Word-of-mouth information evaluation method based on data analysis

Abstract

This invention discloses a data analysis-based method for evaluating word-of-mouth information, belonging to the field of information evaluation technology. It utilizes web crawling technology to acquire content generated in forums; the original sentences are placed into a filter, and after filtering, formal word-of-mouth information is obtained; adjectives, degree adverbs, and negation words in the original sentences are extracted using NLP technology and evaluated; the sentiment assignment of degree adverbs and adjectives is combined to score the positivity of the word-of-mouth evaluation, and the positivity of the word-of-mouth evaluation is determined based on the score results; the obtained word-of-mouth polarity and positivity of the word-of-mouth evaluation are output. This invention divides the entire word-of-mouth mining task into multiple sub-tasks for simplification, greatly alleviating the problems of data scarcity and annotation quality. Furthermore, by calculating the positivity score of the word-of-mouth, it can maintain high accuracy while extracting more word-of-mouth information and judging the positivity of the word-of-mouth evaluation.

Description

Technical Field

[0001] This invention relates to the field of information evaluation technology, specifically a method for evaluating word-of-mouth information based on data analysis. Background Technology

[0002] With the development of the internet, online forums, as an important social media platform, have become a crucial channel for people to obtain and share information, exchange views, and shape public opinion. Online forums contain a large amount of user-generated content, including evaluations and feedback on various products, services, events, and individuals. This content reflects users' genuine attitudes and preferences and has significant reference value. Therefore, extracting useful word-of-mouth information from online forums is of great significance for improving product quality, optimizing service levels, enhancing user experience, and monitoring public opinion dynamics.

[0003] However, word-of-mouth information in online forums has the following characteristics: 1. Massive: The amount of word-of-mouth information in online forums is enormous, making it difficult to process and analyze manually; 2. Heterogeneous: Word-of-mouth information in online forums involves multiple fields and topics, making it difficult to unify standards and norms; 3. Noise: The quality of word-of-mouth information in online forums varies, making it difficult to filter and correct; 4. Dynamic: Word-of-mouth information in online forums changes with time, events, users, etc., making it difficult to capture and update. Summary of the Invention

[0004] The purpose of this invention is to provide a data analysis-based method for evaluating word-of-mouth information in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for evaluating word-of-mouth information based on data analysis, the method comprising the following steps:

[0006] S1. Obtain the content generated in the forum using web crawling technology and define it as the original sentence;

[0007] S2. Place the original sentence into the filter, and obtain the official reputation information after passing through the filter.

[0008] S3. Use NLP technology to extract adjectives from the original sentences, classify the adjectives and evaluate their emotional tendencies to obtain word-of-mouth opinions;

[0009] S4. Use NLP technology to extract degree adverbs and negation words from the original sentence, and evaluate the degree of sentiment enhancement of the extracted degree adverbs;

[0010] S5. Combining the emotional assignments of degree adverbs and adjectives, we score the positiveness of word-of-mouth evaluation and determine the positiveness of word-of-mouth evaluation based on the score results.

[0011] S6. Output the polarity of word-of-mouth and the positivity of word-of-mouth evaluation.

[0012] Furthermore, in step S1: web crawling technology is used to crawl data from the target online forum and extract user-generated content from it; the extracted user-generated content is then stored as original sentences in a local database.

[0013] Furthermore, in step S2: the original sentence is placed into a filter, and after filtering by part of speech and stop words, it is recombined into nouns or noun phrases. The recombined nouns and noun phrases are defined as candidate words for word-of-mouth. The candidate words for word-of-mouth are placed into a second filter, and after being filtered by the fine-tuned Bert model, the official words for word-of-mouth are obtained.

[0014] Furthermore, step S2 includes the following steps:

[0015] S2-1: Import the data required for part-of-speech tagging from NLTK Corpora by loading the nltk library of Python, and import the stop word dataset;

[0016] S2-2: Perform POS Tagging on the original sentence after word segmentation to tag the part of speech of each word;

[0017] S2-3: After obtaining the part of speech of each word, filter out the parts of speech and special symbols including 'CC', 'CD', 'DT', 'JJ', 'JJR', 'JJS', 'MD', 'RB', 'RBR', 'UH', 'VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ';

[0018] S2-4: Use the stop word dataset to perform another round of filtering on the results from the previous step;

[0019] S2-5: Combine the remaining words according to their position in the original sentence, and the combined words or phrases are entered into the candidate queue for word-of-mouth evaluation.

[0020] S2-6: A batch of manually labeled data is used to fine-tune BERT, which calculates the probability that a certain candidate word related to word reputation corresponds to the actual word reputation information in the original sentence. BERT receives the combination of the candidate word related to word reputation and the original sentence as input, and outputs the probability value p of judging it as word reputation. If p is greater than z, then the input candidate word related to word reputation is considered to be a word related to word reputation in the original sentence.

[0021] Furthermore, in step S3: the original sentence is segmented using the Jieba library in natural language processing technology, and adjectives are extracted and classified into three types: positive adjectives, negative adjectives, and neutral adjectives. The sentiment tendency of the original sentence is determined based on the adjective classification, and the number of each type of adjective is counted and recorded. A label value r is assigned to mark the positive or negative of the adjectives. r is assigned a value in three cases: r=1 when the adjective is positive, r=-1 when the adjective is negative, and r=0 when the adjective is neutral. The number of positive adjectives is used to represent positive word-of-mouth evaluation, and the number of negative adjectives is used to represent negative word-of-mouth evaluation. The number of adjectives and the sentiment assignment are used to calculate the polarity of word-of-mouth.

[0022] Furthermore, in step S3: using the Boson NLP sentiment dictionary to assign sentiment values ​​to adjectives of different categories, determining whether an adjective can be used as a word-of-mouth opinion word based on the extracted sentiment values, and calculating the average sentiment value of the extracted group of adjectives by weighting the sentiment values ​​of each adjective when the original sentence contains n adjectives, and setting a standard value s for the sentiment values, comparing the sentiment values ​​of the adjectives or the average sentiment values ​​of a group of adjectives with the standard value s, and determining that an adjective can be used as a word-of-mouth opinion word when the extracted sentiment value of an adjective exceeds the standard value s.

[0023] Furthermore, in step S4: degree adverbs and negation words in the original sentence are extracted using natural language processing technology, and degree adverbs are classified by level. The degree adverbs are graded using the score q of the degree adverb dictionary. Degree adverbs with a score less than 1 are defined as first-level degree adverbs, and degree adverbs with a score greater than 1 are defined as second-level degree adverbs. The degree of intensity of the degree adverb on the emotional tendency is defined. First-level degree adverbs are defined as emotional weakening degree adverbs, and second-level degree adverbs are defined as emotional strengthening degree adverbs.

[0024] Furthermore, in step S4: the number m of negative words in the original sentence is recorded, and after extracting negative words from the original sentence, the negative words are analyzed in combination with the sentiment of the adjectives. Negative words that are odd in number and located before the adjectives can reverse the sentiment of the adjectives.

[0025] Furthermore, in step S5: the adjectives, degree adverbs, and negation words extracted from the original sentence are combined to analyze the positive evaluation level of the original sentence, and a scoring method for the positive evaluation level is established. The positive evaluation level score is calculated by superimposing the sentiment assignment of adjectives and the scores of degree adverbs at different levels, and the positive evaluation level score is calculated using a formula:

[0026] ;

[0027] Where y represents the calculated positive rating of word-of-mouth, q represents the emotional intensification rating of the degree adverb, n represents the number of adjectives, and x represents the positive rating of the word-of-mouth evaluation. i The system assigns sentiment values ​​to different adjectives, where r is the sentiment value marker for the adjective (r = 1, 0, -1), and m represents the number of negative words in the original sentence. It also defines a positive sentiment rating for positive adjectives, a negative sentiment rating for negative adjectives, and a zero sentiment rating for neutral adjectives. In the presence of negative words, the negative words can reverse the sentiment rating. The established sentiment rating method not only obtains the polarity of sentiment opinions but also determines whether the sentiment rating is positive, neutral, or negative, and further determines the degree of positivity based on the rating results.

[0028] Furthermore, in step S6: based on the obtained word-of-mouth aspects and opinions, as well as the initial original sentence, three combinations are arranged: "word-of-mouth aspect + original sentence", "word-of-mouth aspect + word-of-mouth opinion", and "word-of-mouth aspect + original sentence + word-of-mouth opinion". BERT is fine-tuned for each combination, with the current combination as input and the probabilities of the three labels as output. The combination with the highest accuracy among the three is selected as the final solution, and the label calculated by BERT is the word-of-mouth polarity corresponding to the current word-of-mouth aspect. The calculated positive word-of-mouth evaluation score is also output, yielding the evaluation result of the word-of-mouth information. The calculated positive evaluation score has three possibilities: y>0, y=0, and y<0, corresponding to the positive, neutral, and negative states of the word-of-mouth evaluation, respectively.

[0029] Compared with existing technologies, the beneficial effects of this invention are: this invention divides the entire word-of-mouth mining task into multiple sub-tasks for simplification. In the use of BERT, only classification tasks are performed in each sub-task instead of generative tasks. In this way, only less training data is needed to achieve a high accuracy rate, which greatly alleviates the problems of data scarcity and annotation quality. Furthermore, by scoring the positiveness of word-of-mouth, it is possible to maintain a high accuracy rate while extracting more word-of-mouth information, and at the same time judge the positiveness of word-of-mouth evaluation. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow for the reputation information evaluation method based on data analysis of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, this invention provides a technical solution: a method for evaluating word-of-mouth information based on data analysis. This method includes the following steps:

[0033] S1. Obtain the content generated in the forum using web crawling technology and define it as the original sentence;

[0034] S2. Place the original sentence into the filter, and obtain the official reputation information after passing through the filter.

[0035] S3. Use NLP technology to extract adjectives from the original sentences, classify the adjectives and evaluate their emotional tendencies to obtain word-of-mouth opinions;

[0036] S4. Use NLP technology to extract degree adverbs and negation words from the original sentence, and evaluate the degree of sentiment enhancement of the extracted degree adverbs;

[0037] S5. Combining the emotional assignments of degree adverbs and adjectives, we score the positiveness of word-of-mouth evaluation and determine the positiveness of word-of-mouth evaluation based on the score results.

[0038] S6. Output the polarity of word-of-mouth and the positivity of word-of-mouth evaluation.

[0039] In step S1: Data is crawled from the target online forum using web crawling technology, and user-generated content is extracted from it; the extracted user-generated content is then stored as the original sentences in the local database.

[0040] In step S2: The original sentence is placed into a filter, and after filtering by part of speech and stop words, it is recombined into nouns or noun phrases. The recombined nouns and noun phrases are defined as candidate words for word-of-mouth. The candidate words for word-of-mouth are placed into a second filter, and the official words for word-of-mouth are obtained after being filtered by the fine-tuned Bert model.

[0041] Step S2 includes the following steps:

[0042] S2-1: Import the data required for part-of-speech tagging from NLTK Corpora by loading the nltk library of Python, and import the stop word dataset;

[0043] S2-2: Perform POS Tagging on the original sentence after word segmentation to tag the part of speech of each word;

[0044] S2-3: After obtaining the part of speech of each word, filter out the parts of speech and special symbols including 'CC', 'CD', 'DT', 'JJ', 'JJR', 'JJS', 'MD', 'RB', 'RBR', 'UH', 'VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ';

[0045] S2-4: Use the stop word dataset to perform another round of filtering on the results from the previous step;

[0046] S2-5: Combine the remaining words according to their position in the original sentence, and the combined words or phrases are entered into the candidate queue for word-of-mouth evaluation.

[0047] S2-6: A batch of manually labeled data is used to fine-tune BERT, which calculates the probability that a certain candidate word related to word reputation corresponds to the actual word reputation information in the original sentence. BERT receives the combination of the candidate word related to word reputation and the original sentence as input, and outputs the probability value p of judging it as word reputation. If p is greater than z, then the input candidate word related to word reputation is considered to be a word related to word reputation in the original sentence.

[0048] In step S3: The original sentence is segmented using the Jieba library in natural language processing technology. After extracting the adjectives from the sentence, the adjectives are classified into three types: positive adjectives, negative adjectives, and neutral adjectives. The sentiment tendency of the original sentence is determined based on the adjective classification, and the number of each type of adjective is counted and recorded. A label value r is assigned to mark the positive or negative of the adjectives. r is assigned a value in three cases: r=1 when the adjective is positive, r=-1 when the adjective is negative, and r=0 when the adjective is neutral. The number of positive adjectives is used to represent positive word-of-mouth evaluation, and the number of negative adjectives is used to represent negative word-of-mouth evaluation. The number of adjectives and the sentiment assignment are used to calculate the polarity of word-of-mouth.

[0049] In step S3: The Boson NLP sentiment dictionary is used to assign sentiment values ​​to adjectives of different categories. Based on the sentiment values ​​of the extracted adjectives, it is determined whether the adjective can be used as a word-of-mouth opinion word. In the case that the original sentence contains n adjectives, the sentiment values ​​of each adjective are weighted and averaged to calculate the average sentiment value of the extracted adjective group. A standard value s for the sentiment value is set. The sentiment value of the adjective or the average sentiment value of a group of adjectives is compared with the standard value s. When the sentiment value of the extracted adjective exceeds the standard value s, it is determined that the adjective can be used as a word-of-mouth opinion word.

[0050] In step S4: Degree adverbs and negation words in the original sentence are extracted using natural language processing technology, and degree adverbs are classified by level. The degree adverbs are graded using the score q of degree adverbs in the degree adverb dictionary. Degree adverbs with a score less than 1 are defined as first-level degree adverbs, and degree adverbs with a score greater than 1 are defined as second-level degree adverbs. Furthermore, the degree of intensification of the emotional tendency by degree adverbs is defined. First-level degree adverbs are defined as emotional weakening degree adverbs, and second-level degree adverbs are defined as emotional strengthening degree adverbs.

[0051] In step S4: the number m of negative words in the original sentence is recorded. After extracting negative words from the original sentence, the negative words are analyzed in combination with the sentiment of the adjectives. Negative words that are odd in number and located before the adjectives can reverse the sentiment of the adjectives.

[0052] In step S5: The adjectives, degree adverbs, and negation words extracted from the original sentence are combined to analyze the positive evaluation level of the original sentence, and a scoring method for the positive evaluation level of word-of-mouth is established. The positive evaluation level score is calculated by superimposing the sentiment assignment of adjectives and the scores of degree adverbs at different levels. The positive evaluation level score of word-of-mouth is calculated using a formula:

[0053] ;

[0054] Where y represents the calculated positive rating of word-of-mouth, q represents the emotional intensification rating of the degree adverb, n represents the number of adjectives, and x represents the positive rating of the word-of-mouth evaluation. i The system assigns sentiment values ​​to different adjectives, where r is the sentiment value marker for the adjective (r = 1, 0, -1), and m represents the number of negative words in the original sentence. It also defines a positive sentiment rating for positive adjectives, a negative sentiment rating for negative adjectives, and a zero sentiment rating for neutral adjectives. In the presence of negative words, the negative words can reverse the sentiment rating. The established sentiment rating method not only obtains the polarity of sentiment opinions but also determines whether the sentiment rating is positive, neutral, or negative, and further determines the degree of positivity based on the rating results.

[0055] In step S6: Based on the obtained word-of-mouth aspects and opinions, as well as the original sentences, three combinations are arranged: "word-of-mouth aspect + original sentence", "word-of-mouth aspect + word-of-mouth opinion", and "word-of-mouth aspect + original sentence + word-of-mouth opinion". BERT is fine-tuned for each combination. The input is the current combination, and the output is the probability of the three labels. The combination with the highest accuracy among the three is selected as the final solution. The label calculated by BERT is the word-of-mouth polarity corresponding to the current word-of-mouth aspect. The calculated positive evaluation score is also output to obtain the evaluation result of the word-of-mouth information. The calculated positive evaluation score has three cases: y>0, y=0, and y<0, corresponding to the positive, neutral, and negative states of word-of-mouth evaluation, respectively. Example

[0056] In step S1: Data is crawled from the target online forum using web crawling technology, and user-generated content is extracted from it; the extracted user-generated content is then stored as the original sentences in the local database.

[0057] In step S2: The original sentence is placed into a filter, and after filtering by part of speech and stop words, it is recombined into nouns or noun phrases. The recombined nouns and noun phrases are defined as candidate words for word-of-mouth. The candidate words for word-of-mouth are placed into a second filter, and the official words for word-of-mouth are obtained after being filtered by the fine-tuned Bert model.

[0058] Includes the following steps:

[0059] S2-1: Import the data required for part-of-speech tagging from NLTK Corpora by loading the nltk library of Python, and import the stop word dataset;

[0060] S2-2: Perform POS Tagging on the original sentence after word segmentation to tag the part of speech of each word;

[0061] S2-3: After obtaining the part of speech of each word, filter out the parts of speech such as 'CC', 'CD', 'DT', 'JJ', 'JJR', 'JJS', 'MD', 'RB', 'RBR', 'UH', 'VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ', as well as a series of special symbols;

[0062] S2-4: Use the stop word dataset to perform another round of filtering on the results from the previous step;

[0063] S2-5: Combine the remaining words according to their position in the original sentence, and the combined words or phrases are entered into the candidate queue for word-of-mouth evaluation.

[0064] S2-6: A batch of manually labeled data is used to fine-tune BERT, which calculates the probability that a certain candidate word related to word reputation corresponds to the actual word reputation information in the original sentence. The input to BERT is a combination of the candidate word related to word reputation and the original sentence, and the output is the probability value p of judging it as word reputation. If p is greater than z=0.6, then the input candidate word related to word reputation is considered to be a word related to word reputation in the original sentence.

[0065] In step S3: The original sentence is segmented using the Jieba library in natural language processing technology. After extracting the adjectives from the sentence, the adjectives are classified into three types: positive adjectives, negative adjectives, and neutral adjectives. The sentiment tendency of the original sentence is determined based on the adjective classification, and the number of each type of adjective is counted and recorded. A label value r is assigned to mark the positive or negative of the adjectives. r is assigned a value in three cases: r=1 when the adjective is positive, r=-1 when the adjective is negative, and r=0 when the adjective is neutral. The number of positive adjectives is used to represent positive word-of-mouth evaluation, and the number of negative adjectives is used to represent negative word-of-mouth evaluation. The number of adjectives and the sentiment assignment are used to calculate the polarity of word-of-mouth.

[0066] In step S3: The Boson NLP sentiment dictionary is used to assign sentiment values ​​to adjectives of different categories. Based on the extracted sentiment values ​​of adjectives, it is determined whether the adjective can be used as a word-of-mouth opinion word. In the case that the original sentence contains n adjectives, the sentiment values ​​of each adjective are weighted and averaged to calculate the average sentiment value of the extracted adjective group. The standard value of sentiment assignment is set to s=0.9. The sentiment value of the adjective or the average sentiment value of a group of adjectives is compared with the standard value s=0.9. When the sentiment value of the extracted adjective exceeds the standard value s=0.9, it is determined that the adjective can be used as a word-of-mouth opinion word.

[0067] In step S4: Degree adverbs and negation words in the original sentence are extracted using natural language processing technology, and degree adverbs are classified by level. The degree adverbs are graded using the score q of degree adverbs in the degree adverb dictionary. Degree adverbs with a score less than 1 are defined as first-level degree adverbs, and degree adverbs with a score greater than 1 are defined as second-level degree adverbs. Furthermore, the degree of intensification of the emotional tendency by degree adverbs is defined. First-level degree adverbs are defined as emotional weakening degree adverbs, and second-level degree adverbs are defined as emotional strengthening degree adverbs.

[0068] In step S4: the number m of negative words in the original sentence is recorded. After extracting negative words from the original sentence, the negative words are analyzed in combination with the sentiment of the adjectives. Negative words that are odd in number and located before the adjectives can reverse the sentiment of the adjectives.

[0069] In step S5: The adjectives, degree adverbs, and negation words extracted from the original sentence are combined to analyze the positive evaluation level of the original sentence, and a scoring method for the positive evaluation level of word-of-mouth is established. The positive evaluation level score is calculated by superimposing the sentiment assignment of adjectives and the scores of degree adverbs at different levels. The positive evaluation level score of word-of-mouth is calculated using a formula:

[0070] ;

[0071] Where y represents the calculated positive rating of word-of-mouth, q represents the emotional intensification rating of the degree adverb, n represents the number of adjectives, and x represents the positive rating of the word-of-mouth evaluation. i The system assigns sentiment values ​​to different adjectives, where r is the sentiment value marker for the adjective (r = 1, 0, -1), and m represents the number of negative words in the original sentence. It stipulates that positive adjectives receive a positive rating for positive word-of-mouth evaluation, negative adjectives receive a negative rating, and neutral adjectives receive a zero rating. In the presence of negative words, the negative words can reverse the rating of positive word-of-mouth evaluation. The established method for evaluating the positiveness of word-of-mouth evaluation not only obtains the polarity of opinion but also determines whether the evaluation is positive, neutral, or negative, and further assesses the degree of positiveness based on the rating results.

[0072] In step S6: Based on the obtained word-of-mouth aspects and opinions, and the initial original sentence, three combinations are arranged: "word-of-mouth aspect + original sentence", "word-of-mouth aspect + word-of-mouth opinion", and "word-of-mouth aspect + original sentence + word-of-mouth opinion". BERT is fine-tuned for each combination. The input is the current combination, and the output is the probability of three labels. The combination with the highest accuracy is selected as the final solution, and the label calculated by BERT is the word-of-mouth polarity corresponding to the current word-of-mouth aspect. In our experiments, we found that the "word-of-mouth aspect + original sentence" combination had the highest accuracy. The reason for this might be... Word-of-mouth opinions can concisely and quickly reflect the author's emotional inclination, but at the cost of discarding a large amount of contextual information. Therefore, word-of-mouth opinions have a certain degree of bias and may mislead the actual judgment of word-of-mouth polarity. Furthermore, after scoring the information in the original sentence using the evaluation positivity rating method, the evaluation result of the word-of-mouth information is obtained and output. The calculated evaluation positivity rating has three cases: y>0, y=0, and y<0, which correspond to the positive, neutral, and negative situations of word-of-mouth evaluation positivity, respectively. The higher the word-of-mouth evaluation positivity rating, the higher the positiveness of the word-of-mouth information.

[0073] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A data-driven word-of-mouth information evaluation method, characterized by: Includes the following steps: S1. Obtain the content generated in the forum using web crawling technology and define it as the original sentence; S2. Place the original sentence into the filter, and obtain the official reputation information after passing through the filter. S3. Use NLP technology to extract adjectives from the original sentences, classify the adjectives and evaluate their emotional tendencies to obtain word-of-mouth opinions; S4. Use NLP technology to extract degree adverbs and negation words from the original sentence, and evaluate the degree of sentiment enhancement of the extracted degree adverbs; S5. Combining the emotional assignments of degree adverbs and adjectives, we score the positiveness of word-of-mouth evaluation and determine the positiveness of word-of-mouth evaluation based on the score results. S6. Output the polarity of word-of-mouth and the positivity of word-of-mouth evaluation; In step S2: the original sentence is placed into the filter, and after filtering by part of speech and stop words, it is recombined into nouns or noun groups. The recombined nouns and noun groups are defined as candidate words for word-of-mouth. The candidate words for word-of-mouth are placed into the second filter, and the formal words for word-of-mouth are obtained after being filtered by the fine-tuned Bert model. Step S2 includes the following steps: S2-1: Import the data required for part-of-speech tagging from NLTK Corpora by loading the nltk library of Python, and import the stop word dataset; S2-2: Perform POS Tagging on the original sentence after word segmentation to tag the part of speech of each word; S2-3: After obtaining the part of speech of each word, select and filter out the parts of speech containing 'CC', 'CD', 'DT', 'JJ', 'JJR', 'JJS', 'MD', 'RB', 'RBR', 'UH', 'VB', 'VBD', 'VBG', 'VBN', 'VBP', 'VBZ'; S2-4: Use the stop word dataset to perform another round of filtering on the results from the previous step; S2-5: Combine the remaining words according to their position in the original sentence, and the combined words or phrases are entered into the candidate queue for word-of-mouth evaluation. S2-6: Manually label a batch of data to fine-tune Bert, so that it can calculate the probability that a certain word related to reputation is the actual word related information in the original sentence. Bert receives the combination of word related candidate words and the original sentence as input, and outputs the probability value p of judging it as word related. If p is greater than z, it is considered that the input word related to reputation is the word related to reputation in the original sentence. In step S6: Based on the obtained word-of-mouth aspects and opinions, as well as the initial original sentence, three combinations are arranged: "word-of-mouth aspect + original sentence", "word-of-mouth aspect + word-of-mouth opinion", and "word-of-mouth aspect + original sentence + word-of-mouth opinion". BERT is fine-tuned for each combination. The input is the current combination, and the output is the probability of the three labels. The combination with the highest accuracy among the three is selected as the final solution. The label calculated by BERT is the word-of-mouth polarity corresponding to the current word-of-mouth aspect. The calculated positive word-of-mouth evaluation score is also output, resulting in the evaluation result of the word-of-mouth information. The calculated positive evaluation score has three possibilities: y>0, y=0, and y<0, corresponding to the positive, neutral, and negative states of word-of-mouth evaluation, respectively.

2. The word-of-mouth information evaluation method based on data analysis according to claim 1, characterized in that: In step S1: Data is crawled from the target online forum using web crawling technology, and user-generated content is extracted from it; the extracted user-generated content is then stored as the original sentences in the local database.

3. The method for evaluating word-of-mouth information based on data analysis according to claim 1, characterized in that: In step S3: The original sentence is segmented using the Jieba library in natural language processing technology. After extracting the adjectives from the sentence, the adjectives are classified into three types: positive adjectives, negative adjectives, and neutral adjectives. The sentiment tendency of the original sentence is determined based on the adjective classification, and the number of each type of adjective is counted and recorded. A label value r is assigned to mark the positive or negative of the adjective. r is assigned a value in three cases: r=1 when the adjective is positive, r=-1 when the adjective is negative, and r=0 when the adjective is neutral.

4. The method for evaluating word-of-mouth information based on data analysis according to claim 1, characterized in that: In step S3: The Boson NLP sentiment dictionary is used to assign sentiment values ​​to adjectives of different categories. Based on the sentiment values ​​of the extracted adjectives, it is determined whether the adjective can be used as a word-of-mouth opinion word. In the case that the original sentence contains n adjectives, the sentiment values ​​of each adjective are weighted and averaged to calculate the average sentiment value of the extracted adjective group. A standard value s for the sentiment value is set. The sentiment value of the adjective or the average sentiment value of a group of adjectives is compared with the standard value s. When the sentiment value of the extracted adjective exceeds the standard value s, it is determined that the adjective can be used as a word-of-mouth opinion word.

5. The method for evaluating word-of-mouth information based on data analysis according to claim 1, characterized in that: In step S4: Degree adverbs and negation words in the original sentence are extracted using natural language processing technology, and degree adverbs are classified by level. The degree adverbs are graded using the score q of degree adverbs in the degree adverb dictionary. Degree adverbs with a score less than 1 are defined as first-level degree adverbs, and degree adverbs with a score greater than 1 are defined as second-level degree adverbs. Furthermore, the degree of intensification of the emotional tendency by degree adverbs is defined. First-level degree adverbs are defined as emotional weakening degree adverbs, and second-level degree adverbs are defined as emotional strengthening degree adverbs.

6. The method for evaluating word-of-mouth information based on data analysis according to claim 1, characterized in that: In step S4: the number m of negative words in the original sentence is recorded. After extracting negative words from the original sentence, the negative words are analyzed in combination with the sentiment of the adjectives. Negative words that are odd in number and located before the adjectives can reverse the sentiment of the adjectives.

7. The method for evaluating word-of-mouth information based on data analysis according to claim 1, characterized in that: In step S5: The adjectives, degree adverbs, and negation words extracted from the original sentence are combined to analyze the positive evaluation level of the original sentence, and a scoring method for the positive evaluation level of word-of-mouth is established. The positive evaluation level score is calculated by superimposing the sentiment assignment of adjectives and the scores of degree adverbs at different levels. The positive evaluation level score of word-of-mouth is calculated using a formula: ； Where y represents the calculated positive rating of word-of-mouth, q represents the rating of the degree of emotional intensification by the adverb of degree, and n represents the number of adjectives. The sentiment assignments for different adjectives are represented by r, which is the sentiment assignment marker for the adjective. r = 1, 0, -1, and m represents the number of negative words in the original sentence.

Images