WO2019174423A1

WO2019174423A1 - Entity sentiment analysis method and related apparatus

Info

Publication number: WO2019174423A1
Application number: PCT/CN2019/073665
Authority: WO
Inventors: 王天祎
Original assignee: 北京国双科技有限公司
Priority date: 2018-03-16
Filing date: 2019-01-29
Publication date: 2019-09-19
Also published as: CN110287477A; CN110287477B

Abstract

Disclosed in the present invention are an entity sentiment analysis method and a related apparatus, the entity sentiment analysis method comprising: performing word segmentation processing on a text to be predicted to obtain a part of speech sequence of the text to be predicted and to obtain the vector of each participle in the part of speech sequence of the text to be predicted and the vector of a target entity; and carrying out prediction for the vector of the each participle in the part of speech sequence of the text to be predicted and for the vector of the target entity by means of an entity sentiment analysis model so as to be able to obtain a prediction result about the sentiment tendency of the target entity in the text to be predicted. In the described process, the text to be predicted is subject to word segmentation processing to obtain a the part of speech sequence thereof and to obtain the vector of each participle in the part of speech sequence and the vector of the target entity, instead of manual word selection and extraction of word features, which solves the problem of the accuracy of a sentiment tendency result being affected by manual word selection and provision of word features.

Description

Entity sentiment analysis method and related device

The present application claims priority to Chinese Patent Application No. 201810217282.9, entitled "Entity Affective Analysis Method and Related Device", filed on March 16, 2018, the entire contents of which is incorporated herein by reference. .

Technical field

The present invention relates to the field of text analysis technology, and in particular, to a physical sentiment analysis method and related apparatus.

Background technique

Text sentiment analysis is mainly to reflect the emotional orientation of users in certain events, people, companies, products, etc. in social media. Entity sentiment analysis refers to analyzing the sentimental tendency of certain entities in the text, rather than the tendency of the whole text. The advantage of this is to make the analysis of the emotional objects more granular.

Existing solutions generally rely on manual extraction of features for traditional machine learning classification algorithms. Specifically, the words around the target entity in the text are manually selected, and the feature of the word is extracted and input to the classifier, and the classifier performs sentiment analysis to obtain the sentimental tendency result of the text to the target entity.

Selecting words by hand and extracting the characteristics of words will make the feature extraction process have strong subjectivity and affect the accuracy of emotional tendency results.

Summary of the invention

In view of the above problems, the present invention has been made in order to provide an entity sentiment analysis method and related apparatus that overcome the above problems or at least partially solve the above problems.

A method of entity sentiment analysis, including:

Get the text to be predicted;

Performing word segmentation processing on the text to be predicted to obtain a part-of-speech sequence of the text to be predicted;

Obtaining a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity;

Using a physical sentiment prediction model to predict a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the sentiment orientation of the target entity in the text to be predicted; wherein: The entity sentiment prediction model is constructed based on the first principle; the first principle includes: iteratively updating the parameters in the neural network algorithm until the prediction using the neural network algorithm that updates the parameters to predict the feature vector of the training text The result is equivalent to the manual annotation result; the feature vector of the training text is obtained according to the vector of the part of speech sequence of the training text and the vector of the target entity in the part of speech sequence of the training text.

Optionally, the obtaining a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity, including:

Obtaining a word vector of each of the participles in the part of speech sequence of the text to be predicted;

Multiplying a word vector of each word segment in the part-of-speech sequence of the text to be predicted and an attenuation factor to obtain a vector of each word segment in the part-of-speech sequence of the text to be predicted;

A vector of the word segment corresponding to the target entity in the text to be predicted is used as a vector of a target entity in the part-of-speech sequence of the text to be predicted.

Optionally, it also includes:

Obtaining any one or combination of a part of speech vector, a word packet vector, and a vector of a relative target entity distance of each participle of the part of speech sequence of the text to be predicted;

Combining the word vector of each participle in the part-of-speech sequence of the text to be predicted, and the part-of-speech vector of each part of the obtained part-of-speech sequence of the text to be predicted, the word packet vector, and the vector of the relative target entity distance One or a combination, obtaining an initial vector of each of the participles in the part of speech sequence of the text to be predicted;

The multiplying the word vector and the attenuation factor of each participle of the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted includes:

Multiplying an initial vector and an attenuation factor of each participle in the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted.

Optionally, if a part of the to-be-predicted text corresponding to the target entity includes a plurality, the average of the vectors of the plurality of word segments corresponding to the target entity in the text to be predicted is used as the text to be predicted. The vector of the target entity in the part of speech sequence.

Optionally, the entity emotion prediction model is used to predict a vector of each word segment in the part-of-speech sequence of the text to be predicted and a vector of the target entity, and obtain a prediction of the sentiment orientation of the target entity in the text to be predicted. The results include:

Performing weighted averaging processing on the vector of each participle in the part of speech sequence of the text to be predicted, and obtaining a vector weighted by the part of speech sequence of the text to be predicted;

Multiplying a vector of the target entity in the part of speech sequence of the text to be predicted by a first matrix to obtain a derived vector of the target entity;

And obtaining a feature vector according to the vector weighted by the part of speech sequence of the text to be predicted, and/or the derived vector of the target entity in the part of speech sequence of the text to be predicted;

The feature vector is processed by using a softmax function to obtain a probability output vector, wherein the probability output vector includes: a probability value of the target entity in the text to be predicted, respectively, under the sentiment orientation of the preset category.

Optionally,

The construction process of the entity sentiment prediction model includes:

Performing word segmentation processing on the training text to obtain a part-of-speech sequence of the training text;

Obtaining a vector of each participle in the part of speech sequence of the training text and a vector of the target entity;

Performing weighted averaging processing on the vector of each participle in the part of speech of the training text to obtain a vector weighted by the part of speech of the training text;

Multiplying a vector of the target entity in the part-of-speech sequence of the training text with a first matrix to obtain a derived vector of the target entity in the part-of-speech sequence of the training text;

Obtaining a feature vector according to the weighted vector of the part of speech of the training text, and/or the derived vector of the target entity in the part of speech sequence of the training text;

The feature vector is processed by using a softmax function to obtain a probability output vector, wherein the probability output vector includes: a probability value of the target entity in the training text under the sentiment orientation of the preset category;

Performing a cross-entropy operation on the probability output vector and the artificial annotation category of the training text to obtain a loss function;

Optimizing the loss function, and updating the first parameter according to the optimized loss function until the probability output vector obtained by predicting the training text by using the updated feature vector obtained by the updated first parameter and the training text The manual labeling category is equivalent; wherein the first parameter comprises a vector of each of the first matrix, the softmax function, and the part of speech of the training text;

And using the updated second parameter as an entity sentiment prediction model; wherein the second parameter comprises: the first matrix and the softmax function.

A physical emotion analysis device includes:

An obtaining unit, configured to obtain a text to be predicted;

a word segmentation unit, configured to perform word segmentation processing on the text to be predicted, to obtain a part-of-speech sequence of the text to be predicted;

a generating unit, configured to obtain a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity;

a prediction unit, configured to predict, by using an entity sentiment prediction model, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the sentiment orientation of the target entity in the text to be predicted Wherein the entity sentiment prediction model is constructed based on a first principle; the first principle comprises: iteratively updating parameters in the neural network algorithm until the feature vector of the training text is performed by using a neural network algorithm after updating the parameters The prediction result obtained by the prediction is equivalent to the manual annotation result; the feature vector of the training text is obtained according to the vector of the part-of-speech sequence of the training text and the vector of the target entity in the part-of-speech sequence of the training text.

Optionally, the generating unit includes:

a first obtaining unit, configured to respectively obtain a word vector of each participle in the part of speech sequence of the text to be predicted;

a second obtaining unit, configured to multiply a word vector and an attenuation factor of each participle of the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted;

And generating a subunit, configured to use a vector of the word segment corresponding to the target entity in the text to be predicted as a vector of the target entity in the part of speech sequence of the text to be predicted.

A storage medium, the storage medium comprising a stored program, wherein the device in which the storage medium is located is controlled to perform the entity sentiment analysis method according to any one of the above.

A processor for running a program, wherein the program runtime executes the entity sentiment analysis method of any of the above.

According to the above technical solution, in the entity sentiment analysis method and the related device provided by the present invention, the word-prepared text is subjected to word segmentation processing to obtain the part-of-speech sequence of the text to be predicted, and then each of the part of speech sequence of the text to be predicted is obtained. a vector of the word segmentation and a vector of the target entity, and the vector of the target word in the part of speech of the text to be predicted and the vector of the target entity are predicted by the entity sentiment prediction model, and the target entity in the text to be predicted is obtained. The predictive outcome of sentiment orientation. In the above process, the text to be predicted is subjected to word segmentation to obtain a part-of-speech sequence, and the vector of each participle in the part-of-speech sequence and the vector of the target entity are obtained, and the word selection is not selected by the manual selection and the word feature is solved, and the manual selection is solved. Words and questions that provide the characteristics of the words that affect the accuracy of the emotionally biased results.

The above description is only an overview of the technical solutions of the present invention, and the above-described and other objects, features and advantages of the present invention can be more clearly understood. Specific embodiments of the invention are set forth below.

DRAWINGS

Various other advantages and benefits will become apparent to those skilled in the art from a The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting. Throughout the drawings, the same reference numerals are used to refer to the same parts. In the drawing:

1 is a flow chart showing a process of constructing a physical sentiment prediction model disclosed in an embodiment of the present invention;

FIG. 2 is a flowchart showing a specific implementation manner of step S102 disclosed in the embodiment of the present invention;

FIG. 3 is a flowchart of a method for analyzing entity sentiment according to an embodiment of the present invention;

FIG. 4 is a flowchart showing a specific implementation manner of step S303 disclosed in the embodiment of the present invention;

FIG. 5 is a flowchart of a specific implementation manner of step S304 disclosed in the embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a physical sentiment analysis apparatus according to an embodiment of the present invention;

FIG. 7 is a schematic structural diagram of a generating unit disclosed in an embodiment of the present invention;

FIG. 8 is a schematic structural diagram of a prediction unit disclosed in an embodiment of the present invention.

detailed description

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While the embodiments of the present invention have been shown in the drawings, the embodiments Instead, these embodiments are provided to enable

The disclosure is to be thoroughly understood and the scope of the disclosure may be fully conveyed to those skilled in the art.

In the embodiment of the present application, the entity sentiment prediction model needs to be used to predict the predicted text. Therefore, before performing the entity sentiment analysis method disclosed in the embodiment of the present application, the entity sentiment prediction model needs to be constructed first.

Referring to FIG. 1, the process of constructing the entity sentiment prediction model includes:

S101. Perform word segmentation on the training text to obtain a part of speech sequence of the training text.

Wherein, a training document is prepared, and the training document includes at least one training text. The training text is a user's evaluation statement about certain events, people, businesses, products, and so on.

For each training text in the training document, open source tools, such as LTP (Language Technology Platform), are used to segment the words and obtain the part-of-speech sequence of the corresponding participle, wherein the part-of-speech sequence includes the word segmentation sequence and the part-of-speech result. The word segmentation sequence includes each segmentation word obtained after segmentation of the training text; the part-of-speech result includes the part of speech of each segmentation word. For example: the training text is: the car front face design is mighty and domineering. After the word segmentation of the training text, the obtained word segmentation sequence is [car, front face, design, mighty, domineering], the part of speech result is [n, n, v, a, n]; n represents general noun, noun; v Represents verb, verb; a stands for adjective, adjective.

S102. Obtain a vector of each participle in the part of speech sequence of the training text and a vector of the target entity.

Wherein, each participle in the part of speech sequence of the training text needs to be expressed by using a feature vector. Therefore, it is necessary to obtain a vector of the word segmentation for each participle in the part of speech sequence of the training text. The training text includes a target entity, and the part-of-speech sequence after the word segmentation processing of the training text also includes a word segmentation corresponding to the target entity. Therefore, the vector of the word segmentation corresponding to the target entity in the part of speech sequence of the training text is a vector of the target entity.

Optionally, in an implementation manner of step S102, referring to FIG. 2, the step includes:

S1021. Obtain a word vector of each participle in the part of speech sequence of the training text, respectively.

Wherein, for each participle in the part of speech sequence of the training text, the word vector model is separately screened, and the word vector of the current participle in the word vector model is obtained.

The open source tool software is used to segment each text sentence in the text library, and the word vector model is used for word vector training, that is, the word vector model is generated. The text library includes an industry corpus and a general corpus, which refers to a text library that is separated from the industry. The role of the word vector model is to map words to a certain latitude space, which can represent the similarity between words and words. At the same time, the word vector model contains the low-frequency long tail words appearing in the corpus (low-frequency long-tail words refer to words whose frequency is lower than a certain threshold in all vocabulary), and are collectively recorded as UNK (unknown keyword). UNK has a unique word vector in the word vector model.

If a certain part of the part of the training text has no corresponding word vector in the word vector model, then the word vector of the participle uses the UNK word vector.

It should also be noted that each participle in the part-of-speech sequence of the training text, which differs in part of speech, may also result in a different emotional orientation of the target entity. Therefore, the part of speech vector of each part of the part of speech of the training text can also be obtained.

Specifically, a random vector of a certain dimension is used for the part of speech. For example, if there are five kinds of words [a, b, c, d, e], then the random vector Va can be used to represent a. Similarly, the random vector Vb is used to represent b, Va. The dimension of Vb can be arbitrarily specified. For each participle in the part of speech sequence of the training text, a corresponding part of speech vector can be obtained according to its part of speech.

Similarly, the word package to which the word segment belongs also affects the judgment of the sentiment orientation of the target entity. In particular, a part of the participle of the training text does not find the corresponding word vector in the word vector model. The word package vector can be a comprehensive reflection of the word segmentation. Thus, a packet vector for each of the participles in the part of speech sequence of the training text can also be obtained.

Specifically, each word segmentation of the part of speech of the training text and the belonging relationship of the industry domain word package are encoded to obtain a word packet vector of each word segment in the part of speech sequence of the training text. For example, it is determined whether each participle in the part of speech sequence of the training text is in the entity word package and is in the evaluation word package. The result of the judgment is encoded to obtain a packet vector of each part of the part of speech of the training text.

The distance of each participle in the part-of-speech sequence of the training text relative to the target entity may have a different influence on the sentiment orientation of the target entity. In general, the distance between the participle and the target entity is far, and the influence of the sentiment orientation of the target entity is less. Therefore, it is also necessary to obtain a vector of the distance of each word segment relative to the target entity in the part of speech sequence of the training text.

For each participle in the part-of-speech sequence of the training text, the distance is compared with the distance of the target entity to obtain a vector of the relative target entity distance of each participle. For example: [car, front face, design, mighty, domineering], the target entity is the front face design, then the distance between each participle from the target entity is [-2,-1,0,0,1,2,3], The distance sequence is encoded, and -2, -1, 0, 1, 2, and 3 are respectively encoded into vectors of a certain dimension, and a vector of relative target entity distances of each participle is obtained.

If any one or any combination of the part of speech vector, the word packet vector, and the vector of the relative target entity distance of each part of the part of the training text is obtained, it is also necessary to combine the part of speech of the training text. The word vector, the part of speech vector, the word packet vector, and the vector of the relative target entity distance of each of the participles in the middle get the initial vector of each participle in the part of speech sequence of the training text.

Wherein, for each participle in the part of speech sequence of the training text, the word vector, the part of speech vector, the word packet vector, and the vector of the relative target entity distance are respectively spliced and combined to form an initial vector of the participle.

S1022: Multiply a word vector of each part of the part of the training text and an attenuation factor to obtain a vector of each participle in the part of speech sequence of the training text.

Wherein, the attenuation factor of each participle is calculated according to the word vector of the relative target entity distance of each participle in the part of speech sequence of the training text. Specifically, the calculation formula of the attenuation factor e is e=1-d/N, where d represents the absolute distance of the current word segment from the target entity, and N is the length of the part-of-speech sequence of the training document.

The vector of each word segmentation of the part of the training text and the corresponding attenuation factor are multiplied to obtain a vector of the word segmentation.

It should also be noted that the length of the word segmentation of each training text in the training document is counted, and whether the long outlier length text exists in the training document is determined. Specifically, the standard deviation of the mean of the segmentation length of the training text is calculated, and the extra-long outlier length text is the training text other than the multiple of the standard deviation of the mean value. The specific multiple requirements can be set according to the actual situation.

If it is determined that there is no extra long out-of-group text in the training document, the length of the segmentation of the training text having the longest segmentation length in the training document is used as the length of the part-of-speech sequence of the training document. If it is determined that there is an extra long out-of-group length text in the training document, the length of the segmentation length of the training text having the longest segmentation length is included in the training text remaining in the training document except for the extra-long out-group length text. The length of the part of speech sequence of the training document. And, the extra long outlier length text in the training document is intercepted according to the length of the part of speech sequence of the training document. Specifically, it is centered on the target entity in the training text, and is extended forward and backward respectively until the length of the word segmentation reaches the length of the part of speech sequence of the training document.

For example, there are 10 training texts in the training document, and the length of the word segmentation of each training text is not equal, but the length of the segmentation of the longest one training text is 50, then 50 is the length of the part-of-speech sequence of the training document. If there is a training text in the training document with a length of 1000, then the training text is an extra long out-of-group text.

It should be noted that, if an initial vector of each participle in the part-of-speech sequence of the training text is obtained, a vector of each participle in the part-of-speech sequence of the training text is obtained: the part of the training text The initial vector of each participle in the multiplication is multiplied by the attenuation factor.

S1023: A vector of the participle of the training text corresponding to the target entity is used as a vector of the target entity.

It should be noted that, if the word segmentation of the corresponding target entity of the training text includes a plurality, the average of the vectors of the plurality of word segments corresponding to the target entity in the training text is used as a vector of the target entity.

S103. Perform weighted averaging processing on a vector of each participle in the part of speech sequence of the training text, and obtain a vector weighted by the part of speech of the training text.

Wherein, the vector of the target entity is combined on a vector of each participle in the part of speech sequence of the training text to obtain a vector of each participle in the part of speech sequence of the training text. The vector of each participle in the part of speech of the training text is calculated using the attention layer of HAN (Hierarchical Attention Networks), and the weight of each participle is obtained. Specifically, if the participle is farther away from the target entity, the emotional impact on the target entity is not large, and the weight does not need to be paid too much, and the weight is weakened. Otherwise, the weight is strengthened.

And performing weighted averaging processing on the vector of each participle in the part of speech sequence of the training text according to the weight of each participle in the part of speech sequence of the training text, to obtain a vector weighted by the part of speech of the training text.

S104. Multiply a vector of the target entity in the part-of-speech sequence of the training text with the first matrix to obtain a derived vector of the target entity.

The vector of the target entity in the part of speech sequence of the training text is multiplied by the first matrix to obtain a derived vector of the target entity.

It should also be noted that the first matrix is a matrix of m×m, and m is the dimension of the vector of the target entity in the part-of-speech sequence of the training text. The specific value of the first matrix is a randomly initialized value, and each value can select a fraction that is uniformly distributed in the range of -0.1 to 0.1.

S105. Obtain a feature vector according to a vector weighted by the part of speech of the training text, and/or a derived vector of the target entity.

The weighted vector of the training text may be used as the feature vector, or the derived vector of the target entity may be used as the feature vector, and the part of the training text may be weighted. Based on the latter vector, the derived vector of the target entity is added or subtracted to obtain the feature vector.

Specifically, if the participle corresponding to the target entity in the training text itself has sentiment orientation, the derived vector of the target entity may be selected as the feature vector. In addition, based on the weighted vector of the part of speech of the training text, adding or subtracting the derived vector of the target entity, the obtained feature vector may cause the derived vector of the target entity to act on the The weighted vector of the part of speech of the training text.

Optionally, in another embodiment of the present application, steps S103-S104 may be repeatedly performed several times, wherein the number of times to be repeatedly executed may be set according to actual requirements.

Specifically, the feature vector obtained last time in step S104 is used as the vector of the target entity at the next execution of steps S103 and S104, and the latest weighted vector of the part of speech of the training text and the latest of the target entity are obtained. a derived vector, and then a new feature vector is obtained according to the latest weighted vector of the part of speech of the training text and/or the latest derived vector of the target entity.

S106. Process the feature vector by using a softmax function to obtain a probability output vector.

Wherein, the probability output vector includes probability values of three categories, including positive, medium and negative. It is indicated that the training text is positive for the target entity; negative indicates that the training text is negative for the target entity; the middle indicates that the training text is neutral for the target entity. The probability value of each category is used to indicate the probability that the entity emotion of the target entity of the training text belongs to the corresponding category.

S107. Perform cross-entropy operation on the probability output vector and the manual labeling category of the training text to obtain a loss function.

Wherein, for each training text in the training document, the emotion of the training text to the target entity is manually recognized, and the positive, middle and negative emotions of the three categories are marked according to the emotion, and the manual labeling category of the training text is obtained. For example: the training text of “the front face design of the car is mighty and domineering”, the target entity is “front face design”, and the emotion is positive. Therefore, the identifier of the manual label category of the training text can be [1, 0, 0].

And performing the cross entropy operation on the probability output output vector and the artificial annotation category of the training text, and the obtained loss function is used to indicate a difference between the probability output vector and the manual annotation category of the training text.

S108. Optimize the loss function, and update the parameter according to the optimized loss function until the probability output vector obtained by predicting the training text by using the updated feature vector obtained by the updated first parameter and the training text. Manually labeled categories are equivalent.

The first parameter includes a vector of each of the attention layer, the first matrix, the softmax function, and the part of speech of the training text. Specifically, the manner of obtaining the vector of each participle in the part-of-speech sequence of the training text can be referred to in the embodiment corresponding to FIG. 1 , and the content of step S102 is not described herein.

The loss function can be optimized by a stochastic gradient descent method or an Adam optimization algorithm, etc., and an optimized loss function is obtained, and the updated parameter is recursively layer by layer according to the optimized loss function.

It should also be noted that, in this step, the equivalent meaning is that, from the perspective of those skilled in the art, the probability output vector can be regarded as equivalent to the manual labeling category of the training text, and may include not exactly the same.

S109. The updated second parameter is used as an entity sentiment prediction model. The second parameter includes: the attention layer, the first matrix, and the softmax function.

Based on the entity sentiment prediction model constructed by the method of the above embodiment, entity sentiment analysis can be performed on the predicted text. Specifically, referring to FIG. 3, the entity sentiment analysis method includes:

S301. Acquire a text to be predicted.

The text to be predicted is a user's evaluation statement about certain events, people, businesses, products, and the like. The text to be predicted is obtained to analyze the sentiment orientation of the text with respect to the target entity in the text.

S302. Perform word segmentation on the to-be-predicted text to obtain a part-of-speech sequence of the text to be predicted.

For the text to be predicted, the open source tool software is also used for word segmentation, and the part-of-speech sequence of the corresponding word segmentation is obtained. For the specific implementation process of this step, refer to the content of step S101 in the embodiment corresponding to FIG. 1, and details are not described herein again.

S303. Obtain a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity.

Optionally, in an implementation manner of step S303, referring to FIG. 4, the step includes:

S3031: Obtain a word vector of each participle in the part of speech sequence of the text to be predicted, respectively.

Optionally, in addition to obtaining a word vector of each participle in the part of speech sequence of the text to be predicted, a part of speech vector, a word packet vector, and a relative of each part of the part of speech sequence of the text to be predicted may be obtained. The vector of the target entity distance.

For the manner in which the word vector, the part of speech vector, the word packet vector, and the vector of the target entity distance are obtained, refer to the content of step S1021 in the embodiment corresponding to FIG.

S3032: Multiply a word vector of each part of the part of speech of the text to be predicted and an attenuation factor to obtain a vector of each part of the part of speech of the text to be predicted.

For the specific implementation of this step, refer to the content of step S1022 in the embodiment corresponding to FIG. 1, and details are not described herein again.

S3033. A vector of the word segmentation corresponding to the target entity in the text to be predicted is used as a vector of the target entity.

Optionally, in another embodiment of the present application, if the word segmentation corresponding to the target entity in the text to be predicted includes multiple, the vector of the plurality of word segments corresponding to the target entity in the text to be predicted is The average value is used as the vector of the target entity.

S304: predict, by using an entity sentiment prediction model, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the target entity of the text to be predicted; wherein the entity emotion The prediction model is constructed based on the first principle; the first principle includes: iteratively updating the parameters in the neural network algorithm until the prediction result obtained by predicting the feature vector of the training text by using the neural network algorithm after updating the parameter is equivalent The result is manually labeled; the feature vector of the training text is obtained according to a vector of the part of speech of the training text and a vector of the target entity in the part of speech of the training text.

In the entity sentiment analysis method disclosed in this embodiment, after performing word segmentation processing on the predicted text to obtain a part-of-speech sequence of the text to be predicted, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity are obtained. And predicting, by the entity sentiment prediction model, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the sentiment orientation of the target entity in the text to be predicted. In the above process, the text to be predicted is subjected to word segmentation to obtain a part-of-speech sequence, and the vector of each participle in the part-of-speech sequence and the vector of the target entity are obtained, and the word selection is not selected by the manual selection and the word feature is solved, and the manual selection is solved. Words and questions that provide the characteristics of the words that affect the accuracy of the emotionally biased results.

Optionally, in another embodiment of the present application, referring to FIG. 5, step S304 includes:

S3041: Perform weighted average processing on a vector of each participle in the part of speech sequence of the text to be predicted, and obtain a vector weighted by the part of speech sequence of the text to be predicted.

For the specific implementation of this step, refer to the content of step S103 in the embodiment corresponding to FIG. 1 , and details are not described herein again.

S3042: Multiply a vector of the target entity in the part-of-speech sequence of the text to be predicted by a first matrix to obtain a derived vector of the target entity.

The first matrix is a first matrix corresponding to the entity sentiment prediction model in step S109 in the embodiment of FIG. 1. For the specific implementation of this step, refer to the content of step S104 in the embodiment corresponding to FIG. 1 , and details are not described herein again.

S3043. Obtain a feature vector according to the vector weighted by the part-of-speech sequence of the text to be predicted and/or the derived vector of the target entity in the part-of-speech sequence of the text to be predicted.

For the specific implementation of this step, refer to the content of step S105 in the embodiment corresponding to FIG. 1 , and details are not described herein again.

S3044: Processing the feature vector by using a softmax function to obtain a probability output vector.

The softmax function is a softmax function corresponding to the entity sentiment prediction model in step S109 in the embodiment of FIG. 1. For the specific implementation of this step, refer to the content of step S106 in the embodiment corresponding to FIG. 1 , and details are not described herein again.

Another embodiment of the present application further discloses a physical sentiment analysis apparatus, and the specific working process of each unit included in the application can be referred to the content corresponding to the embodiment of FIG. 3. Specifically, referring to FIG. 6, the entity sentiment analysis apparatus includes:

The obtaining unit 601 is configured to obtain the text to be predicted.

The word segmentation unit 602 is configured to perform word segmentation processing on the to-be-predicted text to obtain a part-of-speech sequence of the to-be-predicted text.

The generating unit 603 is configured to obtain a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity.

Optionally, in another embodiment of the present application, the generating unit 603, referring to FIG. 7, includes:

The first obtaining unit 6031 is configured to respectively obtain a word vector of each participle of the part of speech sequence of the text to be predicted.

The second obtaining unit 6032 is configured to multiply the word vector of each part of the part of speech of the text to be predicted and the attenuation factor to obtain a vector of each part of the part of speech of the text to be predicted.

The generating sub-unit 6033 is configured to use a vector of the word segmentation corresponding to the target entity in the text to be predicted as a vector of the target entity in the part-of-speech sequence of the text to be predicted.

For the specific working process of each unit in the generating unit 603 disclosed in this embodiment, refer to the content of the embodiment corresponding to FIG. 4, and details are not described herein again.

Optionally, in another embodiment of the present application, if the participle of the to-be-predicted text corresponding to the target entity includes multiple, the generating sub-unit 6033 executes the corresponding to the target entity in the text to be predicted. The vector of the word segmentation is used as the vector of the target entity in the part of speech sequence of the text to be predicted, and is specifically used to: use an average value of a vector of the plurality of word segments corresponding to the target entity in the text to be predicted as the A vector that predicts the target entity in the part-of-speech sequence of the text.

Optionally, in another embodiment of the present application, the entity sentiment analysis apparatus further includes:

And a third obtaining unit, configured to obtain any one or combination of a part of speech vector, a word packet vector, and a vector of a relative target entity distance of each part of the part of speech of the text to be predicted.

a combination unit, a word vector for combining each of the participles in the part of speech sequence of the text to be predicted, and a part of speech vector, a word package vector, and a relative target of each of the participle sequences of the obtained text to be predicted Any one or combination of the vectors of the entity distances obtains an initial vector of each of the word segments in the part of speech sequence of the text to be predicted.

The second obtaining unit 6032 performs multiplication of the word vector and the attenuation factor of each participle of the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted. Specifically, the method uses: multiplying an initial vector and an attenuation factor of each participle in the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted.

a prediction unit 604, configured to predict, by using an entity sentiment prediction model, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction of the sentiment orientation of the target entity in the text to be predicted a result; wherein the entity sentiment prediction model is constructed based on a first principle; the first principle comprises: iteratively updating parameters in the neural network algorithm until a feature vector of the training text is obtained by using a neural network algorithm after updating the parameter The prediction result obtained by the prediction is equivalent to the manual annotation result; the feature vector of the training text is obtained according to the vector of the part-of-speech sequence of the training text and the vector of the target entity in the part-of-speech sequence of the training text.

Optionally, in another embodiment of the present application, the prediction unit 604, as shown in FIG. 8, includes:

The first calculating unit 6041 is configured to perform weighted averaging processing on the vector of each participle in the part of speech sequence of the text to be predicted, to obtain a vector weighted by the part of speech sequence of the text to be predicted.

The second calculating unit 6042 is configured to multiply a vector of the target entity in the part-of-speech sequence of the text to be predicted by the first matrix to obtain a derived vector of the target entity.

The third calculating unit 6043 is configured to obtain a feature vector according to the vector weighted by the part of speech sequence of the text to be predicted, and/or the derived vector of the target entity in the part of speech sequence of the text to be predicted.

The fourth calculating unit 6044 is configured to process the feature vector by using a softmax function to obtain a probability output vector.

For the specific working process of each unit in the prediction unit 604 disclosed in this embodiment, refer to the content of the embodiment corresponding to FIG. 5, and details are not described herein again.

In this embodiment, the text to be predicted is subjected to word segmentation by the word segmentation unit to obtain a part of speech sequence, and the vector of each word segment in the part of speech sequence and the vector of the target entity are obtained by the generating unit, instead of manually selecting words and extracting word features. The problem of affecting the accuracy of affective tendencies results due to manual word selection and providing word features is solved.

Optionally, in another embodiment of the present application, the entity sentiment analysis apparatus may further predict the training text to obtain an entity sentiment prediction model.

Specifically, the word segmentation unit 602 is further configured to perform word segmentation processing on the training text to obtain a part-of-speech sequence of the training text.

The generating unit 603 is further configured to obtain a vector of each participle in the part of speech sequence of the training text and a vector of the target entity.

The first calculating unit 6041 is further configured to perform weighted averaging processing on the vector of each participle in the part of speech of the training text, to obtain a vector with the partiality of the part of the training text.

The second calculating unit 6042 is further configured to multiply a vector of the target entity in the part-of-speech sequence of the training text by the first matrix to obtain a derived vector of the target entity in the part-of-speech sequence of the training text.

The third calculating unit 6043 is further configured to obtain a feature vector according to the weighted vector of the part of speech of the training text, and/or the derived vector of the target entity in the part of speech sequence of the training text.

The fourth calculating unit 6044 is further configured to process the feature vector by using a softmax function to obtain a probability output vector.

Moreover, the entity sentiment analysis apparatus further includes: an operation unit configured to perform an entropy operation on the probability output output vector and the manual annotation category of the training text to obtain a loss function.

An optimization unit for optimizing the loss function.

And an updating unit, configured to update the first parameter according to the optimized loss function, until the fourth computing unit 6044 predicts the training text by using the updated feature vector obtained by the updated first parameter, and the The manual annotation category of the training text is substantially equivalent; wherein the first parameter comprises a vector of each of the first matrix, the softmax function, and the part of speech of the training text.

a building unit, configured to use the updated second parameter as an entity sentiment prediction model; wherein the second parameter comprises: the first matrix and the softmax function.

For the specific working process of each unit in the foregoing embodiment, refer to the content of the embodiment corresponding to FIG. 1 , and details are not described herein again.

The entity sentiment analysis apparatus includes a processor and a memory, and the above-mentioned acquisition unit, word segmentation unit, generation unit, prediction unit, and the like are all stored as a program unit in a memory, and the processor executes the above-mentioned program unit stored in the memory to implement corresponding The function.

The processor contains a kernel, and the kernel removes the corresponding program unit from the memory. The kernel may set one or more, and adjust the kernel parameters to implement the sentiment analysis process of the text to be predicted to obtain the prediction result of the sentiment orientation of the target entity in the text to be predicted.

The memory may include non-persistent memory, random access memory (RAM), and/or non-volatile memory in a computer readable medium, such as read only memory (ROM) or flash memory (flash RAM), the memory including at least one Memory chip.

Embodiments of the present invention provide a storage medium on which a program is stored, which is implemented by a processor to implement the entity sentiment analysis method.

An embodiment of the present invention provides a processor, where the processor is configured to run a program, where the program executes the method of entity sentiment analysis.

The embodiment of the invention provides a device, which may be a server, a PC, a PAD, a mobile phone or the like. The device includes a processor, a memory, and a program stored on the memory and executable on the processor, and the processor performs the following steps when executing the program:

A method of entity sentiment analysis, including:

Get the text to be predicted;

Optionally, the entity sentiment analysis method further includes:

Optionally, the process of constructing the entity sentiment prediction model includes:

The invention also provides a computer program product, when executed on a data processing device, adapted to perform a process of initializing the method steps as follows:

A method of entity sentiment analysis, including:

Get the text to be predicted;

Multiplying a word vector and an attenuation factor of each participle in the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted;

Optionally, the entity sentiment analysis method further includes:

The feature vector is processed by using a softmax function to obtain a probability output vector, wherein the probability output vector includes: a probability value of the target entity in the text to be predicted, respectively, under a sentiment orientation of a preset species.

Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system, or computer program product. Thus, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment in combination of software and hardware. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (system), and computer program products according to embodiments of the present application. It will be understood that each flow and/or block of the flowchart illustrations and/or FIG. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing device to produce a machine for the execution of instructions for execution by a processor of a computer or other programmable data processing device. Means for implementing the functions specified in one or more of the flow or in a block or blocks of the flow chart.

The computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device. The apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.

These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on the computer or other programmable device to produce the computer

The implemented processing, such as instructions executed on a computer or other programmable device, provides steps for implementing the functions specified in one or more blocks of the flowchart or in a block or blocks of the flowchart.

In a typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include non-persistent memory, random access memory (RAM), and/or non-volatile memory in a computer readable medium, such as read only memory (ROM) or flash memory. Memory is an example of a computer readable medium.

Computer readable media includes both permanent and non-persistent, removable and non-removable media. Information storage can be implemented by any method or technology. The information can be computer readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory. (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disk read only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, Magnetic tape cartridges, magnetic tape storage or other magnetic storage devices or any other non-transportable media can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary storage of computer readable media, such as modulated data signals and carrier waves.

It is also to be understood that the terms "comprises" or "comprising" or "comprising" or any other variations are intended to encompass a non-exclusive inclusion, such that a process, method, article, Other elements not explicitly listed, or elements that are inherent to such a process, method, commodity, or equipment. An element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in a process, method, article, or device that comprises the element, without further limitation.

Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system, or computer program product. Thus, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment in combination of software and hardware. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.

The above is only an embodiment of the present application and is not intended to limit the application. Various changes and modifications can be made to the present application by those skilled in the art. Any modifications, equivalents, improvements, etc. made within the spirit and scope of the present application are intended to be included within the scope of the appended claims.

Claims

A method of entity sentiment analysis, characterized in that it comprises:

Get the text to be predicted;

Performing word segmentation processing on the text to be predicted to obtain a part-of-speech sequence of the text to be predicted;

Obtaining a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity;

Using a physical sentiment prediction model to predict a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the sentiment orientation of the target entity in the text to be predicted; wherein: The entity sentiment prediction model is constructed based on the first principle; the first principle includes: iteratively updating the parameters in the neural network algorithm until the prediction using the neural network algorithm that updates the parameters to predict the feature vector of the training text The result is equivalent to the manual annotation result; the feature vector of the training text is obtained according to the vector of the part of speech sequence of the training text and the vector of the target entity in the part of speech sequence of the training text.
The method according to claim 1, wherein the obtaining a vector of each participle of the part of speech sequence of the text to be predicted and a vector of the target entity comprises:

Obtaining a word vector of each of the participles in the part of speech sequence of the text to be predicted;

Multiplying a word vector of each word segment in the part-of-speech sequence of the text to be predicted and an attenuation factor to obtain a vector of each word segment in the part-of-speech sequence of the text to be predicted;

A vector of the word segment corresponding to the target entity in the text to be predicted is used as a vector of a target entity in the part-of-speech sequence of the text to be predicted.
The method of claim 2, further comprising:

Obtaining any one or combination of a part of speech vector, a word packet vector, and a vector of a relative target entity distance of each participle of the part of speech sequence of the text to be predicted;

Combining the word vector of each participle in the part-of-speech sequence of the text to be predicted, and the part-of-speech vector of each part of the obtained part-of-speech sequence of the text to be predicted, the word packet vector, and the vector of the relative target entity distance One or a combination, obtaining an initial vector of each of the participles in the part of speech sequence of the text to be predicted;

The multiplying the word vector and the attenuation factor of each participle of the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted includes:

Multiplying an initial vector and an attenuation factor of each participle in the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted.
The method according to claim 2, wherein if the word segmentation corresponding to the target entity in the text to be predicted includes a plurality, the vector of the plurality of word segments corresponding to the target entity in the text to be predicted is The average value is used as a vector of the target entity in the part-of-speech sequence of the text to be predicted.
The method according to claim 1, wherein the entity emotion prediction model predicts a vector of each word segment in the part-of-speech sequence of the text to be predicted and a vector of the target entity to obtain the text to be predicted The prediction of the sentiment orientation of the target entity, including:

Performing weighted averaging processing on the vector of each participle in the part of speech sequence of the text to be predicted, and obtaining a vector weighted by the part of speech sequence of the text to be predicted;

Multiplying a vector of the target entity in the part of speech sequence of the text to be predicted by a first matrix to obtain a derived vector of the target entity;

And obtaining a feature vector according to the vector weighted by the part of speech sequence of the text to be predicted, and/or the derived vector of the target entity in the part of speech sequence of the text to be predicted;

The feature vector is processed by using a softmax function to obtain a probability output vector, wherein the probability output vector includes: a probability value of the target entity in the text to be predicted, respectively, under the sentiment orientation of the preset category.
The method according to claim 1, wherein the constructing process of the entity sentiment prediction model comprises:

Performing word segmentation processing on the training text to obtain a part-of-speech sequence of the training text;

Obtaining a vector of each participle in the part of speech sequence of the training text and a vector of the target entity;

Performing weighted averaging processing on the vector of each participle in the part of speech of the training text to obtain a vector weighted by the part of speech of the training text;

Multiplying a vector of the target entity in the part-of-speech sequence of the training text with a first matrix to obtain a derived vector of the target entity in the part-of-speech sequence of the training text;

Obtaining a feature vector according to the weighted vector of the part of speech of the training text, and/or the derived vector of the target entity in the part of speech of the training text;

The feature vector is processed by using a softmax function to obtain a probability output vector, wherein the probability output vector includes: a probability value of the target entity in the training text under the sentiment orientation of the preset category;

Performing a cross-entropy operation on the probability output vector and the artificial annotation category of the training text to obtain a loss function;

Optimizing the loss function, and updating the first parameter according to the optimized loss function until the probability output vector obtained by predicting the training text by using the updated feature vector obtained by the updated first parameter and the training text The manual labeling category is equivalent; wherein the first parameter comprises a vector of each of the first matrix, the softmax function, and the part of speech of the training text;

And using the updated second parameter as an entity sentiment prediction model; wherein the second parameter comprises: the first matrix and the softmax function.
An entity sentiment analysis apparatus, comprising:

An obtaining unit, configured to obtain a text to be predicted;

a word segmentation unit, configured to perform word segmentation processing on the text to be predicted, to obtain a part-of-speech sequence of the text to be predicted;

a generating unit, configured to obtain a vector of each participle in the part of speech sequence of the text to be predicted and a vector of the target entity;

a prediction unit, configured to predict, by using an entity sentiment prediction model, a vector of each participle in the part-of-speech sequence of the text to be predicted and a vector of the target entity, to obtain a prediction result of the sentiment orientation of the target entity in the text to be predicted Wherein the entity sentiment prediction model is constructed based on a first principle; the first principle comprises: iteratively updating parameters in the neural network algorithm until the feature vector of the training text is performed by using a neural network algorithm after updating the parameters The prediction result obtained by the prediction is equivalent to the manual annotation result; the feature vector of the training text is obtained according to the vector of the part-of-speech sequence of the training text and the vector of the target entity in the part-of-speech sequence of the training text.
The device according to claim 7, wherein the generating unit comprises:

a first obtaining unit, configured to respectively obtain a word vector of each participle in the part of speech sequence of the text to be predicted;

a second obtaining unit, configured to multiply a word vector and an attenuation factor of each participle of the part-of-speech sequence of the text to be predicted to obtain a vector of each participle in the part-of-speech sequence of the text to be predicted;

And generating a subunit, configured to use a vector of the word segment corresponding to the target entity in the text to be predicted as a vector of the target entity in the part of speech sequence of the text to be predicted.
A storage medium, comprising: a stored program, wherein the device in which the storage medium is located is controlled to perform entity sentiment analysis according to any one of claims 1-6 while the program is running method.
A processor, wherein the processor is configured to execute a program, wherein the program is executed to perform the entity sentiment analysis method according to any one of claims 1-6.