, represents the semantic primitive features of the current sentence under the global information of all words.

[0065] In this embodiment, the AMR parsing tool is used to parse the global semantic primitive knowledge feature information to obtain the single-root directed acyclic graph of the sentence. Semantic primitive knowledge fusion can carry additional semantic knowledge to the word segmentation results. Using AMR parsing will obtain the single-root directed acyclic graph of the sentence. The single root guarantees the main body of the sentence, while the direction covers the semantic relationships between the sentence entities.

[0066] Word-level and sentence-level semantic information extraction is performed on single-rooted directed acyclic graphs (DAGs). A self-attention neural network is used to extract local semantic information at the word level from all entity labels in the single-rooted DAG. A bidirectional gated recurrent unit neural network is used to extract global semantic information at the sentence level from all entities and relationships between entities in the single-rooted DAG, and the output [10, 200] is used as global information for distance metric calculation (including but not limited to Euclidean distance, Manhattan distance, and cosine similarity).

[0067] The semantic information at the word and sentence levels is spatially semantically measured using Manhattan distance, Euclidean distance, and cosine similarity. The semantic distance measurement results are used as input to the network prediction layer, and finally a binary classification result is output to represent the result of the paraphrase prediction.

[0068] The process of parsing semantic primitives includes: first, installing the dependency package OpenHowNet in the Python environment and initializing an instance object hownet using the OpenHowNet.HowNetDict() method; next, using the get_sememes_by_word() method of the hownet object to obtain all semantic primitive parsing results of the target word; finally, considering that the semantic primitives of a word can contain multiple senses, and senses contain multiple semantic primitives, the senses of the word are ignored, and all the semantic primitives of the word are taken as the semantic primitives of the word for subsequent processing. Following the steps above, to perform semantic primitive analysis on the word "apple": Using `hownet.get_sememes_by_word("apple")` yields a list. By default, each item in the list is a dictionary object, representing the meaning of the word "apple" in different contexts (meanings can be fruit, brand). This dictionary object contains multiple semantic elements, indicating which semantic elements can represent the current meaning (e.g., computer = carry + brand + ability). We ignore all meanings and extract all the semantic elements of "apple" to form our semantic primitive knowledge. That is, the current meaning of "apple" can be represented by (ability, carry, computer, fruit, tree...).

[0069] A paraphrasing recognition system based on semantic primitives knowledge and abstract semantic representation, such as Figure 2 As shown, the system includes: a text acquisition unit, a text segmentation unit, a text representation unit, a semantic primitive knowledge fusion unit, an abstract semantic representation unit, a word-level self-attention feature extraction unit, a sentence-level feature extraction unit, a feature fusion unit, and a feedforward neural network unit;

[0070] The text acquisition unit is used to acquire the paraphrased statement to be detected and send it to the text segmentation unit;

[0071] The text segmentation unit is used to segment the paraphrased statement to obtain sub-words or phrases in the paraphrased statement;

[0072] The text representation unit is used to represent words or phrases after word segmentation, wherein word-level pre-trained word vectors and semantic primitive-level pre-trained vectors are used to represent different dimensions respectively.

[0073] The semantic primitive knowledge fusion unit fuses the initial word-level representation with external semantic primitive knowledge;

[0074] In this embodiment, the semantic primitive fusion module in the system mainly includes two parts: word-level representation and semantic primitive representation. The word-level representation simply uses pre-trained word vectors for representation, while the semantic primitive representation vector involves whether the auxiliary knowledge of global semantic vectors is needed.

[0075] The abstract semantic representation unit is used to perform abstract semantic parsing on the vector results of the fused semantic primitive knowledge representation;

[0076] The local feature extraction unit is used to extract semantic features of entities based on the single rooted directed acyclic graph obtained by the abstract semantic representation unit.

[0077] The global feature extraction unit extracts high-dimensional sentence-level global information based on the results of its abstract semantic representation unit, and uses a bidirectional gated recurrent unit neural network to extract semantic features of a single rooted directed acyclic graph according to the direction between entities.

[0078] The feature fusion unit is used to concatenate word-level semantic features and sentence-level semantic features, and uses a multi-angle distance metric formula to measure the space of the concatenated semantic features;

[0079] The feedforward neural network unit is used to perform spatial paraphrasing and recognition of semantic features to obtain recognition results.

[0080] In one embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function.

[0081] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor, which can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.

[0082] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recognizing paraphrased statements based on semantic primitives and abstract semantic representations, characterized in that, include: S1: Obtain the paraphrased statement to be identified, perform word segmentation on the paraphrased statement, and obtain the semantic knowledge of each word; S2: Represent the vocabulary and semantic primitives into vectors to obtain the vocabulary-level vector representation and the vocabulary-level semantic primitive vector representation. S3: Obtain the global semantic information of the paraphrased statement to be identified, perform mean processing on the lexical-level semantic primitive vector representation and fuse the global semantic information to obtain the global semantic primitive vector representation; S4: Perform abstract semantic parsing on the paraphrased statement to be identified to obtain a single-rooted directed acyclic graph, and use global semantic vectors and lexical-level vectors to represent the single-rooted directed acyclic graph respectively; S5: Extract lexical-level local feature information from lexical nodes in the single-rooted directed acyclic graph by using a self-attention mechanism according to the order of the single-rooted directed acyclic graph. A bidirectional gated recurrent neural network is used to extract sentence-level global feature information from lexical nodes in a single-rooted directed acyclic graph (DAG). Specifically, this includes: extracting local and global feature information from both the DAG represented by global semantic primitives and the DAG represented by lexical vectors; using a self-attention mechanism to calculate the query vector Q, value vector K, and attribute value vector V of the word order in the acyclic graph to obtain lexical-level local feature information; and inputting the words in the acyclic graph into the bidirectional gated recurrent neural network in the specified order to obtain sentence-level global feature information. The final results include: global feature information represented by global semantic primitives, local feature information represented by global semantic primitives, global feature information represented by lexical vectors, and local feature information represented by lexical vectors. S6: All extracted word-level local feature information is fused, all sentence-level global feature information is fused, and the fused word-level local feature information and the fused sentence-level global feature information are concatenated to obtain the recognition result; specifically including: Step 1: Use the Manhatan and Euclidean functions to fuse all local feature information to obtain fused word-level local feature information; Step 2: Use cosine, difference, and multiplication to fuse all global feature information to obtain fused sentence-level global feature information; Step 3: Use a feedforward neural network to concatenate and fuse the fused word-level local feature information and the fused sentence-level global feature information to obtain the recognition result.

2. The paraphrasing recognition method based on semantic primitives and abstract semantic representation according to claim 1, characterized in that, The process of representing word segmentation results and semantic primitives into vectors includes: representing words using pre-trained word vectors to obtain word-level vector representations; parsing semantic primitives to obtain semantic primitive parsing results for each character or word; the semantic primitive parsing results are a three-dimensional matrix, where the first dimension represents the word segmentation length of the sentence, the second dimension represents the number of semantic primitives for each word, and the third dimension represents the vector representation of each semantic primitive; the word segmentation representation is a two-dimensional matrix, where the first dimension represents the word segmentation length of the sentence, and the second dimension represents the vector representation of each word.

3. The paraphrasing recognition method based on semantic primitives and abstract semantic representation according to claim 2, characterized in that, The process of parsing semantic primitives includes: initializing the instance object hownet, obtaining all semantic primitives parsing results of the target word according to the attributes and methods carried by the instance object hownet; wherein the semantic primitives parsing results of a word contain multiple semantic terms, each semantic term contains multiple semantic primitives, and all semantic primitives of the word are used as the semantic primitives set of the word.

4. The paraphrasing recognition method based on semantic primitives and abstract semantic representation according to claim 1, characterized in that, The process of mean averaging and global semantic information fusion of lexical-level semantic primitive vector representations includes: Step 1: Use a bidirectional gated recurrent unit neural network to extract global semantic information from the paraphrased statement to be recognized; Step 2: Calculate the mean of the semantic primitive vector representations at the lexical level using the mean method to obtain the semantic primitive vector representations of individual words. Step 3: Use an interactive attention mechanism to fuse global semantic information and the semantic primitive vector representation of individual words to obtain the global semantic primitive vector representation.

5. The paraphrasing recognition method based on semantic primitives and abstract semantic representation according to claim 1, characterized in that, The process of abstracting and representing the paraphrased statement to be identified includes: Step 1: Use natural language processing tools to perform abstract semantic analysis on the sentence to obtain a single-rooted directed acyclic graph of the sentence; Step 2: Use the feature information represented by the global semantic primitive vector to perform semantic representation on the single-rooted directed acyclic graph, and obtain the single-rooted directed acyclic graph represented by the global semantic primitive vector; use the feature information represented by the lexical vector to perform semantic representation on the single-rooted directed acyclic graph, and obtain the single-rooted directed acyclic graph represented by the lexical vector.

6. A paraphrase recognition system based on semantic primitives and abstract semantic representation, the system being used to execute the paraphrase recognition method based on semantic primitives and abstract semantic representation as described in any one of claims 1 to 5, the system comprising: The system includes a text acquisition unit, a text segmentation unit, a text representation unit, a semantic primitive knowledge fusion unit, an abstract semantic representation unit, a local feature extraction unit, a global feature extraction unit, a feature fusion unit, and a feedforward neural network unit. The text acquisition unit is used to acquire the paraphrased statement to be detected and send it to the text segmentation unit; The text segmentation unit is used to segment the paraphrased statement to obtain the words in the paraphrased statement; The text representation unit is used to represent the words after word segmentation, wherein word-level pre-trained word vectors and semantic primitive-level pre-trained vectors are used for representation respectively. The semantic primitive knowledge fusion unit performs global semantic primitive vector representation on the lexical-level vector representation through external semantic primitive knowledge. The abstract semantic representation unit is used to perform abstract semantic parsing on the vector results of the fused semantic primitive knowledge representation, and is represented by global semantic primitive vectors and lexical vectors respectively. The local feature extraction unit is used to extract local semantic features based on the single-rooted directed acyclic graph obtained by the abstract semantic representation unit. The global feature extraction unit is used to extract global feature information based on the single-rooted directed acyclic graph obtained by the abstract semantic representation unit. The feature fusion unit is used to concatenate local and global features, and uses a multi-angle distance metric formula to measure the spatial characteristics of the concatenated semantic features. The feedforward neural network unit is used to perform spatial paraphrasing and recognition of semantic features to obtain recognition results.

7. The paraphrasing recognition system based on semantic primitive knowledge and abstract semantic representation according to claim 6, wherein the semantic primitive knowledge fusion unit performs global semantic primitive vector representation of lexical vector representation in a manner comprising two parts, the first part being to perform mean processing on the semantic primitive vector representation of lexical to obtain the semantic primitive representation of a single lexical. The second part uses the global semantic information of the current statement to fuse and obtain the global semantic primitive vector representation.

Images