CN116579343A - A Named Entity Recognition Method for Chinese Culture and Tourism - Google Patents

Abstract

The invention discloses a named entity recognition method for Chinese culture and tourism, comprising the following steps: S1, acquiring text data of Chinese culture and tourism, and inputting it into a character embedding layer to obtain a character vector representation; S2, representing the character vector Input to the bidirectional long-term and short-term memory network layer to obtain the context representation; S3. Input the context representation to the CNN layer to obtain a multi-scale local context feature fusion representation; S4. Input the multi-scale local context feature fusion representation to the CRF layer, through The CRF layer performs sequence annotation and completes the named entity recognition of Chinese cultural tourism. The present invention takes into account the problem of less attention paid to the research on named entity recognition of Chinese tourism, and builds a network for Chinese cultural and tourism text data, and uses the second CNN module at the CNN layer to learn multi-scale local context feature fusion Representation, strengthen the correlation between semantics, and improve the feature representation that is beneficial to Chinese recognition.

Description

A Named Entity Recognition Method for Chinese Culture and Tourism

technical field

The invention belongs to the technical field of information extraction, and in particular relates to a named entity recognition method for Chinese culture and tourism.

Background technique

Named entity recognition (NER) is a basic information extraction task that can be applied to many downstream tasks in natural language processing (NLP), such as information extraction, social media analysis, search engine, machine translation, knowledge graph. The goal of NER is to extract some predefined specific entities from a sentence and recognize their correct type like person, place, organization. Early named entity recognition falls into two categories: rule-based methods and statistical-based methods. With the growing strength of deep learning, NER research has made great progress. The fields involved are various: such as the medical field, the financial field, the news field, etc. However, research on named entity recognition for cultural tourism is very scarce, and research on named entity recognition for cultural tourism has not received much attention.

According to the differences between languages, there are also many studies on NER methods for specific languages, such as English, Arabic, Indian and other languages, and many researchers mainly focus on the research of English NER. However, Chinese is an important international language. Compared with English, Chinese has its own characteristics, but the research on Chinese NER is much less than that of English NER, and many studies on Chinese NER have no Make targeted research according to the characteristics of Chinese.

Contents of the invention

In view of the above-mentioned deficiencies in the prior art, the invention provides a named entity recognition method for Chinese culture and tourism, which solves the problem that the current research on named entity recognition pays less attention to Chinese culture and tourism.

In order to achieve the above-mentioned purpose of the invention, the technical solution adopted in the present invention is: a named entity recognition method for Chinese culture and tourism, comprising the following steps:

S1. Obtain the text data of Chinese culture and tourism, and input it into the character embedding layer to obtain the character vector representation;

S2. Input the character vector representation to the bidirectional long short-term memory network layer to obtain the context representation;

S3. Input the context representation to the CNN layer to obtain multi-scale local context feature fusion representation;

S4. Input the multi-scale local contextual feature fusion representation to the CRF layer, and perform sequence annotation through the CRF layer to complete the named entity recognition of Chinese culture and tourism.

Further: in the S1, the character embedding layer includes a parallel ChineseBert module and the first CNN module;

The S1 includes the following sub-steps:

S11. Acquiring Chinese cultural and tourism text data;

S12. Input the Chinese cultural travel text data into the ChineseBert module, and obtain the word embedding vector representation of each word in the Chinese cultural travel text data;

S13. Input the Chinese cultural tourism text data into the first CNN module to obtain a radical-level embedding representation;

S14. Concatenate the word embedding vector representation with the radical-level embedding representation to obtain a character vector representation.

Further: said S12 is specifically:

Input the Chinese cultural tourism text data into the ChineseBert module, encode and express the input Chinese cultural tourism text data through the ChineseBert module, obtain the feature vector, and generate the word embedding vector of each word in the Chinese cultural tourism text data according to the feature vector express;

Wherein, the feature vector includes label embedding, position embedding and segment embedding.

Further: in said S13, the expression for obtaining the radical-level embedding representation _M2 is specifically:

M ₂ =A ₁ (b ₁ +C ₁ (x))

In the formula, x is the radical-level feature of Chinese characters, C ₁ (·) is the first CNN module, A ₁ is the first activation function, and b ₁ is the weight of the first CNN module.

Further: in the said S14, the expression obtained character vector representation Z _concat is specifically:

Z _concat =M ₁ +M ₂

In the formula, M ₁ is the word embedding vector representation.

The beneficial effect of the above further solution is that the character vector representation obtained by concatenating the word embedding vector representation and the radical-level embedding representation can obtain more semantic features, so that the model can better recognize the Chinese meaning in the text.

Further: in said S2, the bidirectional long-short-term memory network layer includes the first to twelfth LSTM units, the first to sixth LSTM units are forward processing the input character vector representation, and the seventh to twelfth The LSTM unit inversely processes the character vector representation of the input;

The method of obtaining the context representation is as follows:

Splicing is performed according to the output results of the first to twelfth LSTM units to obtain the context representation. further:

Further: in said S2, the expression for obtaining the context representation H is specifically:

H＝{h ₁ ,...,h _ti ,...,h _D }

In the formula, h _ti is the splicing of the output results of the first to twelfth LSTM units, ti is the sequence number of splicing, and ti=1,...,D, D is the dimension represented by the character vector;

The first to twelfth LSTM units all include an input gate _it , an output gate o _t and a forgetting gate f _t , the expressions of which are specifically as follows:

i _t = σ(W _xi x _t +W _hi h _t-1 +W _ci c _t-1 +b _i )

f _t ＝σ(W _xf x _t +W _hf h _t-1 +W _cf c _t-1 +b _f )

c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(W _xc x _t +W _hc h _t-1 +b _c )

o _t ＝σ(W _xo x _t +W _ho h _t-1 +W _co c _t +b _o )

h _t ＝o _t ⊙tanh(c _t )

In the formula, σ(·) is the per-element sigmoid function, tanh(·) is the hyperbolic tangent function, ⊙ is the per-element multiplication function, W _xi , W _hi , W _ci , W _xf , W _hf , W _cf , W _xc , W _hc , W _xo , _Who and W _co are weight parameters, b _i , b _f , b _c and b _o are weight parameters, c _t is memory cells, and h _t is output results.

Further: in the S3, the CNN layer is provided with a second CNN module, and the multi-scale local context feature fusion representation M is obtained. The expression of _M3 is specifically:

M ₃ =A ₂ (b ₂ +C ₂ (H))

In the formula, H is the expression below, C ₂ (·) is the second CNN module, A ₂ is the second activation function, and b ₂ is the weight of the second CNN module.

The beneficial effect of the above further solution is: inputting the context representation to the second CNN module can strengthen the correlation between semantics and generate a multi-scale fusion representation of local context features.

The beneficial effects of the present invention are as follows: a method for recognizing named entities of Chinese cultural and tourism categories provided by the present invention solves the problem that less attention is paid to the research on named entity recognition of Chinese tourism categories, and is aimed at Chinese cultural and tourism text data Carry out network construction, use the first CNN module at the character embedding layer to learn Chinese-based radical-level embedded representations, and obtain character vector representations that are conducive to Chinese recognition, and use the second CNN module at the CNN layer to learn multi-scale local context feature fusion representations , strengthen the correlation between semantics, and further improve the feature representation that is beneficial to Chinese recognition.

Description of drawings

FIG. 1 is a flow chart of a named entity recognition method for Chinese culture and tourism in the present invention.

FIG. 2 is a schematic diagram of the overall network structure of the present invention.

Fig. 3 is a schematic structural diagram of the ChineseBert module of the present invention.

FIG. 4 is a schematic structural diagram of the first CNN module of the present invention.

FIG. 5 is a schematic structural diagram of the second CNN module of the present invention.

Detailed ways

The specific embodiments of the present invention are described below so that those skilled in the art can understand the present invention, but it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes Within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

As shown in Figure 1, in one embodiment of the present invention, a named entity recognition method of a Chinese cultural tourism class, comprises the following steps:

S1. Obtain the text data of Chinese culture and tourism, and input it into the character embedding layer to obtain the character vector representation;

S2. Input the character vector representation to the bidirectional long short-term memory network layer to obtain the context representation;

S3. Input the context representation to the CNN layer to obtain multi-scale local context feature fusion representation;

S4. Input the multi-scale local contextual feature fusion representation to the CRF layer, and perform sequence annotation through the CRF layer to complete the named entity recognition of Chinese culture and tourism.

In this embodiment, the present invention provides a named entity recognition method for Chinese cultural tourism based on the fusion representation of radical-level features and multi-scale local context features for the characteristics of Chinese characters and the applicable field is cultural tourism data. The specific structure is shown in Figure 2.

In said S1, the character embedding layer includes a parallel ChineseBert module and the first CNN module;

The S1 includes the following sub-steps:

S11. Acquiring Chinese cultural and tourism text data;

S12. Input the Chinese cultural travel text data into the ChineseBert module, and obtain the word embedding vector representation of each word in the Chinese cultural travel text data;

S13. Input the Chinese cultural tourism text data into the first CNN module to obtain a radical-level embedding representation;

S14. Concatenate the word embedding vector representation with the radical-level embedding representation to obtain a character vector representation.

In this embodiment, the structure of the ChineseBert module is shown in FIG. 3 . The ChineseBert module is a pre-trained model obtained through pre-training Chinese corpus, and is specially processed for Chinese text data.

The S12 is specifically:

Input the Chinese cultural tourism text data into the ChineseBert module, encode and express the input Chinese cultural tourism text data through the ChineseBert module, obtain the feature vector, and generate the word embedding vector of each word in the Chinese cultural tourism text data according to the feature vector express;

Wherein, the feature vector includes label embedding, position embedding and segment embedding.

In said S13, the expression for obtaining the radical-level embedding representation _M2 is specifically:

M ₂ =A ₁ (b ₁ +C ₁ (x))

In the formula, x is the radical-level feature of Chinese characters, C ₁ (·) is the first CNN module, A ₁ is the first activation function, and b ₁ is the weight of the first CNN module.

In this embodiment, CNN is used to perform Radical-level Representation on the input Chinese cultural and tourism text data, and the Radical-level Representation is obtained, where the first CNN module performs Radical Representaion on the input data. The structural diagram is as follows Figure 4 shows.

In said S14, the expression obtained by character vector representing Z _concat is specifically:

Z _concat =M ₁ +M ₂

In the formula, M ₁ is the word embedding vector representation.

The character vector representation obtained by concatenating word embedding vector representation and radical-level embedding representation can obtain more semantic features, so that the model can better recognize the Chinese meaning in the text.

In said S2, the bidirectional long-short-term memory network layer includes the first to twelfth LSTM units, the first to sixth LSTM units forwardly process the input character vector representation, and the seventh to twelfth LSTM units reverse A character vector representation of the input to the process;

The method of obtaining the context representation is as follows:

Splicing is performed according to the output results of the first to twelfth LSTM units to obtain the context representation.

In this embodiment, the context representation blocks obtained by the bidirectional long-short-term memory network layer can improve the semantic representation from both positive and negative directions, and can better recognize the semantics in the paragraph.

In said S2, the expression for obtaining the context representation H is specifically:

H＝{h ₁ ,...,h _ti ,...,h _D }

In the formula, h _ti is the splicing of the output results of the first to twelfth LSTM units, ti is the sequence number of splicing, and ti=1,...,D, D is the dimension represented by the character vector;

The first to twelfth LSTM units all include an input gate _it , an output gate o _t and a forgetting gate f _t , the expressions of which are specifically as follows:

i _t = σ(W _xi x _t +W _hi h _t-1 +W _ci c _t-1 +b _i )

f _t ＝σ(W _xf x _t +W _hf h _t-1 +W _cf c _t-1 +b _f )

c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(W _xc x _t +W _hc h _t-1 +b _c )

o _t ＝σ(W _xo x _t +W _ho h _t-1 +W _co c _t +b _o )

h _t ＝o _t ⊙tanh(c _t )

In the formula, σ(·) is the per-element sigmoid function, tanh(·) is the hyperbolic tangent function, ⊙ is the per-element multiplication function, W _xi , W _hi , W _ci , W _xf , W _hf , W _cf , W _xc , W _hc , W _xo , _Who and W _co are weight parameters, b _i , b _f , b _c and b _o are weight parameters, c _t is memory cells, and h _t is output results.

In said S3, the CNN layer is provided with a second CNN module, and the multi-scale local context feature fusion representation M is obtained. The expression of _M3 is specifically:

M ₃ =A ₂ (b ₂ +C ₂ (H))

In the formula, H is the expression below, C ₂ (·) is the second CNN module, A ₂ is the second activation function, and b ₂ is the weight of the second CNN module.

In this embodiment, the structure of the second CNN module is shown in FIG. 5 . Inputting the context representation to the second CNN module can strengthen the correlation between semantics and generate a multi-scale fusion representation of local context features.

The multi-scale local context feature fusion representation is input to the CRF layer to complete the task of sequence labeling and then complete the named entity recognition of Chinese cultural tourism.

The beneficial effects of the present invention are as follows: a method for recognizing named entities of Chinese cultural and tourism categories provided by the present invention solves the problem that less attention is paid to the research on named entity recognition of Chinese tourism categories, and is aimed at Chinese cultural and tourism text data Carry out network construction, use the first CNN module at the character embedding layer to learn Chinese-based radical-level embedded representations, and obtain character vector representations that are conducive to Chinese recognition, and use the second CNN module at the CNN layer to learn multi-scale local context feature fusion representations , strengthen the correlation between semantics, and further improve the feature representation that is beneficial to Chinese recognition.

In describing the present invention, it is to be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", " The orientation or positional relationship indicated by "radial" and so on is based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred equipment or elements must have a specific orientation, Constructed and operative in a particular orientation and therefore are not to be construed as limitations of the invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the number of technical features. Therefore, a feature defined by "first", "second" and "third" may explicitly or implicitly include one or more of these features.

Claims (8)

1. A named entity recognition method of Chinese culture and travel class, is characterized in that, comprises the following steps: S1. Obtain the text data of Chinese culture and tourism, and input it into the character embedding layer to obtain the character vector representation; S2. Input the character vector representation to the bidirectional long short-term memory network layer to obtain the context representation; S3. Input the context representation to the CNN layer to obtain multi-scale local context feature fusion representation; S4. Input the multi-scale local contextual feature fusion representation to the CRF layer, and perform sequence annotation through the CRF layer to complete the named entity recognition of Chinese culture and tourism. 2. The named entity recognition method of Chinese cultural tourism class according to claim 1, characterized in that, in said S1, the character embedding layer includes a parallel ChineseBert module and the first CNN module; The S1 includes the following sub-steps: S11. Acquiring Chinese cultural and tourism text data; S12. Input the Chinese cultural travel text data into the ChineseBert module, and obtain the word embedding vector representation of each word in the Chinese cultural travel text data; S13. Input the Chinese cultural tourism text data into the first CNN module to obtain a radical-level embedding representation; S14. Concatenate the word embedding vector representation with the radical-level embedding representation to obtain a character vector representation. 3. The named entity recognition method of Chinese cultural tourism according to claim 2, wherein said S12 is specifically: Input the Chinese cultural tourism text data into the ChineseBert module, encode and express the input Chinese cultural tourism text data through the ChineseBert module, obtain the feature vector, and generate the word embedding vector of each word in the Chinese cultural tourism text data according to the feature vector express; Wherein, the feature vector includes label embedding, position embedding and segment embedding. 4. The named entity recognition method of Chinese cultural tourism class according to claim 2, characterized in that, in said S13, the expression of obtaining the radical-level embedding representation _M2 is specifically: M ₂ =A ₁ (b ₁ +C ₁ (x)) In the formula, x is the radical-level feature of Chinese characters, C ₁ (·) is the first CNN module, A ₁ is the first activation function, and b ₁ is the weight of the first CNN module. 5. The named entity recognition method of Chinese cultural tourism class according to claim 4, characterized in that, in said S14, the expression obtained by character vector representation Z _concat is specifically: Z _concat =M ₁ +M ₂ In the formula, M ₁ is the word embedding vector representation. 6. The named entity recognition method for Chinese cultural tourism according to claim 1, wherein in said S2, the two-way long-short-term memory network layer includes first to twelfth LSTM units, and said first to twelfth The six LSTM units forwardly process the input character vector representations, and the seventh to twelfth LSTM units reversely process the input character vector representations; The method of obtaining the context representation is as follows: Splicing is performed according to the output results of the first to twelfth LSTM units to obtain the context representation. 7. The named entity recognition method of Chinese cultural tourism class according to claim 6, characterized in that, in said S2, the expression for obtaining the context representation H is specifically: H＝{h ₁ ,...,h _ti ,...,h _D } In the formula, h _ti is the splicing of the output results of the first to twelfth LSTM units, ti is the sequence number of splicing, and ti=1,...,D, D is the dimension represented by the character vector; The first to twelfth LSTM units all include an input gate _it , an output gate o _t and a forgetting gate f _t , the expressions of which are specifically as follows: i _t = σ(W _xi x _t +W _hi h _t-1 +W _ci c _t-1 +b _i ) f _t ＝σ(W _xf x _t +W _hf h _t-1 +W _cf c _t-1 +b _f ) c _t ＝f _t ⊙c _t-1 +i _t ⊙tanh(W _xc x _t +W _hc h _t-1 +b _c ) o _t ＝σ(W _xo x _t +W _ho h _t-1 +W _co c _t +b _o ) h _t ＝o _t ⊙tanh(c _t ) In the formula, σ(·) is the per-element sigmoid function, tanh(·) is the hyperbolic tangent function, ⊙ is the per-element multiplication function, W _xi , W _hi , W _ci , W _xf , W _hf , W _cf , W _xc , W _hc , W _xo , _Who and W _co are weight parameters, b _i , b _f , b _c and b _o are weight parameters, c _t is memory cells, and h _t is output results. 8. The named entity recognition method of Chinese cultural tourism according to claim 1, wherein in said S3, the CNN layer is provided with a second CNN module to obtain multi-scale local context feature fusion representation _M3 expression The specific formula is: M ₃ =A ₂ (b ₂ +C ₂ (H)) In the formula, H is the expression below, C ₂ (·) is the second CNN module, A ₂ is the second activation function, and b ₂ is the weight of the second CNN module.