Evaluation object extraction method, device, equipment, storage medium and program product

Abstract

The present application discloses a method, device, equipment, storage medium and program product for extracting evaluation objects, the method comprising: obtaining evaluation data to be processed; determining dependency information between words, word information and part-of-speech information of words in the evaluation data to be processed, and inputting them into a preset extraction model; wherein the preset extraction model is obtained by first iterative training of a first preset model to be trained based on first training data marked with a first preset label, the first training data comprising a training sentence and a preset concatenated dependency vector corresponding to the training sentence; object extraction processing is performed on the dependency information, word part-of-speech information and word information of words based on the preset extraction model to obtain a target evaluation object. In the present application, implicit precision analysis is considered from various dimensions, thereby improving the extraction precision of the evaluation object. The technical problem of low extraction precision in the existing extraction of evaluation objects by sequence labeling is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to an evaluation object extraction method, device, equipment, storage medium and program product. Background Art

[0002] With the continuous development of financial technology, especially Internet technology finance, more and more technologies (such as distributed, blockchain, artificial intelligence, etc.) are applied in the financial field, but the financial industry also puts forward higher requirements on technology, such as the financial industry also has higher requirements for the extraction of evaluation objects.

[0003] With the continuous development of computer software and artificial intelligence, the application of artificial intelligence in the Internet industry is becoming more and more extensive. In the Internet industry, each company has many products, and these products have more or less evaluations or comments from users. The comments contain very valuable evaluation object information such as user preferences and product defects. Extracting evaluation objects from comments can, on the one hand, let the company know which aspects of the product users care more about. On the other hand, the extraction of evaluation objects is also part of fine-grained sentiment analysis. Accurately extracting evaluation objects can effectively improve the accuracy of fine-grained sentiment analysis to better optimize products. However, when the evaluation objects are currently sequenced, the dependency between different words is not considered, resulting in the technical problem that the extraction of evaluation objects is difficult to meet the extraction accuracy requirements. Summary of the invention

[0004] The main purpose of this application is to provide an evaluation object extraction method, device, equipment, storage medium and program product, aiming to solve the technical problem of low extraction accuracy of evaluation objects in the prior art.

[0005] To achieve the above object, the present application provides an evaluation object extraction method, which is applied to a first device. The evaluation object extraction method includes:

[0006] Get the evaluation data to be processed;

[0007] Determine the dependency relationship information between words, word information and part-of-speech information of the words in the evaluation data to be processed;

[0008] Inputting the dependency information, word part information and word information into a preset extraction model;

[0009] The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence;

[0010] Based on the preset extraction model, object extraction processing is performed on the dependency information, word part-of-speech information and word information to obtain a target evaluation object.

[0011] Optionally, before the step of performing object extraction processing on the dependency information, word part-of-speech information and word information of a word based on the preset extraction model to obtain a target evaluation object, the method includes:

[0012] Acquire a training sentence to obtain first training data marked with a first preset label;

[0013] Based on the first training data annotated with the first preset label, performing a first iterative training on the first preset model to be trained to obtain a first target model;

[0014] The first target model is set as the preset extraction model.

[0015] Optionally, the step of performing a first iterative training on a first preset model to be trained based on the first training data annotated with a first preset label to obtain a first target model includes:

[0016] Obtain a first preset model to be trained;

[0017] Inputting the first training data into the first preset model to be trained to perform object label analysis on the first training data to obtain a first predicted label;

[0018] Calculating a first model error based on the first predicted label and the first preset label;

[0019] Based on the first model error, the first preset model to be trained is updated until the first preset model to be trained meets a first preset update end condition to obtain a first target model.

[0020] Optionally, the step of acquiring a training sentence to obtain first training data marked with a first preset label includes:

[0021] Obtaining a training sentence, and inputting the training sentence into a preset dependency syntax model;

[0022] Based on the preset dependency syntax model, the training sentence is subjected to dependency syntax analysis between words to obtain a target dependency syntax analysis result;

[0023] Determining a preset concatenated dependency vector of the training sentence based on the target dependency syntactic analysis result;

[0024] First training data is determined based on the preset concatenated dependency vector and the training sentence marked with a first preset label.

[0025] Optionally, before the step of performing dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result, the method includes:

[0026] Acquire second training data and a second preset to-be-trained model annotated with a second preset label, wherein the second preset label is a dependency label;

[0027] Based on the second training data annotated with the second preset label, performing a second iterative training on the second preset model to be trained to obtain a second target model;

[0028] The second target model is set as the preset dependency syntax model.

[0029] Optionally, the step of performing a second iterative training on the second preset model to be trained based on the second training data annotated with the second preset label to obtain a second target model includes:

[0030] Inputting the second training data into the second preset model to be trained to perform dependency relationship label analysis between words on the second training data to obtain a second predicted label;

[0031] Calculating a second model error based on the second predicted label and the second preset label;

[0032] Based on the second model error, the second preset model to be trained is updated until the second preset model to be trained meets a second preset update end condition to obtain a second target model.

[0033] Optionally, the step of determining the dependency relationship information between words in the evaluation data to be processed includes:

[0034] Inputting the evaluation data to be processed into a preset dependency syntax model;

[0035] Based on the preset dependency syntax model, dependency syntax analysis is performed on the evaluation data to be processed to obtain dependency relationship information between words in the evaluation data to be processed.

[0036] Optionally, the preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model.

[0037] The step of performing dependency syntactic analysis on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed includes:

[0038] Vectorizing the evaluation data to be processed to obtain vectorized statements;

[0039] Based on the dependency determination model, performing dependency determination on the vectorized sentence to obtain a dependency determination result;

[0040] Based on the dependency type prediction model and the dependency discrimination result, the dependency type of the vectorized sentence is predicted to obtain dependency information between words in the evaluation data to be processed.

[0041] Optionally, the dependency discriminant model includes a first feature extraction model, a first fully connected network, a second fully connected network and a dual affine transformation network.

[0042] The step of performing dependency determination on the vectorized sentence based on the dependency determination model to obtain a dependency determination result includes:

[0043] Based on the first feature extraction model, perform feature extraction on the vectorized sentence to obtain a first feature extraction result;

[0044] Based on the first fully connected network and the second fully connected network, fully connect the first feature extraction result respectively to obtain a first sentence vector and a second sentence vector;

[0045] Based on the double affine transformation network, performing double affine transformation on the first sentence vector and the second sentence vector to obtain a dependency score matrix;

[0046] Based on the dependency score matrix, a dependency determination result is obtained.

[0047] Optionally, the vectorized sentence includes vectorized words.

[0048] The step of vectorizing the evaluation data to be processed to obtain vectorized statements includes:

[0049] Obtaining a word vector to be processed, a part-of-speech vector to be processed, and a word position vector to be processed corresponding to the evaluation data to be processed;

[0050] The vectorized word is generated based on the word vector to be processed, the part-of-speech vector to be processed and the word position vector to be processed.

[0051] The present application also provides an evaluation object extraction device, the evaluation object extraction device comprising:

[0052] A first acquisition module is used to acquire the evaluation data to be processed;

[0053] A determination module, used to determine the dependency relationship information between words, word information and part of speech information of words in the evaluation data to be processed;

[0054] An input module, used to input the dependency information, word part-of-speech information and word information into a preset extraction model;

[0055] The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence;

[0056] The object extraction module is used to perform object extraction processing on the dependency information, word part-of-speech information and word information of words based on the preset extraction model to obtain a target evaluation object.

[0057] The present application also provides an evaluation object extraction device, which is a physical device. The evaluation object extraction device includes: a memory, a processor, and a program of the evaluation object extraction method stored in the memory and executable on the processor. When the program of the evaluation object extraction method is executed by the processor, the steps of the evaluation object extraction method as described above can be implemented.

[0058] The present application also provides a readable storage medium, on which is stored a program for implementing the evaluation object extraction method. When the program of the evaluation object extraction method is executed by a processor, the steps of the evaluation object extraction method as described above are implemented.

[0059] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned evaluation object extraction method when executed by a processor.

[0060] The present application provides an evaluation object extraction method, device, equipment, storage medium and program product. Compared with the prior art that does not consider the dependency relationship between different words and directly performs sequence annotation on the evaluation objects, resulting in low extraction accuracy of the evaluation objects, the present application obtains evaluation data to be processed; determines the dependency relationship information between words in the evaluation data to be processed, word information and part-of-speech information of the words; inputs the dependency relationship information, word part-of-speech information and word information of the words into a preset extraction model; wherein the preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data annotated with a first preset label, the first training data including training sentences and preset concatenated dependency vectors corresponding to the training sentences; based on the preset extraction model, the dependency relationship information, word part-of-speech information and word information of the words are subjected to object extraction processing to obtain a target evaluation object. In the present application, after obtaining the evaluation data to be processed, the syntactic information of the evaluation data to be processed, i.e., the dependency relationship between words, word information and part-of-speech information of the words in the evaluation data to be processed, is considered. The target evaluation object is determined based on the dependency information, word information and part-of-speech information of the words. Since the dependency relationship between words is considered instead of pure sequence labeling, the target evaluation object is obtained, that is, in the present application, implicit precision analysis is considered, thereby improving the extraction precision of the evaluation object. The technical problem of low extraction precision of the evaluation object extracted by sequence labeling in the prior art is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0062] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0063] Figure 1 This is a flow chart of the first embodiment of the evaluation object extraction method of this application;

[0064] Figure 2 This is a schematic diagram of the detailed process before step S10 in the first embodiment of the evaluation object extraction method of this application;

[0065] Figure 3 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application.

[0066] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0067] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0068] The present application embodiment provides an evaluation object extraction method. In the first embodiment of the evaluation object extraction method of the present application, refer to Figure 1 , the evaluation object extraction method includes:

[0069] Step S10, obtaining evaluation data to be processed;

[0070] Step S20, determining the dependency relationship information between words in the evaluation data to be processed, the word information and the part-of-speech information of the words;

[0071] Step S30: inputting the dependency information, word part information and word information into a preset extraction model;

[0072] The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence;

[0073] Step S40, performing object extraction processing on the dependency information, word part-of-speech information and word information based on the preset extraction model to obtain a target evaluation object.

[0074] The specific steps are as follows:

[0075] Step S10, obtaining evaluation data to be processed;

[0076] In this embodiment, it should be noted that the evaluation object extraction method is applied to the evaluation object extraction device, and the evaluation object extraction device is subordinate to the evaluation object extraction device. At present, with the continuous development of computer software and artificial intelligence, the application of artificial intelligence in the Internet industry is becoming more and more extensive. In the Internet industry, each company has many products, and these products have more or less evaluations or comments from users, that is, evaluation data to be processed. The evaluation data to be processed contains very valuable evaluation object information such as user preferences and product defects. Extracting evaluation objects from comments can, on the one hand, let the company know which aspects of the product users care more about. On the other hand, the extraction of evaluation objects is also a part of fine-grained sentiment analysis. Accurately extracting evaluation objects can effectively improve the accuracy of fine-grained sentiment analysis to better optimize products. Specifically, for example, the evaluation data to be processed is: the food in Restaurant A is unpalatable, where Restaurant A is the object to be evaluated. If the evaluation data to be processed is: the taste of Restaurant A is good, but the service is poor, where the taste and service of Restaurant A are the objects to be evaluated, or if the evaluation data to be processed is: the taste of Restaurant A is good, the hygiene is good, but the service is poor, where the taste, hygiene and service of Restaurant A are the objects to be evaluated. The service is the object to be evaluated. Obviously, taste, hygiene and service as the objects to be evaluated are only finer in granularity than taste as the rating object. Or in the loan field, Weilidai is not only simple to operate, but also quickly arrives at the account, then the evaluation object is Weilidai, that is, the evaluation object in the evaluation data to be processed is part of the fine-grained sentiment analysis, which can derive the user's focus and accurately obtain the evaluation object, which can improve the user experience. However, at present, the extraction of evaluation objects is generally done as a sequence labeling task. The sequence labeling model does not consider the syntactic information in the evaluation data to be processed, that is, the dependency between words in the comments, as well as the part-of-speech information, etc., which makes it difficult to meet the requirements for the accuracy of evaluation object extraction.

[0077] Collect the evaluation data to be processed, wherein the evaluation data to be processed is composed of the evaluation data in the product, for example, multiple evaluation data in the Weilidai product (each of which is composed of sentences), or multiple evaluation data in the micro-loan product.

[0078] Step S20, determining the dependency relationship information between words in the evaluation data to be processed, the word information and the part-of-speech information of the words;

[0079] In this embodiment, in order to accurately obtain the evaluation object, after obtaining the evaluation data to be processed, the dependency information between words in the evaluation data to be processed is determined, specifically, the dependency information between words in the evaluation data to be processed, the part-of-speech information of the words in the evaluation data to be processed and the word information of the words are determined to obtain the target evaluation object, wherein the dependency information includes the syntactic information of the sentence, and the syntactic information of the sentence includes sentence pattern information and word component information. Assuming the sentence is "Who am I", the sentence pattern information indicates that the sentence is an active-object sentence, and the word component information indicates that "I" is the subject, "is" is the predicate, and "who" is the object.

[0080] Determine the dependency information between words in the evaluation data to be processed, the word information and the part-of-speech information of the words; specifically, determine the dependency information between words in the evaluation data to be processed by presetting the dependency syntax model, and obtain the word information and part-of-speech information of the words in the evaluation data to be processed by one-hot encoding, wherein the dependency information can be represented by a dependency vector, the part-of-speech information is represented by a part-of-speech vector, and the word information is represented by a word vector. In this embodiment, the dependency information between words in the evaluation data to be processed can also be determined by other methods, which are not specifically limited here.

[0081] The step of determining the dependency relationship information between words in the evaluation data to be processed includes the following steps: step S21-step S22:

[0082] Step S21, inputting the evaluation data to be processed into a preset dependency syntax model;

[0083] Step S22: performing dependency syntactic analysis on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed.

[0084] In this embodiment, dependency syntactic analysis is performed on the evaluation data to be processed based on a pre-trained dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed.

[0085] In this embodiment, in order to accurately obtain the dependency relationship information between words, a preset dependency syntax model is pre-trained. Specifically, the training method of the preset dependency syntax model can be: obtaining dependency syntax training data (second training data) and a dependency syntax model to be trained (second preset model to be trained), wherein the dependency syntax training data includes dependency syntax training sentences and preset dependency type labels corresponding to the dependency syntax training sentences; inputting the dependency syntax training data into the dependency syntax model to be trained to perform dependency syntax analysis on the dependency syntax training sentences to obtain type dependency syntax prediction labels; calculating a corresponding model error (second model error) based on the type dependency syntax prediction label and the preset dependency type label; and updating the dependency syntax model to be trained based on the corresponding model error until the preset dependency syntax model is obtained. After obtaining the preset dependency syntactic model, the evaluation data to be processed is input into the preset dependency syntactic model, and dependency syntactic analysis is performed on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed.

[0086] After obtaining the dependency information, the sentence information and word component information can be obtained. Specifically: for example, assuming the sentence "ABC", where A, B and C are all words in the sentence, then after obtaining the dependency information, that is, after performing dependency discrimination, it can be determined that B depends on A, and C depends on B. After performing dependency type prediction, it can be determined that the dependency relationship between A and B is a subject-predicate relationship, and the dependency relationship between B and C is a verb-object relationship. After obtaining the dependency information that there is a dependency relationship, the sentence information and word component information of the sentence can be further determined. Specifically, after obtaining the dependency information that there is a dependency relationship, the dependency discrimination vector and the dependency type prediction probability matrix corresponding to the sentence are obtained, and then the dependency vector and the dependency type prediction probability matrix are fused to obtain the dependency type label between words in the sentence, wherein the The dependency type label is an identifier of the dependency type, and based on the dependency type label, the sentence information and word component information of the sentence can be determined, wherein the value on each bit in the dependency type prediction probability matrix is ​​a dependency type label probability prediction vector between one word and another word in the sentence, wherein the value on each bit in the dependency type prediction vector is a probability value that the dependency relationship between one word and another word in the sentence belongs to a preset dependency relationship corresponding to the bit, wherein the preset dependency relationship includes a subject-predicate relationship, a verb-object relationship, etc. For example, assuming that the dependency type label probability prediction vector between word A and word B is (0.1, 0.9), 0.1 indicates that the probability of the subject-predicate relationship between word A and word B is 10%, and 0.9 indicates that the probability of the verb-object relationship between word A and word B is 90%.

[0087] The preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model.

[0088] The step of performing dependency syntactic analysis on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed includes:

[0089] Step M1, vectorizing the evaluation data to be processed to obtain vectorized statements;

[0090] Step M2: performing dependency determination on the vectorized sentence based on the dependency determination model to obtain a dependency determination result;

[0091] Step M3: Based on the dependency type prediction model and the dependency identification result, the dependency type of the vectorized sentence is predicted to obtain the dependency information between words in the evaluation data to be processed.

[0092] Among them, the process of performing dependency syntactic analysis on the evaluation data to be processed by the preset dependency syntactic model includes: vectorizing the evaluation data to be processed to obtain vectorized sentences; based on the dependency discrimination model in the preset dependency syntactic model, performing dependency discrimination on the vectorized sentences to obtain dependency discrimination results; based on the dependency type prediction model in the preset dependency syntactic model and the dependency discrimination results, performing dependency type prediction on the vectorized sentences to obtain the dependency syntactic analysis results. The dependency discrimination results include dependency vectors. Specifically, based on the dependency type prediction model and the dependency discrimination results, the dependency type prediction is performed on the vectorized sentences. The method of obtaining the dependency syntactic analysis results can be: based on the dependency type prediction model, the dependency type prediction is performed on the vectorized sentences to obtain a dependency type probability score matrix; the dependency type probability score matrix and the dependency vector are merged to obtain the dependency syntactic analysis results (dependency information).

[0093] Step S30, inputting the dependency information, word part-of-speech information and word information into a preset extraction model;

[0094] The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence;

[0095] Among them, refer to Figure 2Before the step of performing object extraction processing on the dependency information, word part-of-speech information and word information based on the preset extraction model to obtain the target evaluation object, the method includes the following steps S01-S03:

[0096] Step S01, obtaining a training sentence to obtain first training data marked with a first preset label;

[0097] In this embodiment, a training sentence is first obtained to obtain first training data marked with a first preset label. The first preset label can be a label obtained by annotating the sentence in a BIO manner, that is, if the evaluation object is composed of multiple words, such as the evaluation object in the sentence "WeChat Loan is not only simple to operate, but also arrives quickly", is "WeChat Loan". If the word segmentation tool divides "WeChat Loan" into "WeChat" and "Loan", the first word "WeChat" is marked as BA, the second word "Loan" is marked as IA, and other words in the sentence that do not constitute the evaluation object are marked as O, for example, the labels of the words "operation" and "simple" are "O". If the word segmentation tool divides "WeChat Loan" into "WeChat Loan", the first word "WeChat Loan" is marked as BA, and the labels of other words are "O".

[0098] Step S02, based on the first training data annotated with the first preset label, performing a first iterative training on the first preset model to be trained to obtain a first target model;

[0099] Step S03: setting the first target model as the preset extraction model.

[0100] Based on the first training data annotated with the first preset label, a first iterative training is performed on the first preset model to be trained to obtain a preset extraction model.

[0101] Specifically, the step of performing a first iterative training on the first preset model to be trained based on the first training data annotated with the first preset label to obtain a first target model includes the following steps A1 to A4:

[0102] Step A1, obtaining a first preset model to be trained;

[0103] In this embodiment, it should be noted that the first preset label is the evaluation object identifier in the pre-labeled first training data, and the first preset model to be trained is an untrained extraction model. Obtain the first preset model to be trained, wherein the first training data includes a training sentence and the first preset model to be trained corresponding to the training sentence,

[0104] Specifically, first training data and a first preset model to be trained are obtained, and the first training data are manually annotated to obtain manually annotated first training data. In this embodiment, after the manually annotated first training data is obtained, the first training data can be expanded.

[0105] Step A2, inputting the first training data into the first preset model to be trained to perform object label analysis on the first training data to obtain a first predicted label;

[0106] The first training data is input into the first preset model to be trained to perform object label analysis on the first training data to obtain a first predicted label.

[0107] Step A3, calculating a first model error based on the first predicted label and the first preset label;

[0108] In this embodiment, a first model error is calculated based on the first predicted label and the first preset label. Specifically, a distance between the first preset label and the first predicted label is calculated to obtain a first model error.

[0109] Step A4: Based on the first model error, the first preset model to be trained is updated until the first preset model to be trained meets a first preset update end condition to obtain a first target model.

[0110] In this embodiment, based on the first model error, i.e., the evaluation object extraction error, the first preset model to be trained is updated until the first preset model to be trained meets the first preset update end condition to obtain the first target model. Specifically, based on the first model error, the first gradient information is calculated, and through back propagation, the model parameters of the first preset model to be trained are updated according to the first gradient information to obtain the updated first preset model to be trained, and then it is determined whether the updated first preset model to be trained meets the first preset update end condition. If so, training continues until it meets the first preset update end condition, wherein the first preset update end condition includes reaching the maximum number of iterations and the convergence of the loss function.

[0111] Step S40, performing object extraction processing on the dependency information, word part-of-speech information and word information based on the preset extraction model to obtain a target evaluation object.

[0112] Based on the trained preset extraction model, the dependency information, word part of speech information and word information of the word are subjected to object extraction processing to obtain the target evaluation object. Or more specifically, based on the dependency information between words in the evaluation data to be processed, sentence information and word component information are obtained, and the target evaluation object is determined based on the sentence information and word component information, word part of speech information and word information of the word.

[0113] After obtaining dependency information, word part of speech information and word information, the dependency information, word part of speech information and word information are input into a preset extraction model. Since the preset extraction model is obtained by first training data marked with a first preset label and first iterative training of the first preset model to be trained, the target evaluation object is accurately obtained after object extraction processing is performed on the dependency information, word part of speech information and word information based on the preset extraction model. Specifically, the dependency information, word part of speech information and word information are used to obtain a preset concatenated dependency vector, and the target evaluation object is obtained based on the preset concatenated dependency vector, wherein the preset concatenated dependency vector refers to the word part of speech vector (the word part of speech vector is a coding vector representing the word part of speech) of the word, the word vector (the word coding vector, used to uniquely represent the word), the word vector of the word having a dependency relationship with the word, and the dependency vector (obtained through the dependency information) are concatenated after each word and part of speech in the evaluation data to be processed are vectorized.

[0114] The present application provides an evaluation object extraction method, device, equipment, storage medium and program product. Compared with the prior art that does not consider the dependency relationship between different words and directly performs sequence annotation on the evaluation objects, resulting in low extraction accuracy of the evaluation objects, the present application obtains evaluation data to be processed; determines the dependency relationship information between words in the evaluation data to be processed, word information and part-of-speech information of the words; inputs the dependency relationship information, word part-of-speech information and word information of the words into a preset extraction model; wherein the preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data annotated with a first preset label, the first training data including training sentences and preset concatenated dependency vectors corresponding to the training sentences; based on the preset extraction model, the dependency relationship information, word part-of-speech information and word information of the words are subjected to object extraction processing to obtain a target evaluation object. In the present application, after obtaining the evaluation data to be processed, the syntactic information of the evaluation data to be processed, i.e., the dependency relationship between words, word information and part-of-speech information of the words in the evaluation data to be processed, is considered. The target evaluation object is determined based on the dependency information, word information and part-of-speech information of the words. Since the dependency relationship between words is considered instead of pure sequence labeling, the target evaluation object is obtained, that is, in the present application, implicit precision analysis is considered, thereby improving the extraction precision of the evaluation object. The technical problem of low extraction precision of the evaluation object extracted by sequence labeling in the prior art is solved.

[0115] Further, based on the first embodiment of the present application, in another embodiment of the present application, the step of acquiring a training sentence to obtain first training data marked with a first preset label includes:

[0116] Step B1, obtaining a training sentence, and inputting the training sentence into a preset dependency syntax model;

[0117] Step B2, performing dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result;

[0118] Step B3, determining a preset concatenated dependency vector of the training sentence based on the target dependency syntactic analysis result;

[0119] Step B4: determining first training data based on the preset concatenated dependency vector and the training sentence marked with the first preset label.

[0120] Since the preset extraction model is obtained by training based on the first training data, and obtaining the first training data requires determining the preset splicing dependency vector of the training sentence, it is necessary to first determine the preset splicing dependency vector of the training sentence. In this embodiment, the preset splicing dependency vector of the training sentence is obtained by performing dependency syntactic analysis between words on the training sentence based on a preset dependency syntactic model, and then determining the preset splicing dependency vector of the training sentence based on the target dependency syntactic analysis result. Then, based on the preset splicing dependency vector and the training sentence marked with the first preset label, the first training data is determined.

[0121] Acquire a sentence, and perform dependency syntactic analysis on the sentence based on a preset dependency syntactic model to obtain a dependency syntactic analysis result corresponding to the sentence. Specifically, acquire a sentence, input the sentence into the preset dependency syntactic model, vectorize the sentence to obtain a vectorized sentence, and then perform dependency determination and dependency type prediction on the vectorized sentence, respectively. It should be noted that the purpose of performing dependency determination is to determine the dependency relationship between words, and the purpose of dependency type prediction is to predict the type of pre-stored relationship.

[0122] Before the step of performing dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result, the method includes the following steps C1-C3:

[0123] Step C1, obtaining second training data and a second preset model to be trained that are labeled with a second preset label, wherein the second preset label is a dependency label;

[0124] In this embodiment, it is a specific process of training a preset dependency syntactic model. First, second training data marked with a second preset label and a second preset model to be trained are obtained, wherein the second preset label is a dependency relationship label.

[0125] Step C2, based on the second training data annotated with the second preset label, performing a second iterative training on the second preset model to be trained to obtain a second target model;

[0126] Step C3: setting the second target model as the preset dependency syntax model.

[0127] Based on the second training data annotated with the second preset label, a second iterative training is performed on the second preset model to obtain a second target model, that is, a preset dependency syntactic model.

[0128] The step of performing a second iterative training on the second preset model to be trained based on the second training data annotated with the second preset label to obtain a second target model includes the following steps D1-D3:

[0129] Step D1, inputting the second training data into the second preset model to be trained, so as to perform dependency relationship label analysis between words on the second training data to obtain a second predicted label;

[0130] Step D2, calculating a second model error based on the second predicted label and the second preset label;

[0131] Step D3: Based on the second model error, the second preset model to be trained is updated until the second preset model to be trained meets a second preset update end condition to obtain a second target model.

[0132] In this embodiment, it should be noted that the second preset label is a dependency identifier in the pre-labeled second training data, and the second preset model to be trained is an untrained dependency syntax. The second training data and the second preset model to be trained are obtained, and the second training data are input into the second preset model to be trained to perform dependency label analysis on the second training data to obtain a second predicted label. Specifically, based on the vectorized network in the second preset model to be trained, the second training data is vectorized to obtain a vectorized training sentence, and then based on the dependency discrimination model (already trained) in the second preset model to be trained, the vectorized training sentence (training sentence) is subjected to dependency discrimination to obtain a training dependency vector, and based on the dependency type prediction model (already trained) in the second preset model to be trained, the vectorized training sentence is subjected to dependency type prediction to obtain a training dependency type probability score matrix, and then the training dependency vector and the training dependency type probability score matrix are combined to determine the second predicted label.

[0133] In this embodiment, based on the second predicted label and the second preset label, the second model error is calculated, specifically, the distance between the second preset label and the second predicted label is calculated to obtain the second model error. In this embodiment, based on the second model error, i.e., the evaluation object extraction error, the second preset model to be trained is updated until the second preset model to be trained meets the second preset update end condition to obtain the second target model, specifically, based on the second model error, the second gradient information is calculated, and by back propagation, the model parameters of the second preset model to be trained are updated according to the second gradient information to obtain the updated second preset model to be trained, and then it is determined whether the updated second preset model to be trained meets the second preset update end condition. If so, training is continued until it meets the second preset update end condition, wherein the second preset update end condition includes reaching the maximum number of iterations and the convergence of the loss function.

[0134] In this embodiment, a training sentence is obtained and input into a preset dependency syntax model; dependency syntax analysis is performed on the training sentence based on the preset dependency syntax model to obtain a target dependency syntax analysis result; a preset concatenated dependency vector of the training sentence is determined based on the target dependency syntax analysis result; and first training data is determined based on the preset concatenated dependency vector and the training sentence marked with the first preset label. In this embodiment, the first training data is accurately obtained, laying a foundation for accurately training the extraction model.

[0135] Further, based on the first embodiment of the present application, in another embodiment of the present application, the preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model.

[0136] The step of performing dependency syntactic analysis on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed includes the following steps E1 to E3:

[0137] Step E1, vectorizing the evaluation data to be processed to obtain vectorized statements;

[0138] In this embodiment, the evaluation data to be processed is vectorized to obtain vectorized statements.

[0139] Wherein, the vectorized sentence includes vectorized words,

[0140] The step of vectorizing the evaluation data to be processed to obtain vectorized statements includes steps G1 and G2:

[0141] Step G1, obtaining the word vector to be processed, the corresponding part-of-speech vector to be processed and the corresponding word position vector to be processed corresponding to the evaluation data to be processed;

[0142] Specifically, a word vector to be processed, a part-of-speech vector to be processed and a word position vector to be processed are generated corresponding to each word in the evaluation data to be processed, wherein the word vector to be processed is an encoding vector representing the word, which is used to uniquely represent the word, the part-of-speech vector to be processed is an encoding vector representing the part of speech of the word, and the word position vector to be processed is an encoding vector representing the position of the word in the training sentence, and then based on the word vector to be processed, the corresponding part-of-speech vector to be processed and the corresponding word position vector to be processed corresponding to each word, a vectorized word corresponding to each word is generated, and then the matrix composed of each of the vectorized words is used as the vectorized sentence.

[0143] Step G2, generating the vectorized word based on the word vector to be processed, the part-of-speech vector to be processed and the word position vector to be processed.

[0144] In this embodiment, based on the word vector to be processed (word vector), the part-of-speech vector to be processed (part-of-speech vector) and the word position vector to be processed (word position vector), the vectorized word is generated. Specifically, the word to be processed, the part-of-speech vector to be processed and the word position vector to be processed are input into a preset vectorized word calculation formula to obtain the vectorized word, wherein the preset vectorized word calculation formula is as follows:

[0145]

[0146] Among them, X i is the vectorized word, E w is the word vector (which can be obtained by models such as word2vec, Glove, ELMO, and bert), E t is the part-of-speech vector, E p is the word position vector, It is the concatenation operation between vectors.

[0147] Step E2, based on the dependency determination model, performing dependency determination on the vectorized sentence to obtain a dependency determination result;

[0148] Based on the dependency discrimination model, the vectorized sentence is subjected to dependency discrimination to obtain a dependency discrimination result, wherein the dependency discrimination model includes a first feature extraction model, a first fully connected network, a second fully connected network and a dual affine transformation network,

[0149] The step of performing dependency determination on the vectorized sentence based on the dependency determination model to obtain a dependency determination result includes the following steps H1 to H4:

[0150] Step H1, performing feature extraction on the vectorized sentence based on the first feature extraction model to obtain a first feature extraction result;

[0151] In this embodiment, it should be noted that the first feature extraction model is a neural network that extracts features from the vectorized sentence, and the first feature extraction model includes a Transformer model, etc.

[0152] Based on the first feature extraction model, feature extraction is performed on the vectorized sentence to obtain a first feature extraction result. Specifically, the vectorized sentence is input into the first feature extraction model, feature extraction is performed on the vectorized sentence to obtain a first feature extraction matrix, and the first feature extraction matrix is ​​used as the first feature extraction result.

[0153] Step H2, based on the first fully connected network and the second fully connected network, respectively fully connect the first feature extraction result to obtain a first sentence vector and a second sentence vector;

[0154] In this embodiment, based on the first fully connected network and the second fully connected network, the first feature extraction result is fully connected respectively to obtain a first sentence vector and a second sentence vector. Specifically, the first feature extraction matrix is ​​input into the first fully connected network, the first feature extraction matrix is ​​fully connected to obtain a first sentence vector, and the first feature extraction matrix is ​​input into the second fully connected network, the first feature extraction matrix is ​​fully connected to obtain a second sentence vector. It should be noted that the first sentence vector includes at least a prefix vector, which is used to represent the representation vector of the dependent word in the dependency relationship, and the second sentence vector includes at least a suffix vector, which is used to represent the representation vector of the dependent word in the dependency relationship. For example, assuming that word B depends on word A, the word representation vector corresponding to word A is the prefix vector, and the word representation vector corresponding to word B is the suffix vector.

[0155] Step H3, based on the double affine transformation network, performing double affine transformation on the first sentence vector and the second sentence vector to obtain a dependency score matrix;

[0156] In this embodiment, based on the first double affine transformation network, the first sentence vector and the second sentence vector are subjected to double affine transformation to obtain a dependency score matrix. Specifically, the first sentence vector and the second sentence vector are input into the first double affine transformation network, and the first sentence vector and the second sentence vector are subjected to double affine transformation to calculate the probability score that there is a dependency relationship between each prefix vector in the first sentence vector and each suffix vector in the second sentence vector, so as to obtain the dependency score matrix, wherein the dependency score matrix is ​​a score matrix composed of the probability scores that there is a dependency relationship between each prefix vector and each suffix vector.

[0157] Step H4, obtaining the target dependency syntactic analysis result based on the dependency score matrix.

[0158] In this embodiment, based on the dependency score matrix, the dependency judgment result is determined, specifically based on the preset maximum spanning tree algorithm, the maximum probability score and the maximum probability score that meet the preset score selection condition are selected in the dependency score matrix, and the dependency vector composed of the maximum probability score and the corresponding target probability scores and the corresponding vectorized words is used as the dependency judgment result, wherein the preset score selection condition includes a one-to-one correspondence between the to-be-processed words corresponding to each target probability score and the to-be-processed words in the to-be-processed sentence, etc. For example, assuming that each of the target probability scores is A and B, wherein the target probability score A represents the probability score of word b being attached to word a, the target probability score B represents the probability score of word c being attached to word b, and the vectorized word corresponding to word a is vector X, the vectorized word corresponding to word b is vector Y, and the vectorized word corresponding to word c is vector Z, and then the dependency vector is a vector (X, 1, 0, 0, 1, Y, 1, 0, 0, 1, Z), wherein (1, 0, 0, 1) indicates that there is a dependency relationship between words.

[0159] Step E3: Based on the dependency type prediction model and the dependency identification result, the dependency type of the vectorized sentence is predicted to obtain a target dependency syntactic analysis result.

[0160] Based on the dependency type prediction model and the dependency discrimination result, the dependency type of the vectorized sentence is predicted to obtain a target dependency syntactic analysis result.

[0161] For example, assuming that the sentence is ABC, where the dependency relationship type between word A and word B is a subject-predicate relationship type, and the dependency relationship type between word B and word C is a verb-object relationship type, then it can be determined that the sentence pattern of the sentence is a subject-predicate sentence pattern, the word component corresponding to word A is the subject, the word component corresponding to word B is the predicate, and the word component corresponding to word C is the object.

[0162] In this embodiment, the training sentence is vectorized to obtain a vectorized sentence; based on the dependency discrimination model, the vectorized sentence is subjected to dependency discrimination to obtain a dependency discrimination result; based on the dependency type prediction model and the dependency discrimination result, the vectorized sentence is subjected to dependency type prediction to obtain a target dependency syntactic analysis result. In this embodiment, the target dependency syntactic analysis result is accurately obtained.

[0163] Reference Figure 3 , Figure 3 It is a schematic diagram of the device structure of the hardware operating environment involved in the embodiment of the present application.

[0164] like Figure 3 As shown, the evaluation object extraction device may include: a processor 1001, such as a CPU, a memory 1005, and a communication bus 1002. The communication bus 1002 is used to realize the connection and communication between the processor 1001 and the memory 1005. The memory 1005 may be a high-speed RAM memory or a stable memory (non-volatile memory), such as a disk memory. The memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0165] Optionally, the evaluation object extraction device may also include a rectangular user interface, a network interface, a camera, an RF (Radio Frequency) circuit, a sensor, an audio circuit, a WiFi module, etc. The rectangular user interface may include a display screen (Display), an input submodule such as a keyboard (Keyboard), and the optional rectangular user interface may also include a standard wired interface and a wireless interface. The network interface may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0166] Those skilled in the art will understand that Figure 3 The evaluation object extraction device structure shown in the figure does not constitute a limitation on the evaluation object extraction device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0167] like Figure 3 As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module and an evaluation object extraction method program. The operating system is a program that manages and controls the hardware and software resources of the evaluation object extraction device, and supports the operation of the evaluation object extraction method program and other software and / or programs. The network communication module is used to realize the communication between the components inside the memory 1005, and the communication with other hardware and software in the evaluation object extraction method system.

[0168] exist Figure 3 In the evaluation object extraction device shown, the processor 1001 is used to execute the evaluation object extraction method program stored in the memory 1005 to implement the steps of any of the evaluation object extraction methods described above.

[0169] The specific implementation of the evaluation object extraction device of the present application is basically the same as the various embodiments of the above-mentioned evaluation object extraction method, and will not be repeated here.

[0170] The embodiment of the present application also provides an evaluation object extraction device, the evaluation object extraction device comprising:

[0171] A first acquisition module is used to acquire the evaluation data to be processed;

[0172] A determination module, used to determine the dependency relationship information between words, word information and part of speech information of words in the evaluation data to be processed;

[0173] An input module, used to input the dependency information, word part-of-speech information and word information into a preset extraction model;

[0174] The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence;

[0175] The object extraction module is used to perform object extraction processing on the dependency information, word part-of-speech information and word information of words based on the preset extraction model to obtain a target evaluation object.

[0176] Optionally, the device further comprises:

[0177] A second acquisition module is used to acquire a training sentence to obtain first training data marked with a first preset label;

[0178] A third acquisition module is used to perform a first iterative training on the first preset model to be trained based on the first training data marked with the first preset label to obtain a first target model;

[0179] A setting module is used to set the first target model as the preset extraction model.

[0180] Optionally, the third acquisition module includes:

[0181] A first acquisition unit, used to acquire a first preset model to be trained;

[0182] an object label analysis unit, configured to input the first training data into the first preset model to be trained, so as to perform object label analysis on the first training data and obtain a first predicted label;

[0183] a calculation unit, configured to calculate a first model error based on the first predicted label and the first preset label;

[0184] An updating unit is used to update the first preset model to be trained based on the first model error until the first preset model to be trained meets a first preset update end condition to obtain a first target model.

[0185] Optionally, the second acquisition module includes:

[0186] A second acquisition unit, used for acquiring a training sentence, and inputting the training sentence into a preset dependency syntax model;

[0187] A dependency syntactic analysis unit, configured to perform dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result;

[0188] A first determining unit, configured to determine a preset concatenated dependency vector of the training sentence based on the target dependency syntactic analysis result;

[0189] The second determining unit is used to determine the first training data based on the preset concatenated dependency vector and the training sentence marked with the first preset label.

[0190] Optionally, the second acquisition module further includes:

[0191] A third acquisition unit, used to acquire second training data annotated with a second preset label and a second preset model to be trained, wherein the second preset label is a dependency label;

[0192] A training unit, configured to perform a second iterative training on a second preset model to be trained based on the second training data annotated with a second preset label to obtain a second target model;

[0193] A setting unit is used to set the second target model as the preset dependency syntax model.

[0194] Optionally, the training unit includes:

[0195] A dependency label analysis subunit, configured to input the second training data into the second preset model to be trained, so as to perform dependency label analysis between words on the second training data to obtain a second predicted label;

[0196] A calculation subunit, configured to calculate a second model error based on the second predicted label and the second preset label;

[0197] An updating subunit is used to update the second preset model to be trained based on the second model error until the second preset model to be trained meets a second preset update end condition to obtain a second target model.

[0198] Optionally, the first determining unit includes:

[0199] An input subunit, used for inputting the evaluation data to be processed into a preset dependency syntax model;

[0200] The dependency syntax analysis subunit is used to perform dependency syntax analysis on the evaluation data to be processed based on the preset dependency syntax model to obtain dependency relationship information between words in the evaluation data to be processed.

[0201] Optionally, the preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model.

[0202] The dependency syntax analysis subunit is used to implement:

[0203] Vectorizing the evaluation data to be processed to obtain vectorized statements;

[0204] Based on the dependency determination model, performing dependency determination on the vectorized sentence to obtain a dependency determination result;

[0205] Based on the dependency type prediction model and the dependency discrimination result, the dependency type of the vectorized sentence is predicted to obtain dependency information between words in the evaluation data to be processed.

[0206] Optionally, the dependency discriminant model includes a first feature extraction model, a first fully connected network, a second fully connected network and a dual affine transformation network.

[0207] The dependency syntax analysis subunit is further used to implement:

[0208] Based on the first feature extraction model, perform feature extraction on the vectorized sentence to obtain a first feature extraction result;

[0209] Based on the first fully connected network and the second fully connected network, fully connect the first feature extraction result respectively to obtain a first sentence vector and a second sentence vector;

[0210] Based on the double affine transformation network, performing double affine transformation on the first sentence vector and the second sentence vector to obtain a dependency score matrix;

[0211] Based on the dependency score matrix, a dependency determination result is obtained.

[0212] Optionally, the training sentence includes words to be processed, and the vectorized sentence includes vectorized words.

[0213] The dependency syntax analysis subunit is used to implement:

[0214] Obtaining a word vector to be processed, a part-of-speech vector to be processed and a position vector to be processed corresponding to the word to be processed;

[0215] The vectorized word is generated based on the word vector to be processed, the part-of-speech vector to be processed and the word position vector to be processed.

[0216] Optionally, the determining module includes:

[0217] An input unit, used for inputting the evaluation data to be processed into a preset dependency syntax model;

[0218] The analyzing unit is used to perform dependency syntactic analysis on the evaluation data to be processed based on the preset dependency syntactic model to obtain dependency relationship information between words in the evaluation data to be processed.

[0219] The specific implementation of the evaluation object extraction device of the present application is basically the same as the various embodiments of the evaluation object extraction method described above, and will not be repeated here.

[0220] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned evaluation object extraction method when executed by a processor.

[0221] The specific implementation of the computer program product of the present application is basically the same as the various embodiments of the above-mentioned evaluation object extraction method, and will not be repeated here.

[0222] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent processing scope of the present application.

Claims

1. A method for extracting evaluation objects, characterized in that: The evaluation object extraction method comprises: Get the evaluation data to be processed; Determining dependency information between words, word information and word part-of-speech information in the evaluation data to be processed; Inputting the dependency relationship information and word part-of-speech information into a preset extraction model; The preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence; Obtaining a training sentence, and inputting the training sentence into a preset dependency syntax model; Based on the preset dependency syntactic model, the training sentence is subjected to dependency syntactic analysis between words to obtain a target dependency syntactic analysis result; Determining a preset concatenated dependency vector of the training sentence based on the target dependency syntactic analysis result; Determining first training data based on the preset concatenated dependency vector and the training sentence marked with the first preset label; Based on first training data marked with a first preset label, performing a first iterative training on a first preset model to be trained to obtain a first target model; Setting the first target model to the preset extraction model; Based on the preset extraction model, the dependency information, the word information and the word part-of-speech information are subjected to object extraction processing to obtain a target evaluation object; The preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model. The step of performing dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result includes: Vectorizing the training sentence to obtain a vectorized sentence; Based on the dependency determination model, performing dependency determination on the vectorized sentence to obtain a dependency determination result; Based on the dependency type prediction model and the dependency discrimination result, the dependency type of the vectorized sentence is predicted to obtain a target dependency syntactic analysis result.

2. The evaluation object extraction method according to claim 1, characterized in that: The step of performing a first iterative training on the first preset model to be trained based on the first training data marked with the first preset label to obtain the first target model includes: Acquire first training data marked with a first preset label and a first preset model to be trained; Inputting the first training data into the first preset model to be trained to perform object label analysis on the first training data to obtain a first predicted label; Calculating a first model error based on the first predicted label and the first preset label; Based on the first model error, the first preset model to be trained is updated until the first preset model to be trained meets a first preset update end condition to obtain a first target model.

3. The evaluation object extraction method according to claim 1, characterized in that: Before the step of performing dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result, the method includes: Acquire second training data and a second preset to-be-trained model annotated with a second preset label, wherein the second preset label is a dependency label; Based on the second training data annotated with the second preset label, performing a second iterative training on the second preset model to be trained to obtain a second target model; The second target model is set as the preset dependency syntax model.

4. The evaluation object extraction method according to claim 3, characterized in that: The step of performing a second iterative training on the second preset model to be trained based on the second training data annotated with the second preset label to obtain a second target model includes: Inputting the second training data into the second preset model to be trained to perform dependency relationship label analysis between words on the second training data to obtain a second predicted label; Calculating a second model error based on the second predicted label and the second preset label; Based on the second model error, the second preset model to be trained is updated until the second preset model to be trained meets a second preset update end condition to obtain a second target model.

5. The evaluation object extraction method according to claim 1, characterized in that: The dependency discriminant model includes a first feature extraction model, a first fully connected network, a second fully connected network and a dual affine transformation network. The step of performing dependency determination on the vectorized sentence based on the dependency determination model to obtain a dependency determination result includes: Based on the first feature extraction model, perform feature extraction on the vectorized sentence to obtain a first feature extraction result; Based on the first fully connected network and the second fully connected network, fully connect the first feature extraction result respectively to obtain a first sentence vector and a second sentence vector; Based on the double affine transformation network, performing double affine transformation on the first sentence vector and the second sentence vector to obtain a dependency score matrix; Based on the dependency score matrix, a dependency determination result is obtained.

6. The evaluation object extraction method according to claim 1, characterized in that: The vectorized sentence includes vectorized words, The step of vectorizing the evaluation data to be processed to obtain vectorized statements includes: Obtaining a word vector to be processed, a part-of-speech vector to be processed, and a word position vector to be processed corresponding to the evaluation data to be processed; The vectorized word is generated based on the word vector to be processed, the part-of-speech vector to be processed and the word position vector to be processed.

7. The evaluation object extraction method according to claim 1, characterized in that: The step of determining the dependency relationship information between words and the part-of-speech information of the words in the evaluation data to be processed includes: Inputting the evaluation data to be processed into a preset dependency syntax model; Based on the preset dependency syntax model, dependency syntax analysis is performed on the evaluation data to be processed to obtain dependency relationship information between words in the evaluation data to be processed and part-of-speech information of the words.

8. An evaluation object extraction device, characterized in that: The evaluation object extraction device comprises: A first acquisition module is used to acquire the evaluation data to be processed; A determination module, used to determine the dependency relationship information between words, word information and word part information in the evaluation data to be processed; An input module, used to input the dependency relationship information and word part-of-speech information into a preset extraction model; wherein the preset extraction model is obtained by performing a first iterative training on a first preset model to be trained based on first training data marked with a first preset label, wherein the first training data includes a training sentence and a preset concatenated dependency vector corresponding to the training sentence; A second acquisition unit, used for acquiring a training sentence, and inputting the training sentence into a preset dependency syntax model; A dependency syntactic analysis unit, configured to perform dependency syntactic analysis between words on the training sentence based on the preset dependency syntactic model to obtain a target dependency syntactic analysis result; A first determining unit, configured to determine a preset concatenated dependency vector of the training sentence based on the target dependency syntactic analysis result; A second determining unit, configured to determine first training data based on the preset concatenated dependency vector and the training sentence annotated with a first preset label; A third acquisition module is used to perform a first iterative training on a first preset model to be trained based on first training data marked with a first preset label to obtain a first target model; A setting module, used to set the first target model to the preset extraction model; An object extraction module, for performing object extraction processing on the dependency information, word information and word part-of-speech information based on the preset extraction model to obtain a target evaluation object; The preset dependency syntax model includes a dependency relationship discrimination model and a dependency relationship type prediction model. The dependency parsing subunit is used to implement: The training sentences are vectorized to obtain vectorized sentences; based on the dependency discrimination model, dependency discrimination is performed on the vectorized sentences to obtain dependency discrimination results; based on the dependency type prediction model and the dependency discrimination results, dependency type prediction is performed on the vectorized sentences to obtain target dependency syntactic analysis results.

9. An evaluation object extraction device, characterized in that: The evaluation object extraction device comprises: a memory, a processor, and a program stored in the memory for implementing the evaluation object extraction method. The memory is used to store a program for implementing the evaluation object extraction method; The processor is used to execute a program for implementing the evaluation object extraction method to implement the steps of the evaluation object extraction method as described in any one of claims 1 to 7.

10. A readable storage medium, characterized in that: The readable storage medium stores a program for implementing the evaluation object extraction method, and the program for implementing the evaluation object extraction method is executed by a processor to implement the steps of the evaluation object extraction method as claimed in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images