Text data information extraction-oriented multi-model iterative optimization method

Abstract

The invention provides a multi-model iterative optimization method oriented to text data information extraction. The method comprises the following steps: step 1, constructing a text information extraction model based on multiple models; 2, training and optimizing the text information extraction model constructed in the step 1 by using an iterative optimization training method based on semi-supervision; and step 3, performing text data information extraction based on multi-model iterative optimization by using the text information extraction model trained in the step 2. According to the method provided by the invention, the text entity category in the labeling range can be effectively identified, the algorithm effect can be continuously optimized, and the method is effectively applied to a Chinese text data information extraction scene.

Description

Technical Field

[0001] This invention relates to a multi-model iterative optimization method, and more particularly to a multi-model iterative optimization method for extracting information from text data. Background Technology

[0002] In the era of big data, the Internet and various information systems store massive amounts of data of various types. How to effectively utilize this data and mine its value has become an important topic for exploration in various industries.

[0003] Current entity recognition models mainly use general model methods for common entity recognition, such as BERT model, LSTM model and other end-to-end algorithm recognition models, but they support a limited number of Chinese entity categories (such as personal names, place names, time, location), and the effect of entity recognition on custom corpora extended to professional fields is limited.

[0004] Training algorithm models requires extensive data for continuous trial and error and reconstruction to improve accuracy, F1 score, recall, and other metrics. Therefore, training algorithm models more accurately and efficiently is of paramount importance. Summary of the Invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a multi-model iterative optimization method for extracting information from text data, which addresses the shortcomings of the existing technology.

[0006] To address the aforementioned technical problems, this invention discloses a multi-model iterative optimization method for extracting information from text data, comprising the following steps:

[0007] Step 1: Construct a text information extraction model based on multiple models;

[0008] Step 2: Use a semi-supervised iterative optimization training method to train and optimize the text information extraction model built in Step 1.

[0009] Step 3: Use the text information extraction model trained in Step 2 to perform text data information extraction based on multi-model iterative optimization.

[0010] Furthermore, the construction of the multi-model-based text information extraction model described in step 1 specifically includes the following steps:

[0011] Step 1-1: Preprocess the target text by removing abnormal characters and stop words, and convert it into the corresponding vector to obtain the target text vector;

[0012] Steps 1-2: Input the target text vector into a multi-model-based entity recognition model to obtain the entity label recognition result of the target text; the label result is a sub-result composed of one or more label elements.

[0013] Furthermore, the preprocessing of the target text described in step 1-1 specifically includes:

[0014] Step 1-1-1: Perform preliminary filtering on the original input target text, including at least character filtering and stop word removal, and convert characters not within the preset word vector dictionary into general strings;

[0015] Step 1-1-2: Construct the data preprocessing vector, which is to convert the preprocessed target text into a target text vector, as follows:

[0016] Step 1-1-2-1: Organize the initially filtered target text into sentence pairs;

[0017] Step 1-1-2-2, let the character text sequence of any statement pair be represented as: M = [m1, m2, ..., m i ,…,m N ], where m i This represents the i-th single character. For each single character, a lookup and mapping are performed to obtain the original word vector sequence T = [t1, t2, ..., t]. i ,…,t N ], where t i It is a word vector;

[0018] Steps 1-1-2-3: Based on whether each word belongs to the current sentence or the next sentence, construct a sentence category vector sequence: S = [s1, s2, ..., s...]. i ,…,s N ], where s i The statement type is expressed using binary vectors;

[0019] Step 1-1-2-4: Encode the position of each individual character in the statement to obtain the positional encoding vector sequence: P = [s1, s2, ..., s N ];

[0020] Step 1-1-2-5: Add the above three vector sequences together to obtain the vectorized text result, which is the target text vector, also known as the embedded vector: E = T + S + P.

[0021] Furthermore, the method for organizing the initially filtered target text into sentence pairs as described in step 1-1-2-1 is as follows:

[0022] If statement b follows statement a, and statements a and b have a contextual relationship, then statements a and b constitute a statement pair. A sample statement pair is constructed based on this, by adding an identifier indicating the beginning of the sequence before statement a. <cls>Add a separator to the end of statement a and statement b. <sep>The statement for constructing the sample sequence c is as follows:

[0023] c = [ <cls> ,a, <sep> ,b, <sep>]

[0024] Organize the sample sequences of all constructed statement pairs into a sample set C.

[0025] Furthermore, the multi-model-based entity recognition model described in steps 1-2 specifically includes:

[0026] Step 1-2-1, word vector extraction model, that is: convert the text sentences in the target text vector obtained in Step 1-1 into a sequence of word vectors;

[0027] Step 1-2-2, the first entity feature recognition model, namely: extracting the positional features of the word vector sequence from the forward and reverse order relationships in the word vector sequence;

[0028] Steps 1-2-3, the second entity feature recognition model, namely: updating the weights of the word vector sequence and obtaining the entity label prediction results based on the weights;

[0029] Steps 1-2-4: Fully connected output model, that is: based on the entity label prediction results output by the second entity feature recognition model, obtain the entity label recognition results of the target text, and complete the text data information extraction.

[0030] Furthermore, the word vector extraction model described in step 1-2-1, namely: fine-tuning and prediction based on the BERT Chinese pre-trained model, specifically includes:

[0031] The vector E obtained after processing in steps 1-1-2-5 is then passed through a Transformer multi-layer bidirectional encoder to output the word vector X corresponding to each individual character. i , where i = 1, 2, 3, ..., N.

[0032] Furthermore, the first entity feature recognition model mentioned in step 1-2-2, namely, using the BiLSTM model as the first entity feature recognition model for prediction, specifically includes:

[0033] The BiLSTM model calculates two different hidden layer representations for each word vector sequence using both sequential and reverse methods. These representations are then concatenated to obtain the final hidden layer representation, as detailed below:

[0034] h t =f(w1x t +w2h t-1 )

[0035] h′ t =f(w3x) t +w5h′ t+1 )

[0036] o t =g(w4h) t +w6h′ t )

[0037] Among them, h t h′ represents the output vector at position t of the forward hidden layer. t This represents the output vector at position t of the inverse hidden layer, o t denoted as the output vector after merging the bidirectional hidden layers; f is the calculation formula for a single layer in the LSTM network at a single time step in the BiLSTM model; g is the calculation formula for merging the bidirectional results of BiLSTM; w1 to w6 are the calculation weights, respectively.

[0038] The word vector X calculated in step 1-2-1 i (i = 1, 2, 3, ..., N) are input into the BiLSTM model above to obtain word vectors H containing contextual relationship information. i (i = 1, 2, 3, ..., N).

[0039] Furthermore, the second entity feature recognition model described in steps 1-2-3, namely, using the Attention model as the second entity feature recognition model, specifically includes:

[0040] For the input word vector H i Create three distinct vector matrices: a query matrix Q, a value matrix K, and an index matrix V, and perform the following calculations:

[0041]

[0042] Where, d K The dimension of the value matrix K is represented by , and softmax represents the softmax function.

[0043] The word vector H obtained in step 1-2-2 i After being input into the Attention model described above, the Attention word vector A is obtained. i , where i = 1, 2, 3, ..., N.

[0044] Furthermore, the fully connected output model described in steps 1-2-4 specifically includes:

[0045] Step 1-2-4-1, CRF model prediction, as follows:

[0046] The CRF model uses a linear chain model to learn the associations between labels. The sentence scoring function for the CRF output sequence is as follows:

[0047]

[0048] Where a is the Attention word vector A obtained in steps 1-2-3 of the input sequence. i y is the corresponding label value, which is the numerical encoding of all BIO category labels; Representing the yth i The score of each tag;

[0049] Calculate the probability of the sequence composed of all labels, select and output the final label, as follows:

[0050]

[0051] Among them, Y a P(y|a) represents the set of all possible label sequences; P(y|a) represents the value with the highest output probability, that is, the y-th label in the sequence. i Each label is used as the final output;

[0052] Step 1-2-4-2: After obtaining the tag output results, the tags are converted using the numerical encoding of the BIO category tags and the original tag name dictionary. The start identifier and split identifier tags are removed, and the associated tags are merged according to the tag relationships of BI to obtain the final entity tag recognition results.

[0053] Furthermore, the semi-supervised iterative optimization training method described in step 2 specifically includes:

[0054] Step 2-1: On the server side, configure the multi-model-based text information extraction model described in Step 1.

[0055] Step 2-2: Push N target texts to be extracted to the message queue on the client.

[0056] Step 2-3: The server retrieves data from the message queue, performs data pre-labeling on the target text according to the text information extraction model configured in Step 2-1, and returns it to the client.

[0057] Steps 2-4: The client manually confirms the results of the data pre-labeling process. The manual confirmation includes at least: re-labeling, deletion, and modification; and adds the re-labeled pre-labeling results to the whitelist and the deleted pre-labeling results to the blacklist.

[0058] Steps 2-5 update the pre-labeled data processing structure after manual confirmation into a feature dataset for iterative training;

[0059] Steps 2-6 are repeated from steps 2-3 to 2-5 until all target text pushed by the client has been processed.

[0060] Step 2-7, iterative training of the model: using the feature dataset finally obtained in step 2-5, the text information extraction model based on multiple models configured in step 2-1 is trained.

[0061] Steps 2-8: Repeat steps 2-2 to 2-7. Each execution will create a new version of the text information extraction model.

[0062] Steps 2-9: The client selects and enables the optimal text information extraction model based on all versions, and completes the semi-supervised iterative optimization training.

[0063] Beneficial effects:

[0064] The method proposed in this invention can continuously improve the accuracy of information extraction. This method can effectively identify the text entity categories within the labeled range and continuously optimize the algorithm effect, making it effective for Chinese text data information extraction scenarios. Attached Figure Description

[0065] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0066] Figure 1 This is a schematic diagram illustrating the principle of a multi-model process.

[0067] Figure 2 A schematic diagram of the iterative optimization training process. Detailed Implementation

[0068] This invention proposes a multi-model iterative optimization method for extracting information from text data, as shown in the appendix. Figure 1 As shown, the main steps include: target text preprocessing, word vector extraction, entity feature recognition 1 model prediction, entity feature recognition 2 model prediction, and fully connected output model prediction output.

[0069] Appendix Figure 1 The first step is data preprocessing. This preprocessing step begins by filtering the original text, including character filtering, stop word removal, and converting characters outside the word vector dictionary into generic strings. Next, preprocessed vectors are constructed. Since the original text cannot be directly input into a deep learning model, it needs to be converted into word vectors. This preprocessing step introduces positional information into embedded vectors based on regular word vectors, facilitating input into the subsequent BERT model. The filtered text needs to be organized into multiple sentence pairs. Each sentence pair begins with [CLS] as a start identifier, and each sentence within the pair ends with [SEP] as a sentence segmentation identifier. Then, embedded vectors are extracted from each sentence pair, including word embeddings, segment embeddings, and position embeddings. Assume a sentence pair's character text sequence is represented as M = [m1, m2, ..., m...]. N ], where m i Representing a single character, the WORD2Vec model is used to look up and map each character, resulting in the original word vector sequence T = [t1, t2, ..., t]. N ], in the original word vector sequence t i It can be a WORD2Vec word vector with 200-300 dimensions. The WORD2Vec model can be trained using a large amount of original Chinese corpus. Based on whether each word belongs to the current sentence or the next sentence, a sentence category vector S = [s1, s2, ..., s...] is constructed. N The s i The statement type is represented by a binary vector; the position of each character in the statement is encoded to obtain the position encoding vector P = [s1, s2, ..., s...]. N Finally, the above three vectors are added together to obtain the vectorized text result: embedded vector E = T + S + P.

[0070] Appendix Figure 1 The word vector extraction includes: fine-tuning and model prediction based on the BERT Chinese pre-trained model. The embedded vector E is used as the data input to the BERT Chinese pre-trained model, and after processing and encoding by a Transformer multi-layer bidirectional encoder, the word vector X corresponding to each character is output. i (i = 1, 2, 3, ..., N). The BERT model employs two unsupervised tasks for pre-training: a masked language model, which randomly masks a certain percentage of the input tokens in each sentence and then predicts the masked tokens based on the context (the remaining tokens); and a next-sentence prediction model, which, to train a model capable of understanding sentence relationships, selects two sentences in the training data. A certain percentage of the data represents the next sentence of the first sentence, while the remaining data consists of two randomly selected sentences, and the model determines whether the second sentence follows the first sentence. Based on the existing BERT-Chinese-based pre-trained model, business-related corpora and corresponding preprocessed embedded vectors are input. After training and fine-tuning, a BERT-based Chinese pre-trained fine-tuned model is obtained. This model takes Chinese preprocessed embedded vectors as input and outputs Chinese word vectors containing sentence relationships.

[0071] Appendix Figure 1 The BiLSTM model corresponds to the entity feature recognition model 1 prediction. The BiLSTM model is composed of a forward LSTM and a backward LSTM. For each sentence, the BiLSTM model calculates two different sets of hidden layer representations using both sequential and reverse methods, and then concatenates the vectors to obtain the final hidden layer representation. Its calculation formula is as follows:

[0072] h t =f(w1x t +w2h t-1 )

[0073] h′ t =f(w3x) t +w5h′ t+1 )

[0074] o t =g(w4h) t +w6h′ t )

[0075] Where h t h′ represents the output vector at position t of the forward hidden layer. t This represents the output vector at position t of the inverse hidden layer, o t This represents the output vector after merging the bidirectional hidden layers. f, g, and w1 to w6 are the calculation formulas for a single LSTM layer at a single time step, the calculation formula for merging the bidirectional results of the BiLSTM, and the calculation weights, respectively. The word vector X calculated by the BERT model is then used. i (i = 1, 2, 3, ..., N) are input into the BiLSTM model, and after computation and processing, a BiLSTM vector H containing contextual relationship information is obtained. i (i = 1, 2, 3, ..., N).

[0076] Appendix Figure 1 The Attention model corresponds to the entity feature recognition model 2. The core of the Attention model is the self-attention mechanism, which assigns weights to each input word vector and continuously learns and updates them, giving greater weights to important words, indicating that the word plays a greater role in the sentence, so that named entities and ordinary text labels can be more easily classified. During the Attention calculation process, for each input word vector H... i Create three distinct vector matrices: query matrix Q, value matrix K, and index matrix V, and calculate their corresponding model outputs using the following formula:

[0077]

[0078] In the formula d K The dimension of the value matrix K is represented by H. The BiLSTM vector H... i (i = 1, 2, 3, ..., N) are input into the Attention model and processed to obtain the Attention word vector A. i (i = 1, 2, 3, ..., N).

[0079] The fully connected output model prediction steps include: CRF model prediction, dictionary transformation, and entity label merging. The CRF model uses a linear chain model, which helps learn the relationships between labels and improves the accuracy of entity label merging. The sentence scoring function for the CRF output sequence is defined as follows:

[0080]

[0081] Where 'a' is the input sequence, which is the Attention word vector A obtained after processing by the Attention model. i y is the tag value corresponding to a, which is the numerical encoding of all BIO category tags; P i,yi This represents the score of the first label. The probability of the label sequence is calculated using the following formula:

[0082]

[0083] Among them, Y a represents the set of all possible tag sequences; P(y|a) represents the value with the highest output probability, that is, taking the tag of the sequence yi as the final output. After obtaining the tag output result, it is also necessary to convert it through the numerical encoding of the B-I-O category tags and the original tag name dictionary. Taking the original text "this person Wang Moumou" as an example, after calculation, a B-I-O tag sequence similar to but not limited to "B_NAME I_NAME I_NAME O O O" will be obtained. This tag corresponds one-to-one with the original string. NAME is the entity category, B_ represents the start position of an entity, I_ represents the subsequent string position of the entity, and O represents other categories. Finally, it is also necessary to merge the associated tags according to the B-I tag relationship to obtain the "NAME O O O" tag, corresponding to the text tokenization result "this person Wang Moumou".

[0084] For the above multi-model based text information extraction method, a semi-supervised iterative optimization training method is proposed as shown in the appendix Figure 2 As follows. After preprocessing the original corpus through algorithms and then through manual annotation confirmation and fine-tuning, an incremental training data set can be quickly generated to support iterative training of the algorithm. The whole process forms a complete data closed-loop and can continuously iterate and train to optimize the model. Specifically, it includes the following steps:

[0085] The first step: The server configures the algorithms required for processing the text.

[0086] The second step: The client pushes N txt texts to the message queue.

[0087] The third step: The server consumes the data in the message queue and performs data pre-annotation processing on the text through the algorithms configured in the first step.

[0088] The fourth step: The client opens the txt text preprocessed by the algorithm, highlights the algorithm preprocessing result, and then performs manual confirmation. The algorithm preprocessing result supports operations such as re-annotation, deletion, and modification. Add the re-annotated feature data to the whitelist and the deleted feature data to the blacklist.

[0089] The fifth step: Submit the feature data set after manual confirmation for subsequent model iterative training.

[0090] The sixth step: Repeat the third step, the fourth step, and the fifth step until all the text data pushed in the second step is consumed.

[0091] The seventh step: Model iterative training, push the feature data set to the algorithm model training interface configured in the first step.

[0092] The eighth step: Repeat the second step, the third step, the fourth step, the fifth step, the sixth step, and the seventh step. Each time it is executed, a new algorithm model version is formed.

[0093] Step 9: Browse all model versions on the client and enable the version that performs best.

[0094] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a multi-model iterative optimization method for extracting text data, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0095] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0096] This invention provides a multi-model iterative optimization method for extracting information from text data. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.< / sep> < / sep> < / cls> < / sep> < / cls>

Claims

1. A multi-model iterative optimization method for extracting information from text data, characterized in that, Includes the following steps: Step 1: Construct a text information extraction model based on multiple models; Step 2: Use a semi-supervised iterative optimization training method to train and optimize the text information extraction model built in Step 1. Step 3: Use the text information extraction model trained in Step 2 to perform text data information extraction based on multi-model iterative optimization.

2. The multi-model iterative optimization method for extracting information from text data according to claim 1, characterized in that, The construction of the multi-model-based text information extraction model described in step 1 specifically includes the following steps: Step 1-1: Preprocess the target text by removing abnormal characters and stop words, and convert it into the corresponding vector to obtain the target text vector. Steps 1-2: Input the target text vector into a multi-model-based entity recognition model to obtain the entity label recognition result of the target text; the label result is a sub-result composed of one or more label elements.

3. The multi-model iterative optimization method for extracting information from text data according to claim 2, characterized in that, The preprocessing of the target text described in step 1-1 specifically includes: Step 1-1-1: Perform preliminary filtering on the target text of the original input, including at least: character filtering and removal of stop words, and convert characters not within the preset word vector dictionary into general strings; Step 1-1-2: Construct the data preprocessing vector, which is to convert the preprocessed target text into a target text vector, as follows: Step 1-1-2-1: Organize the initially filtered target text into sentence pairs; Step 1-1-2-2, let the character text sequence of any statement pair be represented as: M = [m1, m2, ..., m i ,…,m N ], where m i This represents the i-th single character. For each single character, a lookup and mapping are performed to obtain the original word vector sequence T = [t1, t2, ..., t]. i ,…,t N ], where t i It is a word vector; Steps 1-1-2-3: Based on whether each word belongs to the current sentence or the next sentence, construct a sentence category vector sequence: S = [s1, s2, ..., s...]. i ,…,s N ], where s i The statement type is expressed using binary vectors; Step 1-1-2-4: Encode the position of each individual character in the statement to obtain the positional encoding vector sequence: P = [p1, p2, ..., p i ,…,p N ], where p i This indicates the position of the i-th word in the sentence; Step 1-1-2-5: Add the above three vector sequences together to obtain the vectorized text result, which is the target text vector, also known as the embedded vector: E = T + S + P.

4. The multi-model iterative optimization method for extracting information from text data according to claim 3, characterized in that, The method for organizing the initially filtered target text into sentence pairs as described in step 1-1-2-1 is as follows: If statement b follows statement a, and statements a and b have a contextual relationship, then statements a and b constitute a statement pair. A sample statement pair is constructed based on this, by adding an identifier indicating the beginning of the sequence before statement a. <cls>Add a separator to the end of statement a and statement b. <sep> The statement for constructing the sample sequence c is as follows:< / sep> < / cls> c＝[ <cls> ,a, <sep> ,b, <sep> ]< / sep> < / sep> < / cls> Organize the sample sequences of all constructed statement pairs into a sample set C.

5. The multi-model iterative optimization method for extracting text data information according to claim 4, characterized in that, The multi-model-based entity recognition model described in steps 1-2 specifically includes: Step 1-2-1, word vector extraction model, that is: convert the text sentences in the target text vector obtained in Step 1-1 into a sequence of word vectors; Step 1-2-2, the first entity feature recognition model, namely: extracting the positional features of the word vector sequence from the forward and reverse order relationships in the word vector sequence; Steps 1-2-3, the second entity feature recognition model, namely: updating the weights of the word vector sequence and obtaining the entity label prediction results based on the weights; Steps 1-2-4: Fully connected output model, that is: based on the entity label prediction results output by the second entity feature recognition model, obtain the entity label recognition results of the target text, and complete the text data information extraction.

6. The multi-model iterative optimization method for extracting information from text data according to claim 5, characterized in that, The word vector extraction model described in step 1-2-1, namely: fine-tuning and prediction based on the BERT Chinese pre-trained model, specifically includes: The vector E obtained after processing in steps 1-1-2-5 is then passed through a Transformer multi-layer bidirectional encoder to output the word vector X corresponding to each individual character. i , where i = 1, 2, 3, ..., N.

7. The multi-model iterative optimization method for extracting information from text data according to claim 6, characterized in that, The first entity feature recognition model mentioned in step 1-2-2, namely, using the BiLSTM model as the first entity feature recognition model for prediction, specifically includes: The BiLSTM model calculates two different hidden layer representations for each word vector sequence using both sequential and reverse methods. These representations are then concatenated to obtain the final hidden layer representation, as detailed below: h t =f(w1x t +w2h t-1 ) h′ t =f(w3x t +w5h′ t+1 ) o t =g(w4h t +W6h′ t ) Among them, h t h′ represents the output vector at position t of the forward hidden layer. t This represents the output vector at position t of the inverse hidden layer, o t denoted as the output vector after merging the bidirectional hidden layers; f is the calculation formula for a single layer in the LSTM network at a single time step in the BiLSTM model; g is the calculation formula for merging the bidirectional results of BiLSTM; w1 to w6 are the calculation weights, respectively. The word vector X calculated in step 1-2-1 i (i = 1, 2, 3, ..., N) are input into the BiLSTM model above to obtain word vectors H containing contextual relationship information. i (i = 1, 2, 3, ..., N).

8. The multi-model iterative optimization method for extracting information from text data according to claim 7, characterized in that, The second entity feature recognition model described in steps 1-2-3, namely, using the Attention model as the second entity feature recognition model, specifically includes: For the input word vector H i Create three distinct vector matrices: a query matrix Q, a value matrix K, and an index matrix V, and perform the following calculations: Where, d K The dimension of the value matrix K is represented by , and softmax represents the softmax function. The word vector H obtained in step 1-2-2 i After being input into the Attention model described above, the Attention word vector A is obtained. i , where i = 1, 2, 3, ..., N.

9. A multi-model iterative optimization method for extracting information from text data according to claim 8, characterized in that, The fully connected output model described in steps 1-2-4 specifically includes: Step 1-2-4-1, CRF model prediction, as follows: The CRF model uses a linear chain model to learn the associations between labels. The sentence scoring function for the CRF output sequence is as follows: Where a is the Attention word vector A obtained in steps 1-2-3 of the input sequence. i y is the corresponding label value, which is the numerical encoding of all BIO category labels; Representing the yth i The score of each tag; Calculate the probability of the sequence composed of all labels, select and output the final label, as follows: Among them, Y a P(y|a) represents the set of all possible label sequences; P(y|a) represents the value with the highest output probability, that is, the y-th label in the sequence. i Each label is used as the final output; Step 1-2-4-2: After obtaining the tag output results, the tags are converted using the numerical encoding of the BIO category tags and the original tag name dictionary. The start identifier and split identifier tags are removed, and the associated tags are merged according to the tag relationships of BI to obtain the final entity tag recognition results.

10. A multi-model iterative optimization method for extracting information from text data according to claim 9, characterized in that, The semi-supervised iterative optimization training method described in step 2 specifically includes: Step 2-1: On the server side, configure the multi-model-based text information extraction model described in Step 1. Step 2-2: Push N target texts to be extracted to the message queue on the client. Step 2-3: The server retrieves data from the message queue, performs data pre-labeling on the target text according to the text information extraction model configured in Step 2-1, and returns it to the client. Steps 2-4: The client manually confirms the results of the data pre-labeling process. The manual confirmation includes at least: re-labeling, deletion, and modification; and adds the re-labeled pre-labeling results to the whitelist and the deleted pre-labeling results to the blacklist. Steps 2-5 update the pre-labeled data processing structure after manual confirmation into a feature dataset for iterative training; Steps 2-6 are repeated from steps 2-3 to 2-5 until all target text pushed by the client has been processed. Step 2-7, iterative training of the model: using the feature dataset finally obtained in step 2-5, the text information extraction model based on multiple models configured in step 2-1 is trained. Steps 2-8: Repeat steps 2-2 to 2-7. Each execution will create a new version of the text information extraction model. Steps 2-9: The client selects and enables the optimal text information extraction model based on all versions, and completes the semi-supervised iterative optimization training.

Images