2.1

Classification of Chinese place names

The classification of place names is the primary task of place name recognition. Place names should be easy to remember and help users to associate the stable associative relationship between place names and geographic entities. Those geographical names that are easy to be associated with specific geographical entities, can vividly reflect local characteristics and reflect regional cultural characteristics, are favored by the society because of their strong pointing effect ^[8]. The characters used in place names are an important part of the composition of place names. It is precisely because of the complexity and particularity of the characters used in place names that it is difficult to use a unified standard to classify Chinese place names. According to different purposes and principles, different place name classification methods can be used ^[9]. Some are named after surnames, while others are named according to dynasties and era names, and some names are not static and constantly changing over time, and there are many place names. The official name is very different from that of the local residents, which makes identification very difficult. Therefore, this paper intends to summarize the place names as a whole by reducing the classification of place names through several major categories, and each major category contains many sub-categories to refine the classification.

According to the analysis of the geographical names data, the data are mainly divided into the following four categories:

1. Organization name 2. Living service place name 3. Special place name, etc. 4. Village name

On this basis, each class is subdivided as shown in Table 1 below:

Table 1

Classification of place names

According to the above-mentioned custom classification method, the data preprocessing of place names can have a general direction. Through the open API of AutoNavi, about 20,000 pieces of address data in Shandong Province can be crawled for this article. The classification of place names in this article refers to AutoNavi map. The categories of poi in the classification collect data. Among them, the main categories are: automobile service, shopping service, catering service, medical care, scenic spot, commercial residence, road, company, government agency. data to form the original corpus of Chinese place name and address named entity recognition.

2.2

Participle of Chinese place names

The word segmentation of Chinese place names is the primary task of place name labeling, and the accuracy of word segmentation affects the accuracy of recognition. The existing word segmentation methods mainly include stuttering word segmentation, Hamp word segmentation, etc. Python’s stammering word segmentation, its word segmentation function is powerful and easy to install ^[10], Bao Shuguang introduced an unconventional data dictionary index table, which greatly improved the word segmentation algorithm. Speed ^[11]. In this paper, stuttering word segmentation is used in combination with a custom dictionary for raw data preprocessing.

As a good word segmentation tool for python, stuttering word segmentation mainly supports three modes of word segmentation:

This paper chooses to use the precise mode of stuttering word segmentation and combines custom dictionary for word segmentation. Before the word segmentation, the introduction of a custom dictionary includes the administrative divisions of Shandong Province, and the custom dictionary is imported before the word segmentation. For Chinese addresses, the idea of this article is to build a custom dictionary through Jieba to segment theinput address into counties and districts. At this level, the remaining parts after the segmentation are then processed for labeling and subsequent processing to improve the accuracy of named entity recognition and labeling. We make subsequent annotations on the part of the segmented reservation that can represent the address category.

2.3

Construction of Chinese place name annotation set

Named entity recognition is the task of sequence labeling in natural language processing, that is, through data preprocessing, text is input into the computer to identify entities and categories and label them. However, the current annotation systems are all aimed at natural language processing.According to the principle of place name classification in Table 1, and reading a large amount of data at the same time, this paper studies the relationship between address components, and constructs an annotation set from the perspective of address elements. According to the manual labeling method of the label set, the BMES labeling method is selected to label the segmented addresses. The labeled data is used to construct the experimental data of place name and address labeling according to the ratio of 8:1:1 of the training set, validation set and evaluation set.

The custom annotation set is shown in Table 2:

Table 2

Address annotation set

3.1

BILSTM model

BILSTM is an extension of the cyclic neural network model. The cyclic neural network model is based on previous experience or just to solve the current or follow-up problems. It can be regarded as a cyclic copy of multiple ordinary neural networks. Through this, the model obtains the above results. information. With the development of recurrent neural networks, the problem of untargeted long-distance dependencies has been shown. Therefore, the BILSTM model has effectively solved the long-distance dependency problem, and can also better capture bidirectional dependencies.

Sometimes the prediction may need to be determined by several previous inputs and several subsequent inputs, so bidirectional LSTM is introduced, which combines the information of the input sequence in the forward and backward directions, which will more accurately learn long-distance dependency information.

3.2

Conditional Random Field (CRF)

Named entity recognition plays an important role in various natural language processing applications, extracting key information from large amounts of unstructured text data. Named entity recognition is the task of identifying and classifying named entities into predefined categories for a given text.CRF uses the Viterbi decoding algorithm. The Viterbi decoding algorithm has a state transition matrix. In CRF, a state transition matrix will be initialized first. This matrix will be actively learned along with the training data. Get the highest score and choose the best solution. The loss function of the CRF layer includes the real path and the total score of all possible paths. The optimal solution is the ratio of the real path to all paths. The larger the ratio, the higher the accuracy. The optimal path score is decoded by the Viterbi algorithm. result. The state transition matrix is continuously updated by continuously training the training data. With this matrix, errors will be reduced. For example, when marking Shandong University, there will be B-org, East M-exc, large M-org, and E-org errors. East should be M-org. Adding CRF will reduce such errors and achieve constraints. These constraints solve the problem that the BILSTM model cannot make the labeling reach the optimal sequence, eliminating the need to manually construct CRF features, which is of great significance for improving the accuracy of the model.

3.3

Detailed explanation of the experimental process

First, the experimental data is crawled through the AutoNavi API, and the AutoNavi data is selectively selected according to the place name classification table 1. There are four categories of about 5,000 pieces of data for each type, a total of 20,000 pieces of data, and the experimental data is preprocessed and cleaned. Then perform stuttering word segmentation on the experimental data to retain the core part, manually label the experimental data according to the labeling set, divide it according to the 8:1:1 training set, testing machine, and evaluation set, input the BILSTM+CRF model for training and testing, and use LSTM at the same time The model and the CRF model were tested, and the experimental results were compared.

4.1

Experiment Evaluation Criteria and Results

The performance of the model is evaluated by the F1 value, the precision rate P, and the recall rate R.

Accuracy rate P = (number of correct predictions / total number of predictions)

Recall rate R = (number of correct predictions/total number of annotations)

F1= 2 * precision rate * recall rate / (precision rate + recall rate)

The experimental results are shown in Table 3 below:

Table 3

Evaluation of experimental results

4.2

Analysis of experimental results

Through the analysis of the experimental results, it can be seen from Table 3 that the BILSTM+CRF model used in this paper has greatly improved the P value, R value and F1 value of place name address recognition, and the accuracy of labeling is improved compared with the CRF model and the LSTM model. Obviously, the dimension of the word vector in the construction model is 64, 128, and 256 dimensions. It is better to choose 128 dimensions for the word vector through experiments. The optimal effect is achieved when the parameter value is 0.2. It is very useful for the task of place name recognition through a custom word segmentation dictionary and a custom label set. Compared with the label set based on part-of-speech Easier tasks that can be based on place names have better results.