Product entity recognition method and system integrating product domain knowledge

Abstract

The present invention provides a product entity recognition method, system, storage medium and electronic device that integrate product domain knowledge, and relates to the field of natural language processing technology. The present invention constructs an index set of product entities in a vocabulary based on product domain knowledge, and obtains character-level candidate product entities corresponding to each character; according to the entity sentence to be recognized, a candidate product entity reference graph is constructed to obtain the global dependency relationship between any two character-level candidate product entities, and obtain the evidence score that each character-level candidate product entity is the reference entity of the entity sentence to be recognized; according to the entity sentence to be recognized, each character-level candidate product entity and its corresponding embedding weight, a joint embedding vector is obtained, and used as the input of a pre-built deep learning model to obtain the entity recognition result. The global dependency relationship of entities can be embedded into the deep learning model, which expands the information source of the model, thereby enhancing its ability to handle Chinese product entity recognition problems.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a product entity recognition method, system, storage medium and electronic device that integrates product domain knowledge. Background Art

[0002] The development of the market has led to an increasing variety of product types, which has also led to the diversity of product entities mentioned in user-generated text data.

[0003] Not only do different texts mention different product entities, but the same text may also mention multiple product entities at the same time. When describing these product entities, users not only use proper nouns in the field, but also use various irregular expressions such as abbreviations and aliases. As a result, when generating character-level candidate product entities, some character sequences representing product entities cannot be linked to the vocabulary. Secondly, when embedding character-level candidate product entities into deep learning models, the global dependencies between candidate product entities are not taken into account, that is, product entities in the same text should be semantically related.

[0004] Although some existing methods use attention mechanisms to learn the weights of candidate product entities, they are still limited to the information of the sentence itself without any additional input, thus ignoring the global dependencies between candidate product entities. Therefore, to address the above problems, it is necessary to propose a product entity recognition method that integrates product domain knowledge. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the deficiencies of the prior art, the present invention provides a product entity recognition method, system, storage medium and electronic device that integrate product domain knowledge, solving the technical problem of ignoring the global dependencies between candidate product entities.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0009] A product entity recognition method integrating product domain knowledge, comprising:

[0010] S1. Build an index set of product entities in the vocabulary based on product domain knowledge;

[0011] S2. Obtain, based on the index set, character-level candidate product entities corresponding to each character of the entity sentence to be identified;

[0012] S3. Based on the entity sentence to be identified, a candidate product entity reference graph is constructed to obtain the global dependency relationship between any two character-level candidate product entities;

[0013] S4. Obtaining, based on each of the global dependencies, an evidence score for each of the character-level candidate product entities being a referent entity of the entity sentence to be identified;

[0014] S5. Obtain a joint embedding vector based on the entity statement to be identified, each of the character-level candidate product entities and their corresponding embedding weights, and use it as input to a pre-built deep learning model to obtain an entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities.

[0015] Preferably, the index construction in S1 includes a word-based index and a pronunciation-based index.

[0016] Preferably, in S2, a character-level candidate product entity generation algorithm is used to obtain the character-level candidate product entity, and the algorithm includes:

[0017] S10, obtain the entity statement to be identified S = {a1, ..., a i ,…a l}, vocabulary D, index set IS, the maximum length of the index in the index set maxLen; where a i Represents the i-th character of sentence S, with a total of l characters;

[0018] S20. Initialize the character-level candidate product entity set

[0019] S30, set i=1;

[0020] S40, initializing the sliding window length ω=1;

[0021] S50, if ω≤maxLen, go to step S60; otherwise, determine whether i is greater than 1. If so, end the algorithm; otherwise, set i=i+1 and go to step S40;

[0022] S60, set the character sequence to chara_seq=a i a i+1 …a i+ω-1 ; If chara_seq is in the index set IS, then find the corresponding entity w in the vocabulary D and put it into the character-level candidate product entity set CE;

[0023] S70. Set ω=ω+1 and go to step S50.

[0024] Preferably, the process of constructing the candidate product entity reference graph in S3 includes:

[0025] Define nodes as two entities: product brand and product type;

[0026] Defining the relationships of a node includes:

[0027] (1) The relationship between product brand and product type

[0028]

[0029] in, Indicates the affiliation between product type product1 and product brand brand1;

[0030] (2) Relationships between product types

[0031]

[0032] in, Indicates the price similarity between any two product types product1 and product2, price1 and price2 are the prices of product1 and product2 respectively;

[0033]

[0034] in, Indicates the size similarity between any two product types, product1 and product2. l1, g1, and 91 represent the length, width, and height of product1 respectively. l2, g2, and 92 represent the length, width, and height of product2 respectively.

[0035] (3) Relationship between product brands

[0036]

[0037] in, Represents the relationship between any two product brands brand1 and brand2, product 1，x Indicates the xth product type of brand1, product 2，y Indicates the y-th product type of brand2; X indicates that it belongs to brand i The number of product types, Y is the brand j The number of product types.

[0038] Preferably, the weighted PageRank algorithm is used in S4 to obtain the evidence score, including:

[0039] S41. Use entity popularity as the initial PR value of each node in the candidate product entity reference graph.

[0040]

[0041] in, Represents the selected product portal or user forum about any product entity w i Number of posts;

[0042] S42. Calculate the PR value of each character-level candidate product entity.

[0043]

[0044] in, Respectively represent the character-level candidate product entity w i 、w j All relevant candidate product entities, Indicates w j to w i , α is the weight factor for balancing the contribution of the edge and product entity relationship; N represents the number of candidate product entities;

[0045] S43, obtain evidence score,

[0046]

[0047] in, Represents the character-level candidate product entity w i The evidence score of the entity to be identified in the entity statement.

[0048] Preferably, the process of obtaining the joint embedding vector in S5 includes:

[0049] S100, BERT-based character embedding

[0050] According to the entity statement to be identified S={a1,…,a i ,…a l}, the formula for generating character vectors using the pre-trained BERT model,

[0051]

[0052] in, Indicates a i A character vector of , d1 represents the embedding dimension of the character vector;

[0053] S200, word embedding based on character-level candidate product entities, including:

[0054] The word2vec model is used to vectorize each character-level candidate product entity.

[0055]

[0056] Among them, wi，j Represents the character a i The j-th character-level candidate product entity, Word2Vec w represents the word embedding lookup table, Indicates w i，j word vectors;

[0057] By weighting the embedding weight of each character-level candidate product entity, the word vector corresponding to each character is obtained.

[0058]

[0059] in, Represents the character a i The j-th character-level candidate product entity is the evidence score of the referent entity of the entity sentence to be identified, as w i，j The embedding weight of k i is the character a i The number of candidate product entities; Represents the character a i The weighted word vectors of all character-level candidate product entities,

[0060] S300, Combination of Character Embedding and Word Embedding

[0061]

[0062] Among them, x i Represents the character a i The joint embedding vector of γ∈(0, 1) is a hyperparameter used to balance the importance of character vectors and word vectors.

[0063] Preferably, the deep learning model adopts a Bi-LSTM-CRF network.

[0064] A product entity recognition system that integrates product domain knowledge, including:

[0065] The index building module is used to build an index set of product entities in the vocabulary based on product domain knowledge;

[0066] An entity acquisition module, which acquires character-level candidate product entities corresponding to each character of the entity sentence to be identified according to the index set;

[0067] The relationship acquisition module is used to construct a candidate product entity reference graph based on the entity sentence to be identified, and obtain the global dependency relationship between any two character-level candidate product entities;

[0068] A score acquisition module, configured to obtain, based on each of the global dependencies, an evidence score for each of the character-level candidate product entities being a referent entity of the entity sentence to be identified;

[0069] A joint embedding module is used to obtain a joint embedding vector based on the entity statement to be identified, each of the character-level candidate product entities and their corresponding embedding weights, and use it as the input of a pre-built deep learning model to obtain the entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities.

[0070] A storage medium stores a computer program for product entity recognition integrating product domain knowledge, wherein the computer program enables a computer to execute the product entity recognition method as described above.

[0071] An electronic device, comprising:

[0072] one or more processors;

[0073] Memory; and

[0074] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including a method for executing the product entity recognition method integrating product domain knowledge as described above.

[0075] (3) Beneficial effects

[0076] The present invention provides a product entity recognition method, system, storage medium, and electronic device that integrates product domain knowledge. Compared with existing technologies, it has the following advantages:

[0077] The product entity recognition method proposed in the present invention, which integrates product domain knowledge, can embed the global dependency of entities into the deep learning model, expand the information source of the model, and thus enhance its ability to handle Chinese product entity recognition problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0079] Figure 1 This is an overall framework diagram of a product entity recognition method that integrates product domain knowledge, provided by an embodiment of the present invention;

[0080] Figure 2A block diagram of a product entity recognition method integrating product domain knowledge provided by an embodiment of the present invention;

[0081] Figure 3 A schematic diagram of a joint embedded deep learning model architecture provided by an embodiment of the present invention;

[0082] Figure 4 An example diagram for constructing a product entity reference diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0083] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0084] The embodiments of the present application solve the technical problem of ignoring the global dependencies between candidate product entities by providing a product entity recognition method, system, storage medium and electronic device that integrate product domain knowledge.

[0085] The technical solution in the embodiments of the present application is to solve the above technical problems, and the overall idea is as follows:

[0086] like Figure 1 As shown, the product entity recognition method for integrating product domain knowledge provided by the embodiments of the present invention mainly includes two parts: 1) constructing a candidate entity reference graph and 2) jointly embedding a deep learning model. It can effectively solve the problems raised in the background technology and can be used for entity recognition at the enterprise, product, and technology levels in specific fields. That is, it can be applied to various sub-industries, including but not limited to the engine, automotive, and medical fields.

[0087] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0088] Example:

[0089] like Figure 2 As shown, an embodiment of the present invention provides a product entity recognition method integrating product domain knowledge, including:

[0090] S1. Build an index set of product entities in the vocabulary based on product domain knowledge;

[0091] S2. Obtain, based on the index set, character-level candidate product entities corresponding to each character of the entity sentence to be identified;

[0092] S3. Construct a candidate product entity reference graph based on the entity statement to be recognized, and obtain the global dependency relationship between any two character-level candidate product entities;

[0093] S4. According to each of the global dependency relationships, obtain the evidence score for each character-level candidate product entity to be the reference entity of the entity statement to be recognized;

[0094] S5. According to the entity statement to be recognized, each character-level candidate product entity and its corresponding embedding weight, obtain a joint embedding vector, and use it as the input of a pre-constructed deep learning model to obtain an entity recognition result; wherein, the embedding weight refers to the evidence score of each character-level candidate product entity.

[0095] The product entity recognition method integrating product domain knowledge proposed in the embodiments of the present invention can embed the global dependency relationship of entities into a deep learning model, expand the information source of the model, and thus enhance its ability to process Chinese product entity recognition problems.

[0096] Next, each step of the above technical solution will be introduced in detail:

[0097] In step S1, according to product domain knowledge, construct an index set of product entities in the vocabulary. The index construction in this step includes word-based indexing and pronunciation-based indexing, including:

[0098] 1) Word-based indexing. The composition of product brands is often relatively simple and fixed. Therefore, the index set can be directly constructed using the full Chinese and English names obtained from encyclopedia websites. Taking automobiles as an example, such as "BMW", "VOLVO", "BYD", etc. The index of product types is generated by designing a series of index generation rules, as shown in Table 1.

[0099] Table 1 Product index generation rules

[0100]

[0101] 2) Pronunciation-based indexing. When users describe product entities, they may incorrectly use other characters with the same pronunciation. For example, the word "欧蓝得" can be used to describe the product type entity "Outlander", where the pronunciation of "得" and "德" is the same, "dé". Therefore, it is necessary to expand the index set according to pronunciation. Specifically, the xpinyin Python package can be used to convert all Chinese characters in the word-based index into pinyin to generate a pronunciation-based index.

[0102] It is not difficult to understand that when constructing the candidate product entity reference graph, the embodiment of the present invention designs a series of word-based and pronunciation-based rules to expand the index range of product entities in the vocabulary and reduce the impact of irregular expressions in the text.

[0103] In step S2, character-level candidate product entities corresponding to each character of the entity sentence to be identified are obtained according to the index set.

[0104] This step aims to generate character-level candidate product entities based on the algorithm described below. The algorithm works as follows: if the character sequence consisting of the current character and its surrounding characters is in the index set, a candidate product entity can be generated based on the correspondence between the vocabulary and the index set. In this algorithm, a sliding window is used to sequentially generate character sequences. Specifically, the sliding window slides from left to right, generating all character sequences that contain the current character. The window length is initialized to 1 and gradually increases to generate character sequences of varying lengths.

[0105] Specifically, the algorithm includes:

[0106] S10, obtain the entity statement to be identified S = {a1, ..., a i ,…a l}, vocabulary D, index set IS, the maximum length of the index in the index set maxLen; where a i Represents the i-th character of sentence S, with a total of l characters;

[0107] S20. Initialize the character-level candidate product entity set

[0108] S30, set i=1;

[0109] S40, initializing the sliding window length ω=1;

[0110] S50, if ω≤maxLen, go to step S60; otherwise, determine whether i is greater than 1. If so, end the algorithm; otherwise, set i=i+1 and go to step S40;

[0111] S60, set the character sequence to chara_seq=a i a i+1 …a i+ω-1 ; If chara_seq is in the index set IS, then find the corresponding entity w in the vocabulary D and put it into the character-level candidate product entity set CE;

[0112] S70. Set ω=ω+1 and go to step S50.

[0113] In step S3, based on the entity sentence to be identified, a candidate product entity reference graph is constructed to obtain the global dependency relationship between any two character-level candidate product entities;

[0114] This step models the global dependencies between candidate product entities by constructing a candidate product entity reference graph for the entity sentence to be identified. The reference graph construction involves two elements: nodes and node relationships. In this embodiment of the present invention, nodes refer to two types of entities: product brand and product type. Node relationships include the following three types:

[0115] (1) The relationship between product brand and product type. Product brand and product type have a subordinate relationship. Therefore, this is a unidirectional relationship from product type to product brand. Given a product brand brand1 and a product type product1, the relationship from product1 to brand1 can be obtained by the following expression:

[0116]

[0117] (2) Relationships between product types. In a given entity statement to be identified, product types mentioned simultaneously often have a high degree of physical similarity, such as price and size, or functional substitutability, such as whether they have certain functions. The similarity of the above indicators can be weighted to obtain the relationship between product types. To simplify the problem, the embodiment of the present invention only uses physical similarity, namely price and size, to measure the relationship between product types. This is a bidirectional relationship.

[0118] Given two product types product1 and product2, the price similarity from product1 to product2 can be calculated by the following expression:

[0119]

[0120] Among them, price1 and price2 are the prices of product1 and product2 respectively.

[0121] The size range is measurable, and the effect of size difference on similarity can be considered a linear relationship. Therefore, the size similarity from product 1 to product 2 can be calculated according to the following expression:

[0122]

[0123] Among them, l1, g1, and 91 represent the length, width, and height of product1 respectively, and l2, g2, and 92 represent the length, width, and height of product2 respectively.

[0124] (3) The relationship between product brands. The relationship between product brands is reflected in the product types they contain and can be measured by summarizing the relationship between the product types contained in each product brand. Since each product type has a different impact on the product brand, and this impact is positively correlated with the popularity of the product type, the embodiment of the present invention uses the number of posts related to the product type on the product portal website or user forum to represent the popularity of the product type. Therefore, given two product brands brand1 and brand2, the relationship from brand1 to brand2 can be measured according to the following formula:

[0125]

[0126] Among them, product 1，x Indicates the xth product type of brand1, product 2，y Indicates the y-th product type of brand2; X indicates that it belongs to brand i The number of product types, Y is the brand j The number of product types.

[0127] In step S4, based on each of the global dependencies, an evidence score is obtained that each of the character-level candidate product entities is a referent entity of the entity sentence to be identified.

[0128] After constructing the product entity referent graph, we can calculate the evidence score for each candidate product entity as the referent entity of a given sentence. This evidence score is used as the embedding weight for each candidate product entity in the deep learning model in the subsequent step S5.

[0129] The embodiment of the present invention uses a weighted PageRank algorithm to measure the evidence score, including:

[0130] S41. Before using the weighted PageRank algorithm, an initial value needs to be assigned to each node. In this embodiment of the present invention, the entity popularity is used as the initial PageRank (PR) value, which is measured by the following expression:

[0131]

[0132] in, Represents the selected product portal or user forum about any product entity w i Number of posts;

[0133] S42. Calculate the PR value of each character-level candidate product entity.

[0134]

[0135] in, Respectively represent the character-level candidate product entity w i 、w j All relevant candidate product entities, represents w j to w i , α is the weight factor for balancing the contribution of the edge and product entity relationship; N represents the number of candidate product entities;

[0136] S43, obtain evidence score,

[0137]

[0138] in, Represents the character-level candidate product entity w i The evidence score of the entity to be identified in the entity statement.

[0139] In step S5, a joint embedding vector is obtained based on the entity statement to be identified, each of the character-level candidate product entities and their corresponding embedding weights, and is used as the input of a pre-built deep learning model to obtain an entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities.

[0140] The goal of this step is to embed the previously generated character-level candidate product entities and their weights into the deep learning model for Chinese product entity recognition. This part uses the classic Bi-LSTM-CRF framework. The core is the improvement of the embedding layer, which can effectively combine character vectors and word vectors. The detailed process of joint embedding is as follows Figure 3 As shown, the following will be introduced in detail.

[0141] The process of obtaining the joint embedding vector in this step includes:

[0142] S100, BERT-based character embedding

[0143] According to the entity statement to be identified S={a1,…,a i ,…a l}, the formula for generating character vectors using the pre-trained BERT model,

[0144]

[0145] in, Indicates a i A character vector of , d1 represents the embedding dimension of the character vector;

[0146] S200, word embedding based on character-level candidate product entities, including:

[0147] The word2vec model is used to vectorize each character-level candidate product entity.

[0148]

[0149] Among them, w i，j Represents the character a i The j-th character-level candidate product entity, Word2Vec w represents the word embedding lookup table, represents w i，j word vectors;

[0150] By weighting the embedding weight of each character-level candidate product entity, the word vector corresponding to each character is obtained.

[0151]

[0152] in, Represents the character a i The j-th character-level candidate product entity is the evidence score of the referent entity of the entity sentence to be identified, as w i，j The embedding weight of k i is the character a i The number of candidate product entities; Represents the character a i The weighted word vectors of all character-level candidate product entities,

[0153] S300, Combination of Character Embedding and Word Embedding

[0154] Character vectors and word vectors represent two different sources of information. The former represents the sentence itself, while the latter represents domain knowledge. These two vectors can be combined using the following expression and then applied to the Chinese product entity recognition task:

[0155]

[0156] Among them, x i Represents the character a i The joint embedding vector of γ∈(0,1) is a hyperparameter used to balance the importance of character vectors and word vectors.

[0157] Finally, the generated joint embedding vector representation is input into the Bi-LSTM-CRF network to obtain the entity type of each character.

[0158] To better illustrate the beneficial effects of the above technical solution, let's take product entity recognition for automobile companies as an example. We analyze the product entity recognition results of the following automobile product review:

[0159] Example 1: After looking at Song MAX, RX5 and H6, I finally bought Song MAX.

[0160] First, the process of constructing the candidate product entity reference graph is as follows: Figure 4 As shown:

[0161] Figure 4 a) First, the index set constructed for the relevant product entities is shown. Then, the candidate product entities for each character in Example 1 are generated, such as Figure 4 b). Next, Figure 4 The left side of c) lists the price and size information of each product type, and the candidate product entity reference graph can be constructed based on this information, such as Figure 4 c) is shown on the right. Finally, Figure 4 d) Shows the evidence score of each candidate product entity measured based on the weighted PageRank algorithm.

[0162] Then, by embedding the above results into the deep learning model, we can obtain three car type entities: "Song MAX", "RX5" and "H6".

[0163] An embodiment of the present invention provides a product entity recognition system that integrates product domain knowledge, including:

[0164] The index building module is used to build an index set of product entities in the vocabulary based on product domain knowledge;

[0165] An entity acquisition module, which acquires character-level candidate product entities corresponding to each character of the entity sentence to be identified according to the index set;

[0166] The relationship acquisition module is used to construct a candidate product entity reference graph based on the entity sentence to be identified, and obtain the global dependency relationship between any two character-level candidate product entities;

[0167] A score acquisition module, configured to obtain, based on each of the global dependencies, an evidence score for each of the character-level candidate product entities being a referent entity of the entity sentence to be identified;

[0168] A joint embedding module is used to obtain a joint embedding vector based on the entity statement to be identified, each of the character-level candidate product entities and their corresponding embedding weights, and use it as the input of a pre-built deep learning model to obtain the entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities.

[0169] An embodiment of the present invention provides a storage medium storing a computer program for product entity recognition integrating product domain knowledge, wherein the computer program enables a computer to execute the product entity recognition method described above.

[0170] An embodiment of the present invention provides an electronic device, including:

[0171] one or more processors;

[0172] Memory; and

[0173] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including a method for executing the product entity recognition method integrating product domain knowledge as described above.

[0174] It can be understood that the product entity recognition system, storage medium and electronic device that integrate product domain knowledge provided in the embodiments of the present invention correspond to the product entity recognition method that integrates product domain knowledge provided in the embodiments of the present invention. The explanations, examples and beneficial effects of the relevant contents can refer to the corresponding parts in the product entity recognition method and will not be repeated here.

[0175] In summary, compared with the existing technology, the present invention has the following beneficial effects:

[0176] 1. The product entity recognition method that integrates product domain knowledge proposed in the embodiment of the present invention can embed the global dependency relationship of entities into the deep learning model, expand the information source of the model, and thus enhance its ability to handle Chinese product entity recognition problems.

[0177] 2. When constructing the candidate product entity reference graph, the embodiment of the present invention designs a series of word-based and pronunciation-based rules to expand the index scope of product entities in the vocabulary, reducing the impact of irregular expressions in the text.

[0178] 3. The embodiments of the present invention can be used for entity recognition at the enterprise level, product level, and technology level for specific fields, that is, it can be applied to various sub-industries, including but not limited to the engine, automobile, and medical fields.

[0179] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0180] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A product entity recognition method integrating product domain knowledge, characterized in that: include: S1. Build an index set of product entities in the vocabulary based on product domain knowledge; S2. Obtain, based on the index set, character-level candidate product entities corresponding to each character of the entity sentence to be identified; S3. Based on the entity sentence to be identified, a candidate product entity reference graph is constructed to obtain the global dependency relationship between any two character-level candidate product entities; S4. Obtaining, based on each of the global dependencies, an evidence score for each of the character-level candidate product entities being a referent entity of the entity sentence to be identified; S5. Obtain a joint embedding vector based on the entity statement to be identified, each of the character-level candidate product entities, and their corresponding embedding weights, and use the joint embedding vector as input to a pre-built deep learning model to obtain an entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities; The process of constructing the candidate product entity reference graph in S3 includes: Define nodes as two entities: product brand and product type; Defining the relationships of a node includes: (1) The relationship between product brand and product type in, Indicates the affiliation between product type product1 and product brand brand1; (2) Relationships between product types in, Indicates the price similarity between any two product types product1 and product2, price1 and price2 are the prices of product1 and product2 respectively; in, Indicates the size similarity between any two product types, product1 and product2. l1, g1, and h1 represent the length, width, and height of product1, respectively. l2, g2, and h2 represent the length, width, and height of product2, respectively. (3) Relationship between product brands in, Represents the relationship between any two product brands brand1 and brand2, product 1,x Indicates the xth product type of brand1, product 2,y Indicates the y-th product type of brand2; X indicates that it belongs to brand i The number of product types, Y is the brand j Number of product types; S4 uses a weighted PageRank algorithm to obtain evidence scores, including: S41. Use entity popularity as the initial PR value of each node in the candidate product entity reference graph. in, Represents the selected product portal or user forum about any product entity w i Number of posts; S42. Calculate the PR value of each character-level candidate product entity. in, Respectively represent the character-level candidate product entity w i 、w j All relevant candidate product entities, represents w j to w i , α is the weight factor for balancing the contribution of the edge and product entity relationship; N represents the number of candidate product entities; S43, obtain evidence score, in, Represents the character-level candidate product entity w i The evidence score of the entity to be identified in the entity statement.

2. The product entity recognition method integrating product domain knowledge according to claim 1 is characterized in that: The index construction in S1 includes a word-based index and a pronunciation-based index.

3. The product entity recognition method integrating product domain knowledge according to claim 1 is characterized in that: In S2, a character-level candidate product entity generation algorithm is used to obtain the character-level candidate product entity. The algorithm includes: S10, obtain the entity statement to be identified S={a1,…,a i ,…a l }, vocabulary D, index set IS, the maximum length of the index in the index set maxLen; where a i Represents the i-th character of sentence S, with a total of l characters; S20. Initialize the character-level candidate product entity set S30, set i=1; S40, initializing the sliding window length ω=1; S50, if ω≤maxLen, go to step S60; otherwise, determine whether i is greater than 1. If so, end the algorithm; otherwise, set i=i+1 and go to step S40; S60, set the character sequence to chara_seq=a i a i+1 …a i+ω-1 ; If chara_seq is in the index set IS, then find the corresponding entity w in the vocabulary D and put it into the character-level candidate product entity set CE; S70. Set ω=ω+1 and go to step S50.

4. The product entity recognition method integrating product domain knowledge according to claim 1 is characterized in that: The process of obtaining the joint embedding vector in S5 includes: S100, BERT-based character embedding According to the entity statement to be identified S={a1,…,a i ,…a l }, the formula for generating character vectors using the pre-trained BERT model, in, Indicates a i A character vector of , d1 represents the embedding dimension of the character vector; S200, word embedding based on character-level candidate product entities, including: The word2vec model is used to vectorize each character-level candidate product entity. Among them, w i,j Represents the character a i The j-th character-level candidate product entity, Word2Vec w represents the word embedding lookup table, represents w i,j word vectors; By weighting the embedding weight of each character-level candidate product entity, the word vector corresponding to each character is obtained. in, Represents the character a i The j-th character-level candidate product entity is the evidence score of the referent entity of the entity sentence to be identified, as w i,j The embedding weight of k i is the character a i The number of candidate product entities; Represents the character a i The weighted word vectors of all character-level candidate product entities, S300, Combination of Character Embedding and Word Embedding Among them, x i Represents the character a i The joint embedding vector of γ∈(0,1) is a hyperparameter used to balance the importance of character vectors and word vectors.

5. The product entity recognition method integrating product domain knowledge according to any one of claims 1 to 4, characterized in that: The deep learning model adopts a Bi-LSTM-CRF network.

6. A product entity recognition system integrating product domain knowledge, characterized in that: The method for performing product entity recognition by integrating product domain knowledge as claimed in claim 1 comprises: The index building module is used to build an index set of product entities in the vocabulary based on product domain knowledge; An entity acquisition module, which acquires, based on the index set, character-level candidate product entities corresponding to each character of the entity sentence to be identified; The relationship acquisition module is used to construct a candidate product entity reference graph based on the entity sentence to be identified, and obtain the global dependency relationship between any two character-level candidate product entities; A score acquisition module, configured to obtain, based on each of the global dependencies, an evidence score for each of the character-level candidate product entities being a referent entity of the entity sentence to be identified; A joint embedding module is used to obtain a joint embedding vector based on the entity statement to be identified, each of the character-level candidate product entities and their corresponding embedding weights, and use it as the input of a pre-built deep learning model to obtain the entity recognition result; wherein the embedding weight refers to the evidence score of each of the character-level candidate product entities.

7. A storage medium, characterized in that: It stores a computer program for product entity recognition that integrates product domain knowledge, wherein the computer program enables a computer to execute the product entity recognition method according to any one of claims 1 to 5.

8. An electronic device, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the programs include a method for executing the product entity recognition method integrating product domain knowledge as described in any one of claims 1 to 5.

Images