A text emotion extraction method, system and electronic device based on granular sphere computing

Abstract

The present invention relates to the field of computer technology and discloses a text sentiment extraction method, system, and electronic device based on granular sphere computing. The method comprises the following steps: S1, extracting high-level semantic representation information of an input text using a Bert model, clustering the input text based on the high-level semantic representation information, dividing the input text into a plurality of granular spheres, and classifying the granular spheres to extract aspect words and opinion words from the input text; S2, positionally embedding each word pair consisting of an aspect word and an opinion word into a BLSTM model based on the distance between the aspect word and the opinion word. After encoding by the BLSTM model, the hidden states of the aspect words and opinion words are connected to perform sentiment classification. The present invention proposes a granular sphere-based computing method for extracting aspect words and opinion words, solving the problem of word overlap.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a text sentiment extraction method, system, and electronic device based on granular sphere computing. Background Art

[0002] In recent years, with the gradual maturity of natural language processing and the rapid development of artificial intelligence, the amount of online information on various social platforms has grown exponentially. Due to the sheer volume and complexity of this data, manual sifting through this information is time-consuming and labor-intensive. Therefore, automated computer-assisted network quality management and sentiment analysis have become increasingly important. Consequently, uncovering user sentiment within this massive data, analyzing individual user preferences, and selecting high-quality content have become key research topics for many researchers.

[0003] Sentiment analysis is currently a key research task in the field of natural language processing. It can help platforms or businesses understand the emotional needs and attitudes of consumers or users, thereby improving and optimizing products and after-sales services, and enhancing their core competitiveness. It can also help government departments understand public demand, guide public opinion, and formulate relevant policies. Traditional sentiment analysis methods are primarily rule-based or dictionary-based. These methods cannot accurately analyze the semantics of sentences, have poor adaptability to new domains and vocabulary, and have low recognition rates. However, aspect-sentiment triple extraction can target specific aspect words and mine the corresponding aspect-sentiment words and the sentiment polarity they express. This task can be defined as (aspect, opinion, sentiment). Among existing sentiment extraction methods, most research methods can only mine shallow semantic information, resulting in poor classification performance and poor performance in span-level word extraction. Furthermore, most sentences contain multiple aspect-sentiment words, resulting in poor recognition performance for such overlapping data. Current technical solutions cannot effectively extract these overlapping elements. Summary of the Invention

[0004] The present invention provides a text sentiment extraction method, system, and electronic device based on granular sphere computing, thereby solving the problem of word overlap.

[0005] The present invention is achieved through the following technical solutions:

[0006] A text sentiment extraction method based on granular sphere computing includes the following steps:

[0007] S1. Extracting high-level semantic representation information of an input text using a Bert model, clustering the input text according to the high-level semantic representation information of the input text, dividing the input text into a plurality of spheres, and classifying the spheres to extract aspect words and opinion words of the input text;

[0008] S2. Based on the distance between the aspect words and the opinion words, each word pair consisting of the aspect words and the opinion words is positionally embedded into the BLSTM model. After the BLSTM model encoding is completed, the hidden states from the aspect words and the opinion words are connected to perform sentiment classification.

[0009] As an optimization, before extracting the input text through the Bert model, all words in the input text are sequence-tagged according to the word labels, and then the tagged input text is input into the input layer of the Bert model to obtain word vectors, and the obtained word vectors are input into the Bert model.

[0010] As an optimization, each of the word tags includes one of a sentiment tag, an aspect tag, an opinion tag, or a meaningless tag.

[0011] As an optimization, the cluster division is specifically to divide the input text into clusters according to the sentiment label of the input text through the K-means clustering method to obtain a number of spheres, each of which has an opinion label or an aspect label or a meaningless label.

[0012] As an optimization, when dividing, the purity of each ball is compared with the preset purity threshold. If the purity of the ball is not less than the purity threshold, the ball stops dividing. Otherwise, the ball continues to divide until the purity of all balls is not less than the purity threshold.

[0013] As an optimization, the specific process of classifying the plurality of spheres is as follows:

[0014] A1, through the softmax function Calculate the aspect probability P of each sphere marked as an aspect label i (ap) , then use To calculate the probability P of each ball marked as an opinion label i (op) , the aspect probability and opinion probability are both called label probability, specifically:

[0015]

[0016]

[0017] Among them, W t (ap) and are the weights and biases of aspect labels, W t (op) 、 are the weight and bias of opinion labels, respectively;

[0018] A2. Determine whether the ball is an aspect word or an opinion word based on the label probability of the ball.

[0019] As an optimization, the specific steps of S2 are:

[0020] S2.1. Filter out the spheres belonging to aspect words and opinion words;

[0021] S2.2. Combining the screened spheres belonging to aspect words and opinion words to form aspect-opinion word pairs, and setting text formats based on the positions of the aspect-opinion word pairs in the input text. Simultaneously, setting text formats different from aspect words and opinion words based on the positions of non-opinion words and non-aspect words in the input text, thereby converting the input text into a specific text format.

[0022] S2.3. Input the input text in a specific text format and the word vector corresponding to the input text into the BLSTM model for prediction to obtain the sentiment label of the aspect-opinion word pair, and then filter out the aspect-opinion-sentiment triples that meet the conditions, thereby completing the sentiment classification, and filter out the attribute-sentiment pairs that meet the conditions, thereby completing the sentiment classification.

[0023] As an optimization, the text format of the aspect words and opinion words is set to 1, and the text format of the non-opinion words and non-aspect words is set to 0.

[0024] The present invention also discloses a text emotion extraction system based on granular sphere computing, comprising:

[0025] An extraction module, configured to extract high-level semantic representation information of an input text using a Bert model, cluster the input text according to the high-level semantic representation information of the input text, divide the input text into a plurality of spheres, and classify the spheres to extract aspect words and opinion words from the input text;

[0026] The classification module is used to positionally embed each word pair consisting of the aspect word and the opinion word into the BLSTM model according to the distance between the aspect word and the opinion word. After the BLSTM model encoding is completed, the hidden states from the aspect word and the opinion word are connected to perform sentiment classification.

[0027] The present invention also discloses an electronic device, comprising at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a text sentiment extraction method based on granular sphere calculation as described above.

[0028] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0029] The present invention proposes a granular sphere-based calculation method to extract aspect words and opinion words, thereby solving the problem of word overlap. At the same time, granular sphere calculation can improve classification performance and accuracy, and ultimately improve the accuracy of sentiment classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings:

[0031] Figure 1 This is a model diagram of a text sentiment extraction system based on granular sphere computing according to the present invention;

[0032] Figure 2 This is an example diagram of extracting aspect words and opinion words through granular spherical calculation. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0034] Example 1

[0035] A text sentiment extraction method based on granular sphere computing includes the following steps:

[0036] S1. Extracting high-level semantic representation information of an input text using a Bert model, clustering the input text according to the high-level semantic representation information of the input text, dividing the input text into a plurality of spheres, and classifying the spheres to extract aspect words and opinion words of the input text;

[0037] In this embodiment, before extracting the input text through the Bert model, all words in the input text are sequence-labeled according to word labels, and then the labeled input text is input into the input layer to obtain word vectors, and the obtained word vectors are input into the Bert model.

[0038] Specifically, step 1: first preprocess the input text data.

[0039] All data in the training set of input text are sequence labeled. There are four main types of data: sentiment labels, aspect words, opinion words, and meaningless words.

[0040] There are four main sentiment tags: {NEG, POS, NEU, NOT} (positive, negative, neutral, none). The aspect word attribute is marked as aspect, the opinion word attribute is marked as opinion, and meaningless words are marked as N.

[0041] For example, if the input text is "My phone screen is big and beautiful", use n to mark useless words, use B-aspect to mark the beginning of aspect words, use I-aspect to mark the inside of aspect words, use B-opinion to mark the beginning of opinion words, and I-opinion to mark the inside of opinion words, then the sentence will be marked as follows:

[0042] n,n,B-aspect,I-aspect,I-aspect,I-aspect B-opinion,I-opinion,n,n,B-opinion,I-opinion.

[0043] Step 2: Word embedding

[0044] The input text is subjected to word embedding. In the input layer, all words are mapped to a high-dimensional vector space. The high-dimensional vector space of the present invention is a pre-trained 300-dimensional word vector, which is used to contextualize all words. Here, all words refer to all the characters that make up the Chinese sentence, or every word in the case of English text. The mapping is based on existing technology, and pre-trained word vector information can be obtained through Glove, which will not be repeated here.

[0045] After obtaining the word vector of the input text, proceed to step three:

[0046] Step 3: Extract high-level semantic representation of text through the Bert model

[0047] The word vector obtained from the input layer is input into the Bert model, through which the high-level semantics of the text can be extracted.

[0048] After extracting the high-level semantic representation information of the input text (the semantic vector of the entire sentence, including all the words in the sentence), the next steps are carried out.

[0049] Step 4: Generate spheres from the high-level semantic representation information, that is, cluster the input text according to the high-level semantic representation information of the input text by clustering, and divide the input text into a number of spheres.

[0050] This step primarily generates several spheres from the high-level semantic representation information obtained in step 3. The spheres primarily use the K-means clustering method to partition semantic information with the same word label (the word labels here only include the three groups of labels: viewpoint, aspect, and meaningless). After partitioning into multiple clusters, each cluster will have a cluster center and radius. The sphere radius is represented by the maximum Euclidean distance from all points in the sphere to the sphere center, which is defined as follows:

[0051] (V i is the semantic vector of each word, and N is the number of samples in the sphere)

[0052] r1=max(dis(x i ,c))

[0053] r1——maximum radius C——center of the sphere

[0054] dis(x i ,c)——semantic vector x in the sentence i Euclidean distance to the center c

[0055] During the sphere generation process, the entire sentence is represented by a sphere. Each sphere is split using a randomly initialized center until the purity threshold T of the sphere exceeds 0.9, at which point the sphere is split. Here, the purity threshold T serves as the stopping condition for sphere generation, determining the quality of the generated spheres. The purity of each sphere is calculated to see if it is less than the purity threshold T. If so, the sphere is split again; if it is greater, the sphere is retained. K-means clustering is used to cluster the input text according to its sentiment label, generating several spheres. Each sphere contains at least one sentiment-aspect-opinion word pair, and the sentiment corresponding to the aspect-opinion word pairs within the same sphere is consistent.

[0056] Step 5: Classify the balls using the softmax function

[0057] The specific process is:

[0058] A1, through the softmax function Calculate the aspect probability P of each sphere marked as an aspect label i (ap) , then use To calculate the probability P of each ball marked as an opinion label i (op) , specifically:

[0059]

[0060]

[0061] Among them, W t (ap) and are the weights and biases of aspect labels, W t (op) 、 are the weights and biases of opinion labels respectively; weights and biases are the coefficients of the softmax classification function, which are randomly initialized, but will be optimized and adjusted according to the learning of each round of data. This is the existing technology and will not be described here.

[0062] A2. Determine whether the ball is an aspect word or an opinion word based on the probability of the ball. For example, if the aspect probability P i (ap) Greater than the opinion probability P i (op) , then the ball is defined as the aspect probability.

[0063] This step uses To calculate the probability P of each ball marked as an aspect i (ap) , then use To calculate the probability P of each ball being marked as an opinion i (op) , where W t (op) ,W t (ap) and are the learnable weights and biases.

[0064] S2. Based on the distance between the aspect words and the opinion words, each word pair consisting of the aspect words and the opinion words is positionally embedded into the BLSTM model. After the BLSTM model encoding is completed, the hidden states from the aspect words and the opinion words are connected to perform sentiment classification.

[0065] Specifically, it includes step 6: sentiment classification of opinion words and aspect words

[0066] S2.1. Filter out the spheres belonging to aspect words and opinion words; Figure 2, the screened balls have 3 opinion words and 2 aspect words.

[0067] S2.2, combining the screened spheres belonging to aspect words and opinion words to form aspect-opinion word pairs, and setting text formats according to the positions of the aspect words and opinion words in the input text. At the same time, setting text formats different from aspect words and opinion words according to the positions of non-opinion words and non-aspect words in the input text, so that the input text is converted into a specific text format; Figure 2 As shown, the non-opinion and aspect words here are "my, and, ".

[0068] S2.3. Input the input text in a specific text format and the word vector corresponding to the input text into the BLSTM model for prediction, obtain the sentiment label of the aspect-opinion word pair, and then filter out the aspect-opinion-sentiment triples that meet the conditions, thereby completing the sentiment classification.

[0069] Find these aspect words and opinion words in the sentence and set them to 1, and set non-these words to 0, to obtain some sequences of 0, 1, 0, 1, 0, 0, and then input them into the Bilstm model together with their word vectors.

[0070] In this step, the N opinion words and M aspect words obtained in step 5 are randomly combined, resulting in N*M possible combinations. These combinations are then screened. After the BLSTM model is encoded, the hidden states of the aspect and opinion words are concatenated (here, the BLSTM model learns the semantic information between the word pairs and then splices them together, with splicing referring to dimensional splicing) to perform sentiment classification. This stage primarily involves filtering out qualified attribute-sentiment pairs from these combinations (sentiment-aspect-opinion triplets). For example, if the BLSTM model predicts (aspect-opinion-positive), it is retained; if it predicts (aspect-opinion-none), it is discarded. When the sentiment prediction is positive, negative, or neutral, these sentiment pairs are retained and compared with the true label pairs to calculate the loss function, thereby optimizing the BLSTM model.

[0071] After the BLSTM model calculates the loss function of the predicted value and the true value, it uses the gradient descent method to optimize the loss function, thereby reversely updating the parameters of the BLSTM model.

[0072] Because the present invention uses spheres to cluster semantic information, gradients can be interrupted during backpropagation. For example, if the map size before sphere clustering is 300*32 and becomes 128*32 after clustering, only 128 gradients are transmitted back to the sphere layer, causing gradient propagation to be interrupted. The calculation of spheres is described as the sphere layer in the neural network. Therefore, during backpropagation, the present invention replicates the gradient at the center of the sphere n times, where n is the number of words clustered within the sphere, and returns it instead of the gradients of other samples within the sphere. This ensures proper gradient propagation.

[0073] For example, during back propagation, if 10 words in a sentence are clustered into 4 spheres, the dimension will change for back propagation. Therefore, the present invention copies n copies of the vector containing the 4 words in the center of a sphere and returns them back, so that 4 becomes 10 again, and the gradient can be propagated and calculated normally.

[0074] With the word vector and the sequence of 0, 1, 0, 1 representing the aspect and opinion words, they are input into the BLSTM model together. (It can be understood that 0, 1, 0, 1 is equivalent to a weight representation, 0 is meaningless and the BLSTM model will not consider it, and 1 is meaningful). Then the BLSTM model will predict these aspect-opinion pairs. The sentiment labels of these aspect-opinion pairs are predicted to be four categories: positive, negative, neutral, and meaningless. The sentiment labels of the predicted results are meaningless and will be discarded, while the other sentiment labels will be paired with the aspect and opinion words of the first stage to form a triplet (aspect-opinion-positive), which is also the ultimate goal of the prediction of the present invention.

[0075] During the previous sequence labeling, the correct triples were marked and then compared with the triples predicted by the BLSTM model. Ultimately, the BLSTM model was optimized. The optimization method is based on existing technology and will not be described in detail here.

[0076] Example 2

[0077] The present invention also discloses a text emotion extraction system based on granular sphere computing, comprising:

[0078] An extraction module, configured to extract high-level semantic representation information of an input text using a Bert model, cluster the input text according to the high-level semantic representation information of the input text, divide the input text into a plurality of spheres, and classify the spheres to extract aspect words and opinion words from the input text;

[0079] The classification module is used to positionally embed each word pair consisting of the aspect word and the opinion word into the BLSTM model according to the distance between the aspect word and the opinion word. After the BLSTM model encoding is completed, the hidden states from the aspect word and the opinion word are connected to perform sentiment classification.

[0080] Example 3

[0081] The present invention also discloses an electronic device, comprising at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a text sentiment extraction method based on granular sphere calculation as described above.

[0082] The present invention proposes a granular sphere-based calculation method to extract aspect words and opinion words. The granular sphere calculation method is used to extract span-level words, and words belonging to an aspect label are clustered together. For example, "mobile phone screen" has four words, which will be clustered together into a granular sphere. In this way, words with span levels can be extracted, and then the aspect labels and opinion labels are combined into NxM types. In this way, the extraction of overlapping words can be solved. For example, "the mobile phone screen is large and good-looking". "Mobile phone screen" should be paired with both "large" and "good-looking". The NxM types include these two cases, thus solving the problem of word overlap.

[0083] Because the dimensionality of sentences during classification is reduced, if a sentence consists of 10 words, each word needs to be predicted. However, granular ball computing can cluster the vectors that can constitute semantic information into four balls, and then prediction and classification are performed on these four balls. Therefore, granular ball computing can improve classification performance and accuracy, and ultimately improve the accuracy of sentiment classification.

[0084] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A text sentiment extraction method based on granular sphere computing, characterized in that: The steps include: S1. Extracting high-level semantic representation information of an input text using a Bert model, clustering the input text according to the high-level semantic representation information of the input text, dividing the input text into a plurality of spheres, and classifying the spheres to extract aspect words and opinion words of the input text; In the process of generating balls, the entire sentence is a ball. When dividing, the purity of each ball is compared with the preset purity threshold. If the purity of the ball is not less than the purity threshold, the ball stops dividing. Otherwise, the ball continues to divide until the purity of all balls is not less than the purity threshold. The specific process of classifying the plurality of pellets is as follows: A1, through the softmax function Calculate the aspect probability P of each sphere marked as an aspect label i (ap) , then use To calculate the probability P of each ball marked as an opinion label i (op) , the aspect probability and opinion probability are both called label probability, specifically: Among them, W t (ap) and are the weights and biases of aspect labels, W t (op) 、 are the weight and bias of opinion labels, respectively; A2. Determine whether the ball is an aspect word or an opinion word based on the label probability of the ball; S2. Positionally embedding each word pair consisting of the aspect word and the opinion word into a BLSTM model based on the distance between the aspect word and the opinion word. After encoding by the BLSTM model, concatenating the hidden states of the aspect word and the opinion word to perform sentiment classification. The specific steps of S2 are: S2.

1. Filter out the spheres belonging to aspect words and opinion words; S2.

2. Combining the screened spheres belonging to aspect words and opinion words to form aspect-opinion word pairs, and setting text formats based on the positions of the aspect-opinion word pairs in the input text. Simultaneously, setting text formats different from aspect words and opinion words based on the positions of non-opinion words and non-aspect words in the input text, thereby converting the input text into a specific text format. S2.

3. Input the input text in a specific text format and the word vector corresponding to the input text into the BLSTM model for prediction to obtain the sentiment label of the aspect-opinion word pair, and then filter out the aspect-opinion-sentiment triples that meet the conditions, thereby completing the sentiment classification, and filter out the attribute-sentiment pairs that meet the conditions, thereby completing the sentiment classification.

2. A text sentiment extraction method based on granular sphere calculation according to claim 1, characterized in that: Before extracting the input text through the Bert model, all words in the input text are sequence-tagged according to the word labels, and then the tagged input text is input into the input layer of the Bert model to obtain word vectors, and the obtained word vectors are input into the Bert model.

3. The text sentiment extraction method based on granular sphere calculation according to claim 2 is characterized in that: Each of the word tags includes one of a sentiment tag, an aspect tag, an opinion tag, or a meaningless tag.

4. The text sentiment extraction method based on granular sphere calculation according to claim 3 is characterized in that: The cluster division is specifically to divide the input text into clusters according to the word labels of the input text by using the K-means clustering method to obtain a number of spheres, each of which has an opinion label or an aspect label or a meaningless label.

5. The text sentiment extraction method based on granular sphere calculation according to claim 1 is characterized in that: The text format of the aspect words and opinion words is set to 1, and the text format of the non-opinion words and non-aspect words is set to 0.

6. A text sentiment extraction system based on granular sphere computing, characterized in that: include: An extraction module, configured to extract high-level semantic representation information of an input text using a Bert model, cluster the input text according to the high-level semantic representation information of the input text, divide the input text into a plurality of spheres, and classify the spheres to extract aspect words and opinion words from the input text; In the process of generating balls, the entire sentence is a ball. When dividing, the purity of each ball is compared with the preset purity threshold. If the purity of the ball is not less than the purity threshold, the ball stops dividing. Otherwise, the ball continues to divide until the purity of all balls is not less than the purity threshold. The specific process of classifying the plurality of pellets is as follows: A1, through the softmax function Calculate the aspect probability P of each sphere marked as an aspect label i (ap) , then use To calculate the probability P of each ball marked as an opinion label i (op) , the aspect probability and opinion probability are both called label probability, specifically: Among them, W t (ap) and are the weights and biases of aspect labels, W t (op) 、 are the weight and bias of opinion labels, respectively; A2. Determine whether the ball is an aspect word or an opinion word based on the label probability of the ball; A classification module is used to positionally embed each word pair consisting of an aspect word and an opinion word into a BLSTM model based on the distance between the aspect word and the opinion word. After encoding by the BLSTM model, the hidden states of the aspect word and the opinion word are connected to perform sentiment classification. The specific operation steps of the classification module are: S2.

1. Filter out the spheres belonging to aspect words and opinion words; S2.

2. Combining the screened spheres belonging to aspect words and opinion words to form aspect-opinion word pairs, and setting text formats based on the positions of the aspect-opinion word pairs in the input text. Simultaneously, setting text formats different from aspect words and opinion words based on the positions of non-opinion words and non-aspect words in the input text, thereby converting the input text into a specific text format. S2.

3. Input the input text in a specific text format and the word vector corresponding to the input text into the BLSTM model for prediction to obtain the sentiment label of the aspect-opinion word pair, and then filter out the aspect-opinion-sentiment triples that meet the conditions, thereby completing the sentiment classification, and filter out the attribute-sentiment pairs that meet the conditions, thereby completing the sentiment classification.

7. An electronic device, characterized in that: It includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute a text sentiment extraction method based on granular sphere calculation as described in any one of claims 1 to 5.

Images