A fine-grained entity classification method based on progressive transfer learning

Abstract

The present invention proposes a fine-grained entity classification method based on progressive transfer learning, which calculates the semantic similarity between the coarse-grained label of the target domain and all the coarse-grained labels of the source domain, and the semantic similarity between the fine-grained label of the target domain and all the fine-grained labels of the source domain; constructs a coarse-grained label mapping set and a fine-grained label mapping set, and calculates the semantic similarity weight; extracts the semantic features of sentences through a BiLSTM network, obtains a source domain coarse-grained label sentence semantic feature set, a fine-grained label sentence semantic feature set and a target domain semantic feature; calculates the overall coarse-grained semantic features and the overall fine-grained semantic features, and determines the overall semantic features; inputs the overall semantic features into a softmax classifier, and predicts the entity label corresponding to the fine-grained label selected by the target domain; trains and updates the parameters of the BiLSTM network and the softmax classifier, and obtains a fine-grained entity classification model based on progressive transfer learning. The present invention improves the accuracy of fine-grained entity classification in the target domain.

Description

Technical Field

[0001] The present invention relates to the field of information processing, and in particular to a fine-grained entity classification method based on progressive transfer learning. Background Art

[0002] In the field of natural language processing (NLP), fine-grained entity classification refers to mapping entity categories in a sentence into a hierarchical and refined entity category system. Since fine-grained entity corpus is difficult to annotate, in practical applications, it is often difficult to obtain large-scale, high-quality annotated corpus related to the domain. Therefore, how to perform fine-grained entity classification in the field of sparsely annotated corpus is a very challenging problem. Traditional transfer learning-based methods are mainly divided into two methods: model transfer and label transfer. Among them, the model transfer method transfers the model parameters of the source domain to the target domain, thereby improving the accuracy of fine-grained entity classification in the target domain; the label transfer method mainly transfers the semantic features of the source domain to the target domain based on the semantic similarity of the labels of the source domain and the target domain, thereby achieving fine-grained entity classification in the target domain. However, traditional methods have the following main problems: (1) Model transfer methods require that both the source domain and the target domain contain a large amount of annotated corpus, but in real applications annotated corpus is very scarce; (2) In the label mapping process, label transfer methods treat fine-grained categories as independent labels and ignore the hierarchical relationship between categories, resulting in inaccurate category semantics after transfer; (3) Due to the lack of annotated corpus in the target domain, label transfer methods cannot provide feedback to the model, resulting in poor domain adaptability of the model. When the similarity between the two domains is low, the fine-grained classification effect is also poor.

[0003] [1]Yang Z,Salakhutdinov R,Cohen W W.Transfer Learning for SequenceTagging with HierarchicalRecurrentNetworks[J].2017.

[0004] [2]Lee JY,Dernoncourt F,Szolovits P.Transfer Learning for Named-Entity Recognition withNeuralNetworks[J].2017.

[0005] [3]Abhishek A, Anand A, Awekar A. Fine-Grained Entity TypeClassification by Jointly Learning Representations and Label Embeddings[C] / / Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers. 2017.

[0006] [4]Hou F, Wang R, Zhou Y. Transfer learning for fine-grained entitytyping[J].Knowledge and Information Systems, 2021:1-22.

[0007] [5]Yoshua Bengio, Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, pages 41–48, 2009. Summary of the invention

[0008] The purpose of the present invention is to propose a fine-grained entity classification method based on progressive transfer learning.

[0009] The technical solution to achieve the purpose of the present invention is: a fine-grained entity classification method based on progressive transfer learning, the steps are as follows:

[0010] Step 1: Calculate entity label mapping weights

[0011] Randomly select the last level label in a target domain as the fine-grained label of the target domain, and obtain the first level label of the fine-grained label as the coarse-grained label of the target domain; calculate the semantic similarity between the coarse-grained label of the target domain and all the coarse-grained labels of the source domain, and the semantic similarity between the fine-grained label of the target domain and all the fine-grained labels of the source domain; sort by similarity, construct a coarse-grained label mapping set and a fine-grained label mapping set, and calculate the semantic similarity weight;

[0012] Step 2: Semantic feature extraction

[0013] For the coarse-grained label mapping set and the fine-grained label mapping set of the source domain, a sentence corresponding to each label is randomly selected, and the semantic features of the sentence are extracted through the BiLSTM network to obtain the semantic feature set of the coarse-grained label sentence and the semantic feature set of the fine-grained label sentence of the source domain; for the fine-grained label selected in the target domain, a sentence corresponding to the label is randomly selected, and the semantic features of the sentence are extracted through the BiLSTM network to obtain the semantic features of the target domain;

[0014] Step 3: Semantic feature integration

[0015] According to the semantic features of the coarse-grained label sentences and the fine-grained label sentences in the source domain, combined with the semantic similarity weight, the overall coarse-grained semantic features and the overall fine-grained semantic features are calculated to determine the overall semantic features;

[0016] Step 4: Fine-grained entity classification

[0017] The overall semantic features are input into the softmax classifier to predict the entity label corresponding to the fine-grained label selected by the target domain;

[0018] Step 5: Model training and prediction

[0019] The training corpus is iterated through steps 1 to 4 multiple times to train and update the parameters of the BiLSTM network and the softmax classifier. After training, a fine-grained entity classification model based on progressive transfer learning is obtained. The sentence to be predicted is input into the model to obtain the output result of the fine-grained entity classification of the sentence.

[0020] Furthermore, in step 1, the semantic similarity between the coarse-grained labels of the target domain and all the coarse-grained labels of the source domain, and the semantic similarity between the fine-grained labels of the target domain and all the fine-grained labels of the source domain are calculated. The specific method is:

[0021] Set the source domain dataset to be {l A |l A ∈L A ,t A |t A ∈T A}, the target domain dataset is {l B |l B ∈L B ,t B |t B ∈T B}, L A and T A is the sentence set and label set of the source domain, L B and T B For the set of statements and labels in the target domain, select any label in the target domain The coarse-grained label is The fine-grained label is Calculate the coarse-grained labels of the target domain separately All coarse-grained labels in the source domain Target domain fine-grained labeling Labels with fine-grained labels from the source domain The similarity is calculated as follows:

[0022]

[0023] Among them, X A With X B They are the word vectors of the source domain label and the target domain label respectively.

[0024] Furthermore, in step 1, sorting is performed by similarity, a coarse-grained label mapping set and a fine-grained label mapping set are constructed, and the semantic similarity weight is calculated. The specific method is:

[0025] Assume the target domain has coarse-grained labels The similarity set with the coarse-grained labels in the source domain is Target domain coarse-grained label The similarity set with the source domain fine-grained labels is Sort by similarity and select the first m coarse-grained labels from the source domain and the first n fine-grained labels Obtain a coarse-grained label mapping set and a fine-grained label mapping set;

[0026] Calculate the weight of each similar label in the coarse-grained label mapping set and the fine-grained label mapping set. The calculation formula is:

[0027]

[0028]

[0029]

[0030]

[0031] in, and are the coarse-grained label weights of the source domain and the fine-grained label weights of the source domain, and is the weight value of the target domain label, is the similarity value of the set target domain label vector.

[0032] Furthermore, in step 3, based on the semantic features of the source domain coarse-grained label sentences and the semantic features of the fine-grained label sentences, combined with the semantic similarity weights, the overall coarse-grained semantic features and the overall fine-grained semantic features are calculated to determine the overall semantic features. The specific method is as follows:

[0033] Calculate the overall coarse-grained semantic feature S c for:

[0034]

[0035] Calculate the overall fine-grained semantic feature S f for:

[0036]

[0037] Calculate the incremental learning parameter ε as:

[0038]

[0039] Calculate the total semantic feature S sum for:

[0040] S sum =(1-ε)S c +εS f (9)

[0041] in, and are the coarse-grained label weights of the source domain and the fine-grained label weights of the source domain, and is the weight value of the target domain label, is the similarity value of the set target domain label vector, and ρ is the ratio of the current number of iterations of progressive transfer learning to the total number of iterations.

[0042] Furthermore, in step 4, the total semantic features are input into the softmax classifier to predict the entity label corresponding to the fine-grained label selected by the target domain. The specific method is:

[0043] p(t|S sum )=softmax(WS sum +b) (10)

[0044] t′=arg max p(t|S sum ) (11)

[0045] Among them, W and b are parameters that need to be trained, and t′ is the predicted label.

[0046] A fine-grained entity classification system based on progressive transfer learning implements fine-grained entity classification based on progressive transfer learning based on the fine-grained entity classification method based on progressive transfer learning.

[0047] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, fine-grained entity classification based on progressive transfer learning is implemented based on the fine-grained entity classification method based on progressive transfer learning.

[0048] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements fine-grained entity classification based on progressive transfer learning based on the fine-grained entity classification method based on progressive transfer learning.

[0049] Compared with the prior art, the present invention has the following significant advantages: 1) It adopts a transfer learning method, so that the idea of ​​transfer learning is applied to fine-grained entity classification, and the semantic features of the source domain containing a large amount of annotated corpus are transferred to the sparsely annotated target domain, which solves the problem that the fine-grained entity classification model cannot be effectively constructed due to the sparse annotated corpus in the target domain, and improves the accuracy of fine-grained entity classification in the target domain; 2) It adopts the idea of ​​progressive learning, and in the process of training the fine-grained entity classification algorithm, training is carried out from coarse-grained to fine-grained and from simple to difficult, which effectively learns the characteristics of label hierarchy and optimizes the performance of the fine-grained entity classification algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flow chart of a fine-grained entity classification method based on progressive transfer learning of the present invention;

[0051] Figure 2 It is the label mapping between the source domain and the target domain. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0053] like Figure 1 As shown, the fine-grained entity classification method based on progressive transfer learning of the present invention comprises the following steps:

[0054] Step 1: Calculate entity label mapping weights.

[0055] The general domain is regarded as the source domain, and the proprietary domain is regarded as the target domain. The source domain contains the fine-grained label system and high-quality annotated data set of the general domain, and the target domain contains the fine-grained label system and a small amount of annotated data of the proprietary domain. Set the source domain dataset as {l A |l A ∈L A ,t A |t A ∈T A}, the target domain dataset is {l B |l B ∈L B ,t B |t B ∈T B}. A and T A is the sentence set and label set of the source domain, L B and T B is the set of statements and labels in the target domain. Select any label in the target domain Can get The coarse-grained label is The fine-grained label is Calculate the coarse-grained labels of the target domain separately All labels of the first level of the source domain Target domain fine-grained labeling All labels of the last level of the source domain The calculation formula is as follows:

[0056]

[0057] Among them, X A With X B are the word vectors of the source domain label and the target domain label respectively. The coarse-grained label of the target domain can be obtained from formula (1): The similarity set with the first-level labels in the source domain is Target domain coarse-grained label The similarity set with the first-level labels in the source domain is

[0058] According to the similarity sorting, the first m source domain coarse-grained labels are selected to obtain the coarse-grained label mapping set. The first n target domain coarse-grained labels, get the fine-grained label mapping set Label mapping diagram Figure 2 As shown, the weight of each similar tag is obtained from the similarity set, and the calculation formula is:

[0059]

[0060]

[0061]

[0062]

[0063] in, and are the coarse-grained label weights of the source domain and the fine-grained label weights of the source domain, and is the weight value of the target domain label, is the similarity value of the set target domain label vector.

[0064] Step 2: semantic feature extraction.

[0065] Select a sentence from each label in the coarse-grained label mapping set and the fine-grained label mapping set, use the BiLSTM network to extract sentence features, and obtain the source domain coarse-grained label sentence semantic feature set And the semantic feature set of source domain fine-grained label sentences

[0066] Similarly, the semantic features of the sentence under the target domain label are s B .

[0067] Step 3: semantic feature integration.

[0068] The weighted sum of the semantic features of the source domain coarse-grained label sentence semantic features and the coarse-grained label weight, and the source domain fine-grained label sentence semantic features and the fine-grained label weight are calculated respectively, and then summed with the target domain semantic features to obtain the overall coarse-grained semantic features and the overall fine-grained semantic features.

[0069] Overall coarse-grained semantic feature S c for:

[0070]

[0071] Overall fine-grained semantic feature S f for:

[0072]

[0073] The weights of the coarse-grained semantic features and the fine-grained semantic features in the incremental learning are obtained according to the training rounds, and the weighted sum of the overall coarse-grained semantic features and the overall fine-grained semantic features is performed to obtain the final semantic features. From the incremental learning, it can be obtained that the incremental learning parameter ε is:

[0074]

[0075] Among them, ρ is the ratio of the current number of iterations to the total number of iterations during training.

[0076] The overall semantic feature S sum for:

[0077] S sum =(1-ε)S c +εS f (9) When the number of iterations is small, the ε value is close to 0, and S sum ≈S c ; When the number of iterations is close to the total number of iterations, the ε value is close to 1, and S sum ≈S f .

[0078] Step 4: Fine-grained entity classification. The obtained total semantic features are input into the softmax classifier to predict its entity label:

[0079] p(t|S sum )=softmax(WS sum +b) (10)

[0080] t′=arg max p(t|S sum ) (11)

[0081] Step 5: Model training and prediction. The training corpus is iterated through steps 1 to 4 multiple times to train and update the parameters of the BiLSTM network and softmax classifier. After training, a fine-grained entity classification model based on progressive transfer learning is obtained. The sentence to be predicted is input into the model to obtain the output result of the fine-grained entity classification of the sentence.

[0082] The present invention also proposes a fine-grained entity classification system based on progressive transfer learning, and implements fine-grained entity classification based on progressive transfer learning based on the fine-grained entity classification method based on progressive transfer learning.

[0083] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, fine-grained entity classification based on progressive transfer learning is implemented based on the fine-grained entity classification method based on progressive transfer learning.

[0084] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements fine-grained entity classification based on progressive transfer learning based on the fine-grained entity classification method based on progressive transfer learning.

[0085] Example

[0086] In order to verify the effectiveness of the solution of the present invention, the following simulation experiment is carried out.

[0087] Step 1: Calculation of entity label mapping weights. For example, a label LOC.Land.Continent is randomly selected in the target domain, whose coarse-grained label is LOC and the fine-grained label is continent. In the source domain, the first-layer label set is [PER,LOC,GPE,FAC,ORG,VAL…], and the last-layer label set is [Governor,Minister,Apartment Building,House,School,Province State,UrbanArea, Region,Country,Team,Corporation…]. According to the word vectors, the similarity between the coarse-grained label of the target domain and the first-layer label of the source domain, and the similarity between the fine-grained label of the target domain and the last-layer label of the source domain are calculated respectively, and sorted according to the similarity to obtain the coarse-grained label mapping set [LOC,GPE,FAC,ORG] and the fine-grained label mapping set [Continent,Country,Organization Of Countries,Province State, UrbanArea,Region,City,Village,Field]. The weights of the target domain coarse-grained label LOC and the coarse-grained mapping set [LOC, GPE, FAC, ORG] are calculated respectively, and the weight set is [0.31, 0.18, 0.13, 0.07]; the weights of the target domain fine-grained label Continent and the fine-grained label mapping set [Continent, Country, Organization OfCountries, Province State, UrbanArea] are calculated respectively, and the weight set is [0.25, 0.17, 0.14, 0.11, 0.08]. Set the similarity parameter of the target domain label vector is 1.0, and the coarse label weight of the target domain can be obtained With fine-grained label weights They are 0.31 and 0.25 respectively.

[0088] Step 2: Semantic feature extraction. Select a sentence from each label in the coarse-grained label mapping set and the fine-grained label mapping set, and obtain the source domain coarse-grained label sentence semantic feature set through the BiLSTM network. And the semantic feature set of source domain fine-grained label sentences Similarly, the semantic features of the sentence under the target domain label are s B For example BThe semantic feature vector is [6.890419 -2.795039 4.2934331.842001 3.672960 -3.585120 -18.438543 3.988691 9.442407 3.132476 13.31136614.396519 10.657969 12.447675 12.598063 -12.850086 -8.458016 -7.10521811.960398 1.979599 -14.365957 -1.148624 -14.423791 -6.047969 10.217813 3.131283 10.387719 -3.936613 -8.450990 -9.578313 6.228574 -5.514681 -2.731924 2.395364 -2.644078 -12.621920 11.8689687.695186 18.216955 11.367293 2.459505 8.968964 17.385582 6.895690 -6.758958 -0.662597 -6.119532].

[0089] Step 3. Semantic feature integration. When the number of iterations just starts, the ε value is 0. At this time, according to formulas (6), (7), (8), and (9), the coarse-grained semantic feature is the total semantic feature, that is, the final semantic feature vector is [-1.235999 -16.372496 18.046347 0.737840 -3.286072 4.972629 -22.899796 2.875767 4.60295923.246979 -5.899751 -9.975878 -8.812774 6.368253 8.421858 -1.366317 6.777222-5.077109 5.373605 -0.951322 -20.269548 -14.224439 -9.978577 -4.476111 -2.221652 8.268779 -1.289561 -3.921106 -0.204428 -0.414611 -6.856577 -1.711689-1.388898 12.953831 -40.208401 0.625537 3.331903 1.312719 0.968604 9.556758 -1.717613 12.953566 2.830797 2.171817 -0.679360 16.886272 8.778729 1.57026719.004416 20.107653].

[0090] Step 4. Fine-grained entity classification: The obtained total semantic features are input into the softmax classifier, and its entity label is predicted to be LOC.Land.Continent.

[0091] In summary, the present invention realizes semantic feature transfer by learning the semantic feature mapping relationship between the source domain and the target domain for a target domain with sparse annotated corpus and a source domain with sufficient annotated corpus; adopts the idea of ​​progressive learning, learns from coarse-grained to fine-grained, and trains the entity classification algorithm from simple to difficult, effectively captures the hierarchical features between labels, and makes up for the shortcomings of traditional methods such as reliance on a large amount of annotated corpus, inaccurate semantic feature transfer, and poor model domain adaptability, thereby improving the accuracy of fine-grained entity classification in the target domain.

[0092] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A fine-grained entity classification method based on progressive transfer learning, characterized in that: Here are the steps: Step 1: Calculate entity label mapping weights Randomly select the last level label in a target domain as the fine-grained label of the target domain, and obtain the first level label of the fine-grained label as the coarse-grained label of the target domain; calculate the semantic similarity between the coarse-grained label of the target domain and all the coarse-grained labels of the source domain, and the semantic similarity between the fine-grained label of the target domain and all the fine-grained labels of the source domain; sort by similarity, construct a coarse-grained label mapping set and a fine-grained label mapping set, and calculate the semantic similarity weight; Step 2: Semantic feature extraction For the coarse-grained label mapping set and the fine-grained label mapping set of the source domain, a sentence corresponding to each label is randomly selected, and the semantic features of the sentence are extracted through the BiLSTM network to obtain the semantic feature set of the coarse-grained label sentence and the semantic feature set of the fine-grained label sentence of the source domain; for the fine-grained label selected in the target domain, a sentence corresponding to the label is randomly selected, and the semantic features of the sentence are extracted through the BiLSTM network to obtain the semantic features of the target domain; Step 3: Semantic feature integration According to the semantic features of the coarse-grained label sentences and the fine-grained label sentences in the source domain, combined with the semantic similarity weight, the overall coarse-grained semantic features and the overall fine-grained semantic features are calculated to determine the overall semantic features; Step 4: Fine-grained entity classification The overall semantic features are input into the softmax classifier to predict the entity label corresponding to the fine-grained label selected by the target domain; Step 5: Model training and prediction The training corpus is iterated through steps 1 to 4 multiple times to train and update the parameters of the BiLSTM network and the softmax classifier. After training, a fine-grained entity classification model based on progressive transfer learning is obtained. The sentence to be predicted is input into the model to obtain the output result of the fine-grained entity classification of the sentence. Step 1: Calculate the semantic similarity between the coarse-grained labels of the target domain and all the coarse-grained labels of the source domain, as well as the semantic similarity between the fine-grained labels of the target domain and all the fine-grained labels of the source domain. The specific method is: Set the source domain dataset to be {l A |l A ∈L A ,t A |t A ∈T A }, the target domain dataset is {l B |l B ∈L B ,t B |t B ∈T B }, L A and T A is the sentence set and label set of the source domain, L B and T B For the set of statements and labels in the target domain, select any label in the target domain The coarse-grained label is The fine-grained label is Calculate the coarse-grained labels of the target domain separately All coarse-grained labels in the source domain Target domain fine-grained labeling Labels with fine-grained labels from the source domain The similarity is calculated as follows: Among them, X A With X B are the word vectors of the source domain label and the target domain label respectively; Sort by similarity, build a coarse-grained label mapping set and a fine-grained label mapping set, and calculate the semantic similarity weight. The specific method is: Assume the target domain has coarse-grained labels The similarity set with the coarse-grained labels in the source domain is Target domain coarse-grained label The similarity set with the source domain fine-grained labels is Sort by similarity and select the first m coarse-grained labels from the source domain and the first n fine-grained labels Obtain a coarse-grained label mapping set and a fine-grained label mapping set; Calculate the weight of each similar label in the coarse-grained label mapping set and the fine-grained label mapping set. The calculation formula is: in, and are the coarse-grained label weights of the source domain and the fine-grained label weights of the source domain, and is the weight value of the target domain label, is the similarity value of the set target domain label vector; Step 2: For the coarse-grained label mapping set and fine-grained label mapping set of the source domain, randomly select a sentence corresponding to each label, extract the semantic features of the sentence through the BiLSTM network, and obtain the semantic feature set of the coarse-grained label sentence in the source domain. and a collection of fine-grained label sentence semantic features For the fine-grained label selected for the target domain, a sentence is randomly selected from the labeled sentence corresponding to the label, and the semantic features of the sentence are extracted through the BiLSTM network to obtain the target domain semantic feature s B ; Step 3: According to the semantic features of the coarse-grained label sentences and the fine-grained label sentences in the source domain, combined with the semantic similarity weight, the overall coarse-grained semantic features and the overall fine-grained semantic features are calculated to determine the overall semantic features. The specific method is as follows: Calculate the overall coarse-grained semantic feature S c for: Calculate the overall fine-grained semantic feature S f for: Calculate the incremental learning parameter ε as: Calculate the total semantic feature S sum for: S sum =(1-ε)S c +εS f (9) in, and are the coarse-grained label weights of the source domain and the fine-grained label weights of the source domain, and is the weight value of the target domain label, is the similarity value of the set target domain label vector, and ρ is the ratio of the current number of iterations of progressive transfer learning to the total number of iterations.

2. The fine-grained entity classification method based on progressive transfer learning according to claim 1, characterized in that: Step 4: Input the total semantic features into the softmax classifier to predict the entity label corresponding to the fine-grained label selected by the target domain. The specific method is: p(t|S sum )=softmax(WS sum +b) (10) t′=arg maxp(t|S sum ) (11) Among them, W and b are the parameters that need to be trained, and t′ is the predicted label.

3. A fine-grained entity classification system based on progressive transfer learning, characterized in that: Based on the fine-grained entity classification method based on progressive transfer learning according to any one of claims 1 to 2, fine-grained entity classification based on progressive transfer learning is implemented, including: The entity label mapping weight calculation unit is used to randomly select the last level label in a target domain as the fine-grained label of the target domain, obtain the first level label of the fine-grained label as the coarse-grained label of the target domain; calculate the semantic similarity between the coarse-grained label of the target domain and all the coarse-grained labels of the source domain, and the semantic similarity between the fine-grained label of the target domain and all the fine-grained labels of the source domain; sort by similarity, construct a coarse-grained label mapping set and a fine-grained label mapping set, and calculate the semantic similarity weight; The semantic feature extraction unit randomly selects a sentence corresponding to each label in the coarse-grained label mapping set and the fine-grained label mapping set of the source domain, extracts the semantic features of the sentence through the BiLSTM network, and obtains the semantic feature set of the coarse-grained label sentence and the semantic feature set of the fine-grained label sentence in the source domain; randomly selects a sentence corresponding to the fine-grained label selected in the target domain, extracts the semantic features of the sentence through the BiLSTM network, and obtains the semantic features of the target domain; A semantic feature integration unit is used to calculate the overall coarse-grained semantic features and the overall fine-grained semantic features according to the semantic features of the coarse-grained label sentences and the semantic features of the fine-grained label sentences in the source domain and combine the semantic similarity weights to further determine the overall semantic features; The fine-grained entity classification unit is used to input the overall semantic features into the softmax classifier to predict the entity label corresponding to the fine-grained label selected by the target domain; The model training and prediction unit is used to iterate the training corpus multiple times to train and update the parameters of the BiLSTM network and the softmax classifier. After training, a fine-grained entity classification model based on progressive transfer learning is obtained. The sentence to be predicted is input into the model to obtain the output result of the fine-grained entity classification of the sentence.

4. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the fine-grained entity classification method based on progressive transfer learning according to any one of claims 1 to 2 is used to implement fine-grained entity classification based on progressive transfer learning.

5. A computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the fine-grained entity classification method based on progressive transfer learning is implemented based on the fine-grained entity classification method based on progressive transfer learning as described in any one of claims 1-2.

Images