Chinese geographic entity prediction method and system based on multilayer deep learning framework

Abstract

The invention discloses a Chinese geographic entity prediction method and system based on a multilayer deep learning framework, and belongs to the technical field of named entity recognition, and the method comprises the following steps: 1, obtaining a geographic entity data set in the field of Chinese geographic entity recognition, the geographic entity data set comprising a training sample; 2, training and predicting a Chinese geographic entity recognition model through the training sample; 3, adjusting and optimizing model parameters according to the prediction performance of the Chinese geographic entity recognition model; wherein ERNIE-Gram is used as a language preprocessing model in the training and prediction of the Chinese geographic entity recognition model, a tagged corpus is preprocessed, a corresponding dynamic word vector is obtained, a BiLSTM output result is decoded through IDCNN and BiLSTM training, emission score output and CRF, a tagged sequence with the maximum probability is predicted, and the Chinese geographic entity recognition model is obtained. And then classifying each entity in the tag sequence so as to complete the whole process of Chinese geographic entity identification. According to the invention, the accuracy of geographic entity identification is improved from a multi-dimensional perspective.

Description

Technical Field

[0001] The present invention belongs to the technical field of named entity recognition, and in particular, relates to a Chinese geographic entity prediction method and system based on a multi-layer deep learning framework. Background Art

[0002] With the emergence of Web2.0 and the widespread application of 3S (GPS, GIS, RS) technology, the amount of information such as global positioning system information acquisition and map information open access by mobile devices is increasing day by day. Massive amounts of geospatial data are generated every day through social media, map platforms, encyclopedias, etc. Most of them exist in the form of unstructured natural language text and need to be formalized for further analysis and mining. As the most basic work of geographic information formalization, geographic entity recognition, its accuracy will directly affect subsequent work such as spatial relationship extraction and knowledge graph construction.

[0003] Named entity recognition is an important task in the field of natural language processing. Its purpose is to locate the boundaries of named entities in text and classify them into a set of predefined types. Geographic entity recognition is a special case of named entity recognition. Its purpose is to identify geographical location-related entities from text, such as mountains, rivers, administrative divisions, roads, buildings and other geographical elements.

[0004] In recent years, the amount of data and importance of social media messages in various fields have increased dramatically. Identifying geographic entities from social media texts is of great significance. At the same time, identifying Chinese geographic entities from social media messages faces problems such as the irregularity of text, the complexity and diversity of geographic entity structures, and the ambiguity of geographic entities. In response to these problems, researchers have proposed a variety of geographic entity recognition methods, which can be divided into four categories: (1) Dictionary-based methods: Many scholars have proposed matching geographic entities with information in a predefined static geographic entity dictionary to identify geographic entities. This type of method is based on a fixed dictionary and ignores the influence of the context of geographic entities. It is difficult to handle geographic entity variants and polysemy in natural language texts. In addition, the geographic entity dictionary is difficult to update and cannot solve the problem of unregistered words. (2) Rule-based methods: Establish word formation rules and syntactic rules for geographic entities and use matching algorithms to achieve geographic entity recognition. In social media texts, the same geographic entity often has different meanings in different contexts. It is difficult to handle this situation by matching only with fixed rules. In addition, manually designed rules are extremely cumbersome and not universal. (3) Machine learning-based methods: using a large amount of annotated corpus to identify geographic entities based on their intrinsic features. This type of method still requires a lot of manual participation in feature extraction. For complex Chinese geographic entity recognition tasks, it is difficult to fully express the semantic and grammatical information in the context by manually designing features. (4) Deep learning-based methods: such as the BERT model, which has been widely used in the field of natural language processing in recent years. For geographic entity recognition tasks, BERT is used as a pre-trained model and then fine-tuned. It can make full use of contextual information, dynamically represent semantic and grammatical information, and handle the problem of polysemy, thereby achieving better recognition results. However, due to the masking mechanism during pre-training, BERT has certain limitations when recognizing Chinese texts with complex semantics. Summary of the invention

[0005] The purpose of the present invention is to provide a Chinese geographic entity prediction method and system based on a multi-layer deep learning framework to solve the problem that the existing geographic entity prediction method in the above-mentioned technology ignores the context or semantics of the context when identifying complex Chinese geographic entities in irregular Chinese syntax, and cannot dynamically represent the geographic entities in the context, resulting in inaccurate geographic entity recognition.

[0006] To achieve the above object, the present invention provides a Chinese geographic entity prediction method based on a multi-layer deep learning framework, comprising the following steps:

[0007] Step 1: Obtain a geographic entity dataset in the field of Chinese geographic entity recognition, where the geographic entity dataset includes training samples;

[0008] Step 2: Train and predict the Chinese geographic entity recognition model using training samples;

[0009] Step 3: According to the prediction performance of the Chinese geographic entity recognition model, tune the model parameters.

[0010] Preferably, the training samples include a training set, a test set and a validation set.

[0011] Preferably, the specific process of training and predicting the Chinese geographic entity recognition model in step 2 is as follows:

[0012] S21, vectorize the training samples through the ERNIE-Gram model to obtain vectorized representations of all geographic entities;

[0013] S22, extracting geographic entity features and context features from the vector representation of the text by feeding it into the IDCNN and BiLSTM models;

[0014] S23, using the CRF model to combine the previous and next adjacent labels and the context information of the word, predict the label of the word, and adjust the model parameters through the back propagation algorithm and optimization strategy to improve the ability to accurately predict geographic entities;

[0015] S24. Use the validation set in the training sample to test the model to evaluate the performance of the model in predicting geographic entities.

[0016] Preferably, the geographic entity dataset in the field of Chinese geographic entity recognition is a public Chinese named entity dataset, and the public Chinese named entity dataset includes MSRA and People's Daily datasets.

[0017] Preferably, the features extracted by the IDCNN model include local features of geographic entities in the sequence data, and the IDCNN model extracts contextual features by expanding the perception field of view; the BiLSTM model extracts contextual features by extracting long-term dependencies in the sequence data, and at the same time uses the gating mechanism of the BiLSTM model to delete the noise information introduced by the attention matrix of the ERNIE-Gram model.

[0018] Preferably, the local features of geographic entities extracted by the IDCNN model include location features, boundary features, and part-of-speech features.

[0019] Preferably, the specific process of the IDCNN model extracting contextual features by expanding the perception field of view is as follows: the dilated convolution operation of IDCNN when processing text enables the convolution kernel to cover a wide range of input data, thereby integrating contextual information.

[0020] The present invention also provides a system for a Chinese geographic entity prediction method based on a multi-layer deep learning framework, comprising a microprocessor and a memory connected to each other.

[0021] Therefore, the present invention adopts the above-mentioned Chinese geographic entity prediction method and system based on a multi-layer deep learning framework, which has the following beneficial effects:

[0022] (1) The model adopted by the present invention makes full use of the context in natural language and combines multiple deep learning models to improve the accuracy of geographic entity recognition from a multi-dimensional perspective and is used to identify geographic entities from various types of Chinese texts;

[0023] (2) The present invention proposes to use the ERNIE-Gram model to effectively encode and represent geographic entities, improve the representation and recognition capabilities of Chinese geographic entities, and solve the polysemy problem to a certain extent;

[0024] (3) The present invention proposes to add the IDCNN model to expand the perception field of the dynamic word vector and capture longer context information. And adopt a parallel computing method to minimize the training time of the IDCNN model;

[0025] (4) The model combines multi-layer convolution and recurrent neural networks, which can fully utilize the contextual information of the input text to extract richer feature representations and help identify complex geographic entities;

[0026] (5) The present invention improves the recognition performance of Chinese geographic entities in view of the irregularity of Chinese social media texts and the complexity of Chinese geographic entities, and provides a good foundation for downstream tasks such as Chinese geographic entity relationship extraction and the construction of Chinese geographic knowledge graphs.

[0027] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a model framework diagram of a Chinese geographic entity prediction method and system based on a multi-layer deep learning framework of the present invention. DETAILED DESCRIPTION

[0029] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] See also Figure 1, a Chinese geographic entity prediction method based on a multi-layer deep learning framework, comprising the following steps:

[0031] Step 1: Obtain a geographic entity dataset in the field of Chinese geographic entity recognition. The geographic entity dataset contains training samples, which include training sets, test sets, and validation sets. The geographic entity dataset in the field of Chinese geographic entity recognition is a public Chinese named entity dataset, which includes MSRA and People's Daily datasets.

[0032] Step 2: Train and predict the Chinese geographic entity recognition model using training samples. The specific process is as follows:

[0033] S21, vectorize the training samples through the ERNIE-Gram model to obtain vectorized representations of all geographic entities;

[0034] S22. By feeding the vector representation of the text into the IDCNN and BiLSTM models, geographic entity features and contextual features are extracted from the vector representation. The features extracted by the IDCNN model include local features of geographic entities in the sequence data. The local features of geographic entities include location features, boundary features, and part-of-speech features. The IDCNN model extracts contextual features by expanding the perception field of view. The specific process is as follows: the dilated convolution operation of IDCNN when processing text enables the convolution kernel to cover a wide range of input data, thereby integrating contextual information; the BiLSTM model extracts contextual features by extracting long-term dependencies in the sequence data, and at the same time uses the gating mechanism of the BiLSTM model to delete the noise information introduced by the attention matrix of the ERNIE-Gram model;

[0035] S23, using the CRF model to combine the previous and next adjacent labels and the context information of the word, predict the label of the word, and adjust the model parameters through the back propagation algorithm and optimization strategy to improve the ability to accurately predict geographic entities;

[0036] S24. Use the validation set in the training sample to test the model to evaluate the performance of the model in predicting geographic entities.

[0037] Step 3: According to the prediction performance of the Chinese geographic entity recognition model, tune the model parameters.

[0038] A system used in a Chinese geographic entity prediction method based on a multi-layer deep learning framework includes a microprocessor and a memory connected to each other.

[0039] Example

[0040] The specific process of the Chinese geographic entity prediction method is as follows:

[0041] S1) obtaining a Chinese geographic entity recognition text dataset as a training sample and inputting it into the model, and starting to loop and iterate each training epoch, wherein two public NER datasets: MSRA and People's Daily are the main datasets of the present invention, and the number of geographic entities contained in the dataset is shown in Table 1. The dataset adopts BIO annotation mode, and after the training sample is input into the model, the vector representation of the training sample is obtained through the language preprocessing model ERNIE-Gram;

[0042] Table 1 Number of geographic entities contained in the dataset

[0043] Dataset Training set Test Set Validation set People's Daily 206160 26033 25374 MSRA 36517 2877 2877

[0044] S2) The geographic entity word vector obtained by the ERNIE-Gram model is input into the IDCNN to increase the receptive field of view of the convolution kernel. Specifically, when processing text, the dilated convolution operation of IDCNN can increase the receptive field of view of the convolution kernel, exponentially expanding the receptive field of view without losing resolution and coverage, so that the convolution kernel can cover more input data and capture more geographic entity context information. The embedded representation of the text is first input into the IDCNN model to expand the receptive field of view and extract contextual features; the IDCNN model is composed of 4 dilated convolution blocks of the same size and structure, each of which has four dilated convolution layers with dilation widths of 1, 1, 2, and 4, respectively. Let X = [x1, x2,…, x r ] is the vector matrix input to the IDCNN module, the internal structure of the dilated convolution block is implemented as follows:

[0045]

[0046]

[0047]

[0048] in is a dilated convolutional neural network with a dilation distance of δ in the jth layer; is the feature obtained by convolution of the j-th layer network (j>1); r() represents the ReLU activation function.

[0049] S3) Let the output of IDCNN be X, and then input X into the BiLSTM model. BiLSTM further extracts the correlation features between Chinese geographic entities and deletes noise information. Among them, the word vector is passed through the forget gate f t =σ(W xf x t +W hf h t-1 +W cf C t-1 +bf ) to control the degree of forgetting of the input X and the output h of the previous hidden layer. Then, through the input gate i t =σ(W xi x t +W hi h t-1 +W ci C t-1 +b i ) to control the input X and the current calculation state to update to the memory unit c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c ) is the degree of the output gate. t =σ(W xo x t +W ho h t-1 +W co c t +b o ), h t =o t tanh(c t ) to control the final output. Where σ is the activation function; W is the weight matrix; b is the bias vector; c t is the content to be added; tanh is the hyperbolic tangent function; h t is the output of the entire LSTM unit at time t. Let X = (x1, x2, x3, ..., x t ), then for x t Learn the features from the previous and next contexts respectively, and get and Integrate the information of the previous and following texts to obtain H = [h1,h2,…,h t ], the output of the BiLSTM unit at time t can be expressed as:

[0050] S4) In the geographic entity recognition task, BiLSTM is good at processing long-distance text information, but cannot handle the dependencies between labels. Therefore, when there is no CRF layer, BiLSTM will independently predict the label corresponding to each word. CRF can obtain the global optimal prediction sequence through the relationship between neighboring labels to make up for the shortcomings of BiLSTM. Therefore, the CRF model is called to calculate the negative log-likelihood loss and update the model parameters through the correlation features output by BiLSTM, the label sequence and the mask; where, for any sequence X = (x1, x2, x3, ..., x n) and the corresponding prediction sequence Y = (y1, y2, y3, …, y n ), after passing through the IDCNN and BiLSTM layers, the score of each label can be obtained through linear mapping: P i =W s h (t) +b s , where h (t) is the input data x at time t of the previous layer (t) The output, W s and b are linear mapping parameters. CRF defines a label transfer score, assuming that P is the label score matrix of BiL STM, the size of P is n×k, where n is the number of words, k is the number of labels, and P i,j represents the score of the jth label of the i-th character. Then the score from the input sequence X to the predicted sequence Y can be expressed as: Where: W is the transfer matrix; W i,j is the fraction of label i transferred to label j, the size of W is k+2; the probability of predicting sequence Y is Taking the logarithm of both sides gives the likelihood function of the predicted sequence: in represents the true label sequence, Y x Represents all possible labeled sequences, and after Viterbi decoding, the output sequence with the maximum score is obtained:

[0051] S5) If the loop iteration is completed, the model is tested using the validation set to evaluate the performance of the model. If the training is not completed, jump to step S1.

[0052] The specific verification process is as follows:

[0053] The parameter settings of this experiment are shown in Table 2

[0054] Table 2 Parameter settings

[0055] parameter value Transformer Layers 768 Max_seq_len 130 Batch Values 100 LSTM-dim 200 Dilated convolutional block layers 4 Expansion distance 1,1,2,4 Number of dilated convolution blocks 4 Convolution kernel window size 3 Dropout 0.5 Learning Rate 0.001 Optimizer Adam

[0056] Evaluation indicators: This experiment uses three commonly used evaluation indicators in natural language processing to evaluate the model, namely precision (P), recall (R) and F1 value. Their respective definitions are as follows:

[0057]

[0058]

[0059] The F1 value is a commonly used evaluation method to measure the overall classification effect. Its formula is:

[0060]

[0061] Experimental results and conclusions:

[0062] In order to objectively evaluate the model proposed in this paper, this paper conducted multiple experiments on the People's Daily corpus and the MSRA corpus. The specific experimental results are shown in Table 3. (Note: Eg in the table refers to the pre-training model ERNIE-Gram; all evaluation indicators are the classification effects of geographic entities LOC)

[0063] Table 3 Experimental results of different models (unit: %)

[0064]

[0065] In both datasets, the model proposed in this paper achieves the best performance in terms of precision, recall and F1 value. EIBC_1 is the ERNIE-Gram-IDCNN-BiLSTM-CRF model, and EIBC_2 is the ERNIE-Gram-IDCNN-BiGRU-CRF model (EIBC refers to EIBC_1 if not specified below). The difference between EIBC_1 and EIBC_2 lies in the choice of recurrent neural network unit. EIBC_2 uses BiGRU (bidirectional gated recurrent unit) as the recurrent neural network unit. Compared with BiLSTM, BiGRU has fewer parameters and higher computational efficiency, but its feature extraction ability is slightly inferior to BiLSTM. Compared with the effect of the Bert-BiLSTM-CRF model on the MSRA dataset, the proposed model EIBC_1 is 1.005%, 0.94%, and 1% higher in precision, recall, and F1 value than the former, while EIBC_2 is 0.65%, 1.13%, and 0.89% higher in precision, recall, and F1 value than the Bert-BiLSTM-CRF model. Compared with other deep learning models, the proposed model achieves the best results in the three evaluation indicators, proving the effectiveness of the proposed model.

[0066] In order to demonstrate the role of each module in the EIBC model in this paper, four groups of ablation experiments were designed. Tables 4 to 7 present the results of the ablation experiments on the MSRA dataset.

[0067] Table 4 CRF ablation experiment (unit: %)

[0068] Model Precision Recall F1-score BiLSTM 82.24 83.21 82.72 BiLSTM-CRF 92.39 87.31 89.78

[0069] First, we compare the two sets of experiments, BiLSTM and BiLSTM-CRF. The results show that the F1 value of the latter is 7.06% higher than that of the former on the MSRA corpus, which proves the necessity of introducing the CRF model. The reason is that the CRF model can make full use of the dependency between adjacent labels, avoid invalid prediction sequences such as "B-PER O-LOC", and thus obtain the global optimal label sequence.

[0070] Table 5 ERNIE-Gram ablation experiment (unit: %)

[0071] Model Precision Recall F1-Score BiLSTM-CRF 92.39 87.31 89.78 BERT-BiLSTM-CRF 96.32 94.72 95.51 RoBERTa-BiLSTM-CRF 96.45 95.04 95.73 Eg-BiLSTM-CRF 96.22 95.55 95.89

[0072] Comparing the results of the three groups of experiments, BERT-BiLSTM-CRF, RoBERTa-BiLSTM-CRF, and Eg-BiLSTM-CRF, the F1 value of the latter is 0.38% and 0.2% higher than that of the former on the MSRA corpus, respectively. The reason is that BERT only focuses on the representation of fine-grained text units, and RoBERTa is based on BERT and has made several adjustments. The training time is longer and the amount of training data is larger, but it only focuses on the representation of fine-grained text units. The ERNIE-Gram model integrates the representation of both coarse-grained and fine-grained text units, and learns the representation of text more fully during pre-training. By comparing the results of the two groups of experiments, BiLSTM-CRF and Eg-BiLSTM-CRF, the F1 value of the latter is 6.11% higher on the MSRA corpus, proving the value of using the ERNIE-Gram pre-training model. The ERNIE-Gram model fully models the meaning of words, and the word vectors of geographic entities generated by it can better reflect the complex characteristics of geographic entities.

[0073] Table 6 IDCNN and BiLSTM ablation experiments (unit: %)

[0074] Model Precision Recall F1-score Eg-IDCNN-CRF 96.63 95.23 95.93 Eg-BiLSTM-CRF 96.22 95.55 95.89 Eg-IDCNN-BiLSTM-CRF 97.37 95.66 96.51

[0075] Comparing the results of the two sets of experiments, Eg-BiLSTM-CRF and Eg-IDCNN-BiLSTM-CRF, the F1 value of the latter is 0.62% higher than that of the former on the MSRA corpus, which proves the value of introducing the IDCNN model. The reason is that IDCNN can capture longer context information by expanding the perception field, thereby learning the local features of geographic entities and the global features of sentences.

[0076] Comparing the results of the two groups of experiments, Eg-IDCNN-CRF and Eg-IDCNN-BiLSTM-CRF, the F1 value of the latter is 0.58% higher than that of the former on the MSRA corpus. This proves the necessity of introducing the BiLSTM model. The reason is that ① the attention mechanism of ERNIE-Gram encodes each word, resulting in a very large attention matrix and the introduction of noise. ② The disadvantage of IDCNN is that although it increases the perception field of view, it also introduces a lot of irrelevant information. BiLSTM deletes the information it obtains through its gating mechanism and retains useful information, so that the model can learn the features of the text more effectively.

[0077] Table 7 BiGRU ablation experiment (unit: %)

[0078] Model Precision Recall F1-score Eg-IDCNN-BiLSTM-CRF 97.37 95.66 96.51 Eg-IDCNN-BiGRU-CRF 96.98 95.85 96.41

[0079] Comparing the results of the two groups of experiments, Eg-IDCNN-BiLSTM-CRF and Eg-IDCNN-BiGRU-CRF, the F1 values ​​of the two are not much different, which proves that the use of BiGRU and BiLSTM can improve the recognition effect of the model, but the difference between the two is small. Through the above experiments and analysis, we can know that:

[0080] (1) The ERNIE-Gram model uses explicit n-gram masked language modeling, so that the pre-trained word vectors can effectively encode and represent geographic entities, thereby improving the recognition performance of geographic entities.

[0081] (2) IDCNN makes full use of contextual information to extract rich features by expanding the receptive field of view, thereby improving the accuracy of geographic entity recognition.

[0082] (3) The BiLSTM model is good at processing long-distance text information and can learn the features of the entire sentence, thereby grasping the contextual information of the entity and removing the noise information introduced by the previous model. Integrating the features obtained by the IDCNN and BiLSTM models can make better use of contextual information compared to a single model.

[0083] (4) The CRF model improves the performance of geographic entity recognition by introducing constraint relationships between labels.

[0084] Therefore, the present invention adopts the above-mentioned Chinese geographic entity prediction method and system based on a multi-layer deep learning framework. First, the vector representation of geographic entities is obtained through ERNIE-Gram to improve the model's representation ability of geographic entities; secondly, the geographic entity context information is considered from different angles through IDCNN and BiLSTM to extract rich geographic entity context features; finally, CRF is used to construct the constraint relationship between geographic entity labels, generate the best label sequence, and reduce geographic entity ambiguity. Comparative experiments and ablation experiments were carried out with multiple groups of advanced Chinese geographic entity recognition models based on MSRA and People's Daily datasets, which proved the performance improvement of the model in this paper and the rationality of the combination. The present invention improves the recognition performance of Chinese geographic entities in view of the irregularity of Chinese social media texts and the complexity of Chinese geographic entities.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A Chinese geographic entity prediction method based on a multi-layer deep learning framework, It is characterized in that The following steps are involved: Step 1: Obtain a geographic entity dataset in the field of Chinese geographic entity recognition, where the geographic entity dataset includes training samples; Step 2: Train and predict the Chinese geographic entity recognition model using training samples; Step 3: According to the prediction performance of the Chinese geographic entity recognition model, tune the model parameters.

2. According to the Chinese geographic entity prediction method based on the multi-layer deep learning framework according to claim 1, Features: The training samples include a training set, a test set and a validation set.

3. According to the Chinese geographic entity prediction method based on the multi-layer deep learning framework according to claim 2, It is characterized in that The specific process of training and predicting the Chinese geographic entity recognition model in step 2 is as follows: S21, vectorize the training samples through the ERNIE-Gram model to obtain vectorized representations of all geographic entities; S22, extracting geographic entity features and context features from the vector representation of the text by feeding it into the IDCNN and BiLSTM models; S23, predicting the word label by combining the adjacent labels and the context information of the word through the CRF model, and adjusting the model parameters through the back propagation algorithm and optimization strategy; S24. Use the validation set in the training sample to test the model to evaluate the performance of the model in predicting geographic entities.

4. According to claim 3, the Chinese geographic entity prediction method based on a multi-layer deep learning framework, Features: The geographic entity dataset in the field of Chinese geographic entity recognition is a public Chinese named entity dataset, and the public Chinese named entity dataset includes MSRA and People's Daily datasets.

5. According to claim 4, the Chinese geographic entity prediction method based on the multi-layer deep learning framework, Features: The features extracted by the IDCNN model include local features of geographic entities in the sequence data, and the IDCNN model extracts contextual features by expanding the perception field of view; The BiLSTM model extracts context features by extracting long-term dependencies in sequence data, and uses the gating mechanism of the BiLSTM model to delete the noise information introduced by the attention matrix of the ERNIE-Gram model.

6. The Chinese geographic entity prediction method based on a multi-layer deep learning framework according to claim 5, Features: The local features of geographic entities extracted by the IDCNN model include location features, boundary features, and part-of-speech features.

7. The Chinese geographic entity prediction method based on a multi-layer deep learning framework according to claim 6, It is characterized in that The specific process of the IDCNN model to extract contextual features by expanding the receptive field of view is as follows: the dilated convolution operation of IDCNN when processing text enables the convolution kernel to cover a wide range of input data, thereby integrating contextual information.

8. A system for use in a Chinese geographic entity prediction method based on a multi-layer deep learning framework as described in any one of claims 1 to 7, Features: It consists of a microprocessor and memory connected to each other.

Images