A quantitative method for evaluating the quality of online e-commerce reviews based on AHP-fuzzy comprehensive evaluation.

Abstract

The present invention relates to a quantitative method for the quality of online reviews of e-commerce companies based on the AHP-fuzzy comprehensive evaluation method, which belongs to the technical field of data analysis and mining. First, an online review quality evaluation framework is constructed, the framework includes several dimensions, each dimension includes several indicators; the dimensions and indicators are quantitatively calculated; the review quality levels are divided, and the weights of different indicators under the same dimension are determined to calculate the quantitative values ​​of each dimension; for the quantitative values ​​of each dimension, the weights of different dimensions are determined, and the fuzzy comprehensive evaluation method is combined to calculate the quality values ​​of the online reviews, and then the quality level evaluation of the reviews is obtained according to the corresponding relationship between the values ​​and the quality levels. The present invention constructs a relatively comprehensive online review quality evaluation system, integrates the quantitative values ​​of these dimensions and indicators, and calculates the quantitative representation of the quality of online reviews.

Description

Technical Field

[0001] The present invention relates to a method for quantifying the quality of online reviews of e-commerce companies based on an AHP-fuzzy comprehensive evaluation method, and belongs to the technical field of data analysis and mining. Background Art

[0002] Existing technologies mainly use natural language processing (NLP), machine learning, and deep learning to evaluate the quality of reviews. These technologies evaluate the quality of online reviews from multiple dimensions. For example, through sentiment analysis, topic extraction, grammar and semantic analysis, the usefulness, authenticity, completeness, and timeliness of reviews are evaluated.

[0003] Although existing related studies have proposed evaluation dimensions and indicators for online review quality, on the one hand, these dimensions and indicators are not comprehensive, and on the other hand, these studies do not provide specific quantitative evaluation methods. Most of the existing review quantification studies are also carried out on a single dimension or a few dimensions such as review sentiment and product features, and their final output results cannot objectively reflect the evaluation quality. Summary of the invention

[0004] In order to solve the above problems, the present invention discloses a method for quantifying the quality of online reviews of e-commerce companies based on AHP-fuzzy comprehensive evaluation method, and its specific technical scheme is as follows:

[0005] A quantitative method for the quality of online reviews of e-commerce companies based on AHP-fuzzy comprehensive evaluation method includes the following steps:

[0006] Step 1: Construct an online review quality evaluation framework, which includes several dimensions, and each dimension includes several indicators;

[0007] Step 2: Quantitatively calculate the dimensions and indicators, use natural language processing and mathematical statistics methods to calculate the quantitative values ​​of each indicator, and standardize the indicator values ​​in a targeted manner so that the value range of each indicator falls within the interval [0~1] and its value is proportional to the quality, so as to unify the measurement indicators and the quality of the comments;

[0008] Step 3: Use the fuzzy comprehensive evaluation method to first classify the review quality level, and then determine the weights of different indicators under the same dimension according to the characteristics of the indicators, so as to calculate the quantitative value of each dimension;

[0009] Step 4: For the quantitative value of each dimension, use the hierarchical analysis method to determine the weights of different dimensions, and then use the fuzzy comprehensive evaluation method to calculate the quality value of the online reviews. Then, based on the correspondence between the value and the quality level, the quality level evaluation of the reviews is obtained.

[0010] Furthermore, the dimensions in step 1 include credibility, usefulness, completeness, accuracy, freshness, user expertise, richness of user follow-up reviews, and merchant response acceptance.

[0011] Furthermore, the indicators of each dimension in step 1 are:

[0012] The indicators of credibility include objectivity and similarity;

[0013] The indicators of usefulness include useful word coverage, user recognition, sentiment intensity and useful tags;

[0014] The indicators of completeness include review length, number of pictures and product feature word coverage;

[0015] Accuracy includes indicators such as grammatical accuracy and text readability;

[0016] Freshness includes indicators such as the time interval between the current comment and the most recent comment, the time interval between placing an order and confirming receipt, the time interval between placing an order and publishing a comment, and the time interval between additional comments and the main comment.

[0017] The indicators of user professionalism include anonymous mark, registered mark and member mark;

[0018] The indicators of user review richness include review length, objectivity, emotional intensity, number of pictures and user recognition;

[0019] The indicators of merchant response acceptance include response length, objectivity, emotional intensity and merchant response recognition.

[0020] Furthermore, the specific method for quantitatively calculating each indicator in step 2 is:

[0021] (1) The quantitative calculation method of the credibility is:

[0022] ① Objectivity: Use TextBlob to load the comment text and call its subjectivity property. The output value range is [0~1], where 0 represents complete objectivity and 1 represents complete subjectivity. Subtract the subjectivity value from 1. The larger the value, the more objective it is.

[0023] ②Similarity: Use SpaCy to load the text and call the similarity method to get the similarity, which ranges from [0 to 1]; calculate the median similarity between the comment text and other comments in the same set, and then subtract the similarity value from 1. The result is proportional to the high quality;

[0024] (2) The quantitative calculation method of the usefulness is:

[0025] ① Useful word coverage: Build a useful word library and calculate the ratio of useful words in the comments, that is, the number of useful words divided by the length of the comment;

[0026] ② User recognition: Build a vocabulary of positive and negative words, and calculate the absolute value of the difference between the coverage rate of positive words and the coverage rate of negative words;

[0027] ③ Sentiment intensity: Use TextBlob to load the text, call the polarity attribute, and take the absolute value as the sentiment intensity, ranging from [-1 to 1];

[0028] ④ Useful Mark: Calculate the number of likes or useful marks, and use it as the final value after normalization;

[0029] (3) The quantitative calculation method of the integrity is:

[0030] ① Comment length: Use regular expressions to remove punctuation marks and count the number of words; set the length interval and convert the length value to the [0~1] interval. The larger the value, the higher the quality.

[0031] ②Number of images: Use regular expressions to identify the number of img tags in HTML; divide the interval and convert the number of images into a [0~1] interval. The larger the value, the higher the quality.

[0032] ③ Product feature word coverage: Build a feature word library and calculate the ratio of feature words in reviews; the calculation method is the same as the useful word coverage;

[0033] (4) The quantitative calculation method of the accuracy is:

[0034] ① Grammatical accuracy: Use LanguageTool to detect grammatical errors and count the number of errors; take the reciprocal as the accuracy value. If there are no grammatical errors, the value is 1;

[0035] ② Text readability: The readability is evaluated by combining the Fog Index (FOG), the Coleman-Liu Index (CLI), and the Automatic Readability Index (ARI). The arithmetic mean of the three indexes is taken as the readability value. The formula is used for statistical analysis to convert the value into a [0~1] interval. The larger the value, the higher the quality.

[0036] (5) The quantitative calculation method of the freshness is:

[0037] Convert the time interval between the previous comment and the latest comment, the time interval between placing an order and confirming receipt, the time interval between placing an order and publishing a comment, and the time interval between additional comments and the main comment into seconds, calculate the absolute value of the difference in seconds, and then convert it into days as the interval value between the two times; if the time interval is the same second, it is considered to be 0; the initial value is a real number greater than or equal to 0; since the longer the comment time, the lower the reading value; therefore, take the reciprocal of the interval as the freshness, the smaller the interval, the larger the value, and the higher the quality;

[0038] (6) The quantitative calculation method of user expertise is as follows:

[0039] According to the importance of user identities of anonymous tags, registered tags, and member tags, the importance weights of different identities are obtained, and then the weighted sum is normalized and finally the corresponding scores are obtained;

[0040] (7) The quantitative calculation method for the richness of user reviews is:

[0041] For the content of user follow-up comments, we calculated them one by one according to the indicators of follow-up comment length, objectivity, emotional intensity, number of pictures and user recognition, using the calculation method of the completeness comment length indicator, the credibility objectivity indicator, the usefulness emotional intensity indicator, the completeness picture number indicator and the usefulness user recognition indicator;

[0042] (8) The quantitative calculation method for the merchant response acceptance is:

[0043] For the merchant's reply content, the reply length, objectivity, emotional intensity and merchant reply recognition indicators are used to calculate the review length index of completeness, the objectivity index of credibility, the emotional intensity index of usefulness and the user recognition of usefulness. If there is no merchant reply, the dimension value is 0.

[0044] Furthermore, each indicator is assigned a weight, and a weighted average is performed according to the weight of the indicator to obtain the quantitative value of each dimension.

[0045] Furthermore, the calculation process of the comment quality level in step 3 is:

[0046] (1) Determine the evaluation factors and levels:

[0047] The evaluation factors include 8 dimensions, denoted as V=[V1, V2, ..., V8], and the quantitative value of each dimension is denoted as Vi. The quality of comments is divided into three levels: low, medium, and high, corresponding to the evaluation set T=[low, medium, high]. The quality value of each comment is calculated through the fuzzy relationship matrix and weight of the evaluation factors;

[0048] (2) Constructing the fuzzy relationship matrix:

[0049] The fuzzy comprehensive evaluation method uses membership functions to construct a fuzzy relationship matrix to express the degree of compliance of each dimension with the quality grade;

[0050] (3) Determination of evaluation factor weights based on the AHP method:

[0051] The AHP method is used to calculate the weights of each dimension. The steps are as follows:

[0052] ① Establish a hierarchical model: online review quality is the target layer, and the eight dimensions are the criterion layer;

[0053] ② Construct a judgment matrix: Use the 1-9 scaling method to compare the dimensions of the criterion layer in pairs to form a judgment matrix;

[0054] ③ Consistency test: Calculate the eigenvector and maximum eigenroot of the matrix and perform consistency test. After passing the test, the weight vector w is finally obtained, which is used as the weight of each dimension.

[0055] Furthermore, the step 4 calculates the fuzzy comprehensive evaluation set for quality evaluation as follows:

[0056] Multiply the fuzzy relationship matrix U by the eigenvector w to obtain the fuzzy comprehensive evaluation set S:

[0057]

[0058] in, , and Respectively represent The three quality level values ​​of low, medium and high in the fuzzy relationship matrix of the dimension are The The weight value of each dimension, 1≤ ≤8;

[0059] The fuzzy comprehensive evaluation set S has three values, corresponding to the three quality levels of low, medium and high. According to the rules of fuzzy comprehensive evaluation, the maximum value of these three values ​​is used as the quantitative representation of the quality of online reviews, and the quality level corresponding to the maximum value is the qualitative representation of the quality of online reviews.

[0060] Furthermore, the trapezoidal function is selected as the membership function to construct the fuzzy relationship matrix. The trapezoidal function is a piecewise function divided into small, medium and large sizes. Each segment of the piecewise function represents the quality level of low, medium and high, as follows:

[0061] The trapezoidal function formula for the small segment is: ,

[0062] The trapezoidal function formula of the intermediate segment is: ,

[0063] The trapezoidal function formula for the large segment is: ,

[0064] Among them, the independent variable That is, each dimension The quantitative value of ≤8, and the determination of the parameter values ​​of a, b, c, and d in each piecewise function needs to be adjusted according to the specific dimension value so that the adjusted dimension value falls between 0 and 1; and these three-segment divisions are the basic rules of the fuzzy comprehensive evaluation method. According to this rule, three values ​​will eventually be obtained, and the grade corresponding to the largest value is the corresponding quality; if the maximum value appears in a small size, then the quality of the review is low quality, if the maximum value appears in an intermediate size, then the quality of the review is medium quality, and if the maximum value appears in a large size, then the quality of the review is high quality.

[0065] The beneficial effects of the present invention are:

[0066] The present invention constructs a relatively comprehensive online review quality evaluation system, which includes relatively comprehensive dimensions and indicators.

[0067] The present invention provides a detailed quantitative calculation method for dimensions and indicators, integrates the quantitative values ​​of these dimensions and indicators, and calculates a quantitative representation of the quality of online reviews, which is more advanced than existing evaluation and analysis methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a framework diagram of online review quality evaluation of the present invention,

[0069] Figure 2 is the online review quality evaluation framework diagram of the present invention,

[0070] Figure 3 It is the trapezoidal membership function formula diagram of the online review quality of the present invention,

[0071] Figure 4 4 is a distribution diagram of manual evaluation results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0073] 1. This paper constructs an online review quality evaluation framework, which includes 8 dimensions and 27 indicators, such as Figure 1 shown.

[0074] The present invention aims to evaluate the quality of online reviews by quantitatively calculating indicators and dimensions. This is a multi-dimensional and multi-indicator quantitative calculation problem. Since the quality of the review text itself has a certain degree of ambiguity, it is an object that is difficult to quantify. Therefore, the present invention combines AHP and fuzzy comprehensive evaluation method to quantitatively calculate the quality of online reviews. The overall structure is as follows: Figure 2 shown.

[0075] The present invention uses natural language processing and mathematical statistics methods to calculate the quantitative value of each indicator in order to eliminate the quantitative differences in the indicator values ​​between different methods. The indicator values ​​will be standardized in a targeted manner so that the value range of each indicator value falls within the interval [0~1], and its value is proportional to the quality level, so as to unify the measurement indicators and the quality of the comments. Secondly, the fuzzy comprehensive evaluation method is mainly used to first divide the comment quality level, and then the expert survey method is applied according to the characteristics of the indicator to determine the weights of different indicators under the same dimension, so as to calculate the quantitative value of each dimension. For each dimension value, the hierarchical analysis method is used to determine the weights of different dimensions, and then the fuzzy comprehensive evaluation method is combined to calculate the quality value of the online comments, and then the quality level evaluation of the comments is obtained according to the corresponding relationship between the value and the level.

[0076] 2. Quantitative calculation method of each indicator value

[0077] (1) Credibility

[0078] The credibility dimension mainly measures whether the comments are true and reliable, and its evaluation indicators are: objectivity and similarity.

[0079] ① Objectivity: Use TextBlob to load the comment text and call its subjectivity property. The output value range is [0~1], where 0 represents complete objectivity and 1 represents complete subjectivity. Subtract the subjectivity value from 1. The larger the value, the more objective it is.

[0080] Code:

[0081] def get_objectivity(text):

[0082] if text == "This user have not give any comment !":

[0083] return 0

[0084] blob = TextBlob(text)

[0085] objectivity = 1 - blob.sentiment.subjectivity

[0086] return objectivity.

[0087] ②Similarity: Use SpaCy to load the text and call the similarity method to get the similarity, which ranges from [0 to 1]. Calculate the median similarity between the comment text and other comments in the same set, and then subtract the similarity value from 1. The result is proportional to the high quality.

[0088] Key code:

[0089] Feature extraction:

[0090] def feature_extraction(texts):

[0091] vectorizer = TfidfVectorizer(tokenizer=text_preprocess, token_pattern=None)

[0092] X = vectorizer.fit_transform(texts)

[0093] return X.toarray()

[0094] Similarity calculation:

[0095] def similarity_calculation(text1, text2):

[0096] X = feature_extraction([text1, text2])

[0097] similarity = cosine_similarity(X[0].reshape(1, -1), X[1].reshape(1, -1))

[0098] return similarity[0][0]

[0099] Similarity matrix construction:

[0100] for i in range(start_row, end_row):

[0101] for j in range(start_row, i + 1):

[0102] text1 = df.iloc[i, df.columns.get_loc('Comment content')]

[0103] text2 = df.iloc[j, df.columns.get_loc('Comment content')]

[0104] similarity = similarity_calculation(text1, text2)

[0105] similarities_matrix[i - start_row, j - start_row] =similarity

[0106] similarities_matrix = similarities_matrix + similarities_matrix.T -np.diag(np.diag(similarities_matrix))

[0107] (2) Usefulness

[0108] The usefulness dimension mainly measures whether the review content is useful to potential users. The evaluation indicators are: useful word coverage, user recognition, sentiment intensity and useful tags.

[0109] ① Useful word coverage: Build a useful word library and calculate the ratio of useful words in the comments, that is, the number of useful words divided by the length of the comment.

[0110] Key code:

[0111] (1) Text segmentation and stop word filtering

[0112] for text in text_column:

[0113] if isinstance(text, str):

[0114] words = [word.strip() for part in text.replace(';', ',').replace(',', ',').split(',') for word in part.split()]

[0115] words = [word for word in words if word not in stopwords]

[0116] word_set.update(words)

[0117] (2) Calculate the coverage of “useful words”

[0118] def calculate_ratio(txt_file_path, comments):

[0119] txt_words = set(jieba.lcut(open(txt_file_path, 'r', encoding='utf-8').read()))

[0120] word_freq = {}

[0121] for word in comments:

[0122] if word in txt_words:

[0123] word_freq[word] = word_freq.get(word, 0) + 1

[0124] return len(word_freq) / len(comments) if len(comments) != 0 else0

[0125] ② User recognition: Build a vocabulary of positive and negative words, and calculate the absolute value of the difference between the positive word coverage and the negative word coverage.

[0126] Key code:

[0127] (1) Word segmentation and stop word filtering

[0128] for text in text_column:

[0129] if isinstance(text, str):

[0130] words = [word.strip() for part in text.replace(';', ',').replace(',', ',').split(',') for word in part.split()]

[0131] words = [word for word in words if word not in stopwords]

[0132] word_set.update(words)

[0133] (2) Calculate coverage

[0134] df['User recognition of useful words'] = df.apply(lambda row: row[column_a] - row[column_b] if row[column_b] != 0 else row[column_a], axis=1)

[0135] ③Sentiment intensity: Use TextBlob to load text, call the polarity attribute, and take the absolute value as the sentiment intensity, ranging from [-1 to 1].

[0136] Key code:

[0137] s = SnowNLP(text)

[0138] polarity = s.sentiments

[0139] return polarity

[0140] ④ Useful Mark: Calculate the number of times the post is liked or marked as useful, and normalize it as the final value.

[0141] Key code:

[0142] df['usefulness statistics'] = df[column_name].apply(lambda x: 1 if x == Trueelse 0)

[0143] (3) Integrity

[0144] The completeness dimension mainly measures whether the review content is rich and whether it provides the content that potential users want. The evaluation indicators are: review length, number of pictures, and coverage of product feature words.

[0145] ① Comment length: Use regular expressions to remove punctuation marks and count the number of words. Set the length interval and convert the length value to the [0~1] interval. The larger the value, the higher the quality.

[0146] Key code:

[0147] df['comment length'] = df['comment content'].apply(lambda x: len(x.replace(' ','')) if isinstance(x, str) else 0)

[0148] ②Number of images: Use regular expressions to identify the number of img tags in HTML. Divide the interval and convert the number of images into a [0~1] interval. The larger the value, the higher the quality.

[0149] Key code:

[0150] df['Number of pictures'] = df['Comment pictures'].apply(count_commas_in_csv_column)

[0151] ③ Product feature word coverage: Build a feature word library and calculate the ratio of feature words in reviews. The calculation method is the same as the useful word coverage.

[0152] Key code:

[0153] (1) Segment each text and add it to the vocabulary, while removing stop words

[0154] for text in text_column:

[0155] if isinstance(text, str):

[0156] words = [word.strip() for part in text.replace(';', ',').replace(',', ',').split(',') for word in part.split()]

[0157] words = [word for word in words if word not in stopwords]

[0158] word_set.update(words)

[0159] (2) Calculate coverage

[0160] def calculate_ratios_and_write(csv_file_path, txt_file_path, column_name):

[0161] df = pd.read_csv(csv_file_path, encoding='utf_8_sig')

[0162] ratios = []

[0163] with open(txt_file_path, 'r', encoding='utf-8') as txt_file:

[0164] txt_words = set(jieba.lcut(txt_file.read()))

[0165] for comment in df[column_name]:

[0166] if comment != 'This user has not filled in a comment!':

[0167] comments = jieba.lcut(comment)

[0168] ratio = calculate_ratio(txt_file_path, comments)

[0169] ratios.append(ratio)

[0170] else:

[0171] ratios.append(0)

[0172] df['Character word coverage of similar products'] = ratios

[0173] df.to_csv(csv_file_path, index=False, encoding='utf_8_sig')

[0174] (4) Accuracy

[0175] The accuracy dimension mainly measures the organization of the review content, and its evaluation indicators are: grammatical accuracy and text readability.

[0176] ①Grammar accuracy: Use LanguageTool to detect grammar errors and count the number of errors. The reciprocal is taken as the accuracy value. If there are no grammar errors, the value is 1.

[0177] Key code:

[0178] def check_grammar(text_list):

[0179] results = []

[0180] for text in text_list:

[0181] matches = tool.check(text)

[0182] results.append(len(matches))

[0183] time.sleep(5)

[0184] return results

[0185] ② Text readability: The Fog Index (FOG), Coleman-Liu Index (CLI) and Automatic Readability Index (ARI) are used to evaluate readability, and the arithmetic mean of the three indexes is taken as the readability value. The formula is used for statistical analysis to convert the value into the [0~1] interval. The larger the value, the higher the quality.

[0186] Key code:

[0187] def get_language_feature(text):

[0188] feature_result = []

[0189] for index in text:

[0190] fog_num = ts.gunning_fog(index)

[0191] ari_num = ts.automated_readability_index(index)

[0192] cli_num = ts.coleman_liau_index(index)

[0193] feature_num = (fog_num + ari_num + cli_num) / 3

[0194] feature_result.append(feature_num)

[0195] return feature_result

[0196] (5) Freshness

[0197] The freshness dimension mainly evaluates the time interval between comment publishing, and its evaluation indicators are: the time interval between the current comment and the most recent comment, the time interval between placing an order and confirming receipt, the time interval between placing an order and comment publishing, and the time interval between additional comments and the main comment.

[0198] The calculation rule is: convert the two times into seconds, calculate the absolute value of the difference in seconds, and then convert it into days as the interval value of the two times. For the same second, it is regarded as 0. The initial value is a real number greater than or equal to 0. Since the longer the comment time is, the lower the value of reading can be considered. Therefore, the reciprocal of the interval is taken as the freshness. The smaller the interval is, the larger the value is, and the higher the quality is.

[0199] Implementation method: Calculate the difference between two times and take the reciprocal as the freshness key code:

[0200] selected_rows = df.loc[start_row:end_row - 1]

[0201] max_comment_time = selected_rows['Comment time'].max()

[0202] selected_rows['Comment time interval'] = max_comment_time - selected_rows['Comment time']

[0203] selected_rows['Comment time interval'] = selected_rows['Comment time interval'] / pd.Timedelta(days=1)

[0204] Implementation method: Calculate the difference between two times and take the reciprocal as the freshness key code:

[0205] with open('D: / Acode / Quantitative Analysis / Time Interval / time.csv', 'r', encoding='utf-8') as csvfile:

[0206] csvreader = csv.DictReader(csvfile)

[0207] time_intervals = []

[0208] for row in csvreader:

[0209] start_time_data = ast.literal_eval(row['Transaction start time'])

[0210] end_time_data = ast.literal_eval(row['Transaction end time'])

[0211] start_time = start_time_data.get('time')

[0212] end_time = end_time_data.get('time')

[0213] start_time = datetime.fromtimestamp(start_time / 1000)

[0214] end_time = datetime.fromtimestamp(end_time / 1000)

[0215] time_interval = (end_time - start_time) / pd.Timedelta('1day')

[0216] absolute_time_interval = abs(time_interval)

[0217] time_intervals.append(absolute_time_interval)

[0218] Implementation method: Calculate the difference between two times and take the reciprocal as the freshness key code:

[0219] for index, row in df.iterrows():

[0220] start_time_data = eval(row['Transaction start time'])

[0221] start_time = start_time_data.get('time')

[0222] start_time = datetime.fromtimestamp(start_time / 1000)

[0223] time_start.append(start_time)

[0224] comment_times = []

[0225] for index, row in df.iterrows():

[0226] comment_time_str = row['Comment time']

[0227] comment_time = pd.to_datetime(comment_time_str)

[0228] comment_times.append(comment_time.to_pydatetime())

[0229] time_diff = [abs((comment - start).total_seconds()) / (24 * 60 * 60)for comment, start in zip(comment_times, time_start)]

[0230] df['Transaction time and comment publishing time interval'] = time_diff

[0231] Implementation method: Calculate the difference between two times and take the reciprocal as the freshness key code:

[0232] df["Add Comment Time"] = pd.to_datetime(df["Add Comment Time"])

[0233] df["Comment Time"] = pd.to_datetime(df["Comment Time"])

[0234] df["Time interval between additional comments and main comments"] = (df["Additional comment time"] - df["Comment time"]) / pd.Timedelta('1 day')

[0235] df["Time interval between additional comments and main comments"].fillna(0, inplace=True)

[0236] (6) User professionalism

[0237] The user professionalism dimension mainly measures the differences in comments posted by users of different levels, and its evaluation indicators are: anonymous mark, registered mark and member mark.

[0238] The calculation rule is: according to the results of the questionnaire on the importance of user identities, the importance weights of different identities are obtained, and then the weighted sum is normalized. Finally, the corresponding scores are: anonymous users 0.0317, registered users 0, and member users 0.3837. In the actual calculation, if a certain mark appears, the corresponding score is taken, and if not, it is counted as 0.

[0239] Implementation method: Use questionnaires to assign weights to user identity tags (anonymous tags, registered tags, member tags) and calculate them, combined with operations on the data frame pandas.

[0240] Key code:

[0241] def calculate_score(row):

[0242] if row['Member Statistics'] == 0 and row['Anonymous Statistics'] == 0:

[0243] return row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535

[0244] elif row['Member Statistics'] == 1 and row['Anonymous Statistics'] == 1:

[0245] return (row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535) / 3

[0246] else:

[0247] return (row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535) / 2

[0248] df['User identity score'] = df.apply(calculate_score, axis=1)

[0249] (7) Richness of user reviews

[0250] The richness of user follow-up reviews mainly measures user satisfaction with the product from the perspective of user additional reviews. This dimension is a supplement to the main review. The richer the content, the more it reflects the user's satisfaction with the purchase, and the greater the reference value for potential users. Therefore, it is also part of the quality of the main review. The evaluation indicators are: follow-up review length, objectivity, emotional intensity, number of pictures, and user recognition. If there is no follow-up review, the value of this dimension is 0.

[0251] Length of follow-up review

[0252] Implementation method: Use len() to calculate the length of a character.

[0253] Key code:

[0254] def get_review_length(text):

[0255] review_length = []

[0256] for item in text:

[0257] if pd.notna(item) and isinstance(item, str):

[0258] review_length.append(len(item.replace(' ', '')))

[0259] else:

[0260] review_length.append(0)

[0261] return review_length

[0262] Emotional intensity

[0263] Implementation: Use the sentiments method of the SnowNLP library to calculate the sentiment polarity of each comment. The output range of sentiments is [0, 1]: the closer to 1, the more positive it is. The closer to 0, the more negative it is. Key code:

[0264] def get_sentiment(text1):

[0265] text = str(text1)

[0266] text = " ".join(seg_list)

[0267] s = SnowNLP(text)

[0268] return polarity

[0269] Number of pictures

[0270] Implementation method: Define a function that receives a text as input, converts it to a string, then counts the number of commas in the string and returns the number of commas. Key code:

[0271] def count_commas_in_csv_column(text1):

[0272] text = str(text1)

[0273] return text.count(',')

[0274] User recognition

[0275] Implementation method: (1) Calculate the ratio of bad words to good words in each comment. The bad words and good words are stored in a txt file. (2) The recognition is measured by calculating the difference between the good word rate and the bad word rate in the comment. Key code:

[0276] df['Additional comment user approval'] = df.apply(lambda row: row[column_a] - row[column_b] if row[column_b] != 0 else row[column_a], axis=1)

[0277] objectivity

[0278] Implementation: Use the sentiments method of the SnowNLP library to calculate the sentiment polarity of each comment. The output range of sentiments is [0, 1]: the closer to 1, the more positive it is. The closer to 0, the more negative it is. Key code:

[0279] def get_sentiment(text1):

[0280] text = str(text1)

[0281] seg_list = jieba.cut(text, cut_all=False)

[0282] text = " ".join(seg_list)

[0283] s = SnowNLP(text)

[0284] polarity = s.sentiments

[0285] return polarity

[0286] (8) Merchant response acceptance

[0287] Merchant response acceptance measures the user's acceptance of the merchant's response, which serves as a supplement to the quality of the main review. The following indicators are set: response length, objectivity, emotional intensity, and merchant response acceptance. Merchant response acceptance is reflected in the merchant's response content, and the calculation method is the same as user acceptance. If there is no merchant response, the dimension value is 0.

[0288] 3. Quantitative calculation methods for each dimension

[0289] (1) Determine the weight of indicators in each dimension

[0290] In order to calculate the quantitative value of each dimension, all calculated indicators need to be normalized first, and the indicator values ​​are unified to the range of [0,1]. Then, the normalized values ​​are weighted averaged according to the weight of each indicator to obtain the final dimension quantitative value.

[0291] Key code:

[0292] (1) Credibility Dimension

[0293] df['credibility dimension mean'] = df['objectivity'] * 0.9302 + df['similarity mean'] * 0.0698

[0294] (2) Usefulness

[0295] df['usefulness dimension'] = df['useful word coverage'] * 0.3023 + df['user recognition'] * 0.5349 + df['usefulness statistics'] * 0.0581 + df['polarity'] * 0.1047

[0296] (3) Completeness Dimension

[0297] df['Completeness dimension'] = df['Comment length'] * 0.2209 + df['Feature word coverage of similar products'] * 0.5465 + df['Number of pictures'] * 0.2326

[0298] (4) Accuracy Dimension

[0299] df['Accuracy dimension'] = df['Grammar accuracy'] * 0.3488 + df['Text readability'] * 0.6512

[0300] (5) Time dimension:

[0301] df['time dimension'] = df['comment interval'] * 0.2209 + df['time interval between transaction start and end'] * 0.1512 + df['time interval between transaction time and comment release'] * 0.2674 + df['interval between additional comments and main comments'] * 0.3605

[0302] (6) User identity dimension

[0303] def calculate_score(row):

[0304] if row['Member Statistics'] == 0 and row['Anonymous Statistics'] == 0:

[0305] return row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535

[0306] elif row['Member Statistics'] == 1 and row['Anonymous Statistics'] == 1:

[0307] return (row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535) / 3

[0308] else:

[0309] return (row['member statistics'] * 1*0.3837 + row['anonymous statistics'] *0.031746032 *0.4535) / 2

[0310] (7) Add comment dimension

[0311] df['Additional comment dimension'] = df['Additional content length'] * 0.1395 + df['Additional comment user recognition'] * 0.2442 + df['Additional content objectivity'] * 0.407 + df['Additional image number'] * 0.093 + df['Additional comment polarity'] * 0.1163

[0312] (8) Merchant response dimensions

[0313] df['Merchant response dimension'] = df['Merchant response length'] * 0.1163 + df['Merchant response user recognition'] * 0.3605 +df['Merchant response objectivity'] * 0.4651 + df['Merchant response polarity'] * 0.0581

[0314] The present invention determines the weights of various indicators under different dimensions, as shown in Table 1.

[0315] Table 1 Weight of each indicator

[0316]

[0317] (2) Calculation of quantized values ​​of each dimension

[0318] According to the weights shown in Table 1, the indicator values ​​are weighted averaged by dimension to obtain the quantitative value of each dimension.

[0319] 4. Online review quality evaluation based on AHP-fuzzy comprehensive evaluation method

[0320] The present invention uses the fuzzy comprehensive evaluation method to integrate the values ​​of each dimension to calculate the final review quality, specifically:

[0321] (1) Determine the evaluation factors and levels

[0322] The evaluation factors of online review quality include the aforementioned 8 dimensions, denoted as V=[V1, V2, ..., V8], and the quantitative value of each dimension is denoted as Vi. Review quality is divided into three levels: low, medium, and high, corresponding to the evaluation set T=[low, medium, high]. The quality value of each review is calculated through the fuzzy relationship matrix and weight of the evaluation factors.

[0323] (2) Constructing the fuzzy relationship matrix

[0324] The fuzzy comprehensive evaluation method uses the membership function to construct a fuzzy relationship matrix to indicate the degree of compliance of each dimension with the quality grade. The present invention selects the trapezoidal function as the membership function. The specific formula is as follows: Figure 3 As shown, the independent variable x is the dimension The quantitative value of ≤8. According to this function, each dimension of the review will get three membership degrees, and finally form a fuzzy relationship matrix U with 8 rows and 3 columns, as shown in formula (1).

[0325] (1)

[0326] (3) Determination of evaluation factor weights based on the analytic hierarchy process (AHP)

[0327] The AHP method is used to calculate the weights of each dimension. The steps are as follows:

[0328] ① Establish a hierarchical model: the quality of online reviews is the target layer, and the eight dimensions are the criterion layer.

[0329] ② Construct a judgment matrix: Use the 1-9 scaling method to compare the dimensions of the criterion layer in pairs to form a judgment matrix, see Table 2.

[0330] Table 2 Importance judgment matrix

[0331]

[0332] ③ Consistency test: Calculate the eigenvector and maximum eigenroot of the matrix and perform consistency test. After passing the test, the weight vector w is finally obtained, which is used as the weight of each dimension.

[0333] (2)

[0334] Normalize the indicators of each dimension (such as through Min-Max normalization) so that all indicator values ​​are in the range of [0, 1]. This is to ensure that the scores of each dimension are compared on the same scale.

[0335] First, a discriminant matrix is ​​constructed, and the eigenvalues ​​and eigenvectors are used to calculate the weight of each dimension. The code in this section already includes the calculation of the eigenvector and the consistency check (CR < 0.1). If the consistency check passes, the weight of each dimension can be obtained.

[0336] Key code:

[0337] b = np.array([

[0338] [1, 3, 5, 3, 6, 7, 6, 8],

[0339] [1 / 3, 1, 4, 2, 4, 5, 4, 7],

[0340] [1 / 5, 1 / 4, 1, 1 / 3, 2, 3, 2, 3],

[0341] [1 / 3, 1 / 2, 3, 1, 4, 4, 3, 6],

[0342] [1 / 6, 1 / 4, 1 / 2, 1 / 4, 1, 2, 2, 3],

[0343] [1 / 7, 1 / 5, 1 / 3, 1 / 4, 1 / 2, 1, 1 / 2, 2],

[0344] [1 / 6, 1 / 4, 1 / 2, 1 / 3, 1 / 2, 2, 1, 3],

[0345] [1 / 8, 1 / 7, 1 / 3, 1 / 6, 1 / 3, 1 / 2, 1 / 3, 1] ])

[0347] a=np.transpose(b)

[0348] w = np.linalg.eig(a)

[0349] tzz = np.max(w[0])

[0350] t = np.argwhere(w[0] == tzz)

[0351] tzx = w[1][::-1, t[0]]

[0352] RILIST = [0, 0, 0.58, 0.9, 1.12, 1.24, 1.32, 1.41, 1.45, 1.49, 1.52,1.54, 1.56, 1.58, 1.59]

[0353] n = a.shape[0]

[0354] RI = RILIST[n]

[0355] CI = (tzz - n) / (n - 1)

[0356] print(float(CI))

[0357] CR = CI / RI

[0358] print(CR)

[0359] print("Passed consistency check") if CR < 0.1 else print("Failed consistency check")

[0360] Normalize the indicators of each dimension (such as through Min-Max normalization) so that all indicator values ​​are in the range of [0, 1]. This is to ensure that the scores of each dimension are compared on the same scale.

[0361] Key code:

[0362] columns_to_normalize = ['Credibility dimension mean', 'Usefulness dimension', 'Completeness dimension', 'Accuracy dimension', 'Time dimension', 'User identity dimension', 'Additional comment dimension', 'Merchant response dimension']

[0363] scaler = MinMaxScaler()

[0364] df[columns_to_normalize] = scaler.fit_transform(df[columns_to_normalize])

[0365] (4) Calculate the fuzzy comprehensive evaluation set for quality evaluation

[0366] Multiply the fuzzy relationship matrix U by the eigenvector w to obtain the fuzzy comprehensive evaluation set S:

[0367] (3)

[0368] in, and is a value in formula (1) and formula (2), 1≤ ≤8. This fuzzy comprehensive evaluation set has three values, corresponding to the three quality levels of low, medium and high. According to the rules of fuzzy comprehensive evaluation, the maximum value among these three values ​​is used as the quantitative representation of the quality of online reviews, and the quality level corresponding to the maximum value is the qualitative representation of the quality of online reviews.

[0369] The trapezoidal function is selected as the membership function to construct the fuzzy relationship matrix. The trapezoidal function is a piecewise function divided into small, medium and large parts. Figure 3 shown.

[0370] The degree of membership of each dimension is calculated using the trapezoidal fuzzy set method. The degree of membership indicates the degree of fuzzy belonging of a comment on each dimension. Different types of trapezoidal fuzzy set calculation methods are used, namely, small trapezoidal fuzzy sets, intermediate trapezoidal fuzzy sets, and large trapezoidal fuzzy sets.

[0371] Key code:

[0372] def calculate_memberships(value_to_fuzzify, params):

[0373] a, b, c, d = params

[0374] def skewed_trapezoid_membership_small(x, a, b):

[0375] if x <= a:

[0376] return 1

[0377] elif a < x <= b:

[0378] return (b - x) / (b - a)

[0379] else:

[0380] return 0.0

[0381] def trapezoid_membership_medium(x, a, b, c, d):

[0382] if a <= x <= b:

[0383] return (x - a) / (b - a)

[0384] elif b <= x <= c:

[0385] return 1.0

[0386] elif c <= x <= d:

[0387] return (d - x) / (d - c)

[0388] else:

[0389] return 0.0

[0390] def skewed_trapezoid_membership_large(x, c, d):

[0391] if x < c:

[0392] return 0.0

[0393] elif c <= x <= d:

[0394] return (x - c) / (d - c)

[0395] else:

[0396] return 1.0

[0397] membership_results = []

[0398] for val in value_to_fuzzify:

[0399] membership_small = skewed_trapezoid_membership_small(val, a,b)

[0400] membership_medium = trapezoid_membership_medium(val, a, b, c,d)

[0401] membership_large = skewed_trapezoid_membership_large(val, c,d)

[0402] membership_results.append([membership_small, membership_medium, membership_large])

[0403] return membership_results

[0404] Calculate dimension merging results

[0405] In the fuzzy comprehensive evaluation, the system uses the dot multiplication method to calculate the comprehensive score of the review in each dimension. Through matrix dot multiplication, the membership degree of each review in each dimension is multiplied by the pre-set weight vector to obtain the comprehensive score.

[0406] Key code:

[0407] for index, row in data.iterrows():

[0408] row_values ​​= row.values.tolist()

[0409] value_to_fuzzify = row_values[:8]

[0410] row_results = []

[0411] for i in range(len(value_to_fuzzify)):

[0412] membership_results = calculate_memberships([value_to_fuzzify[i]], dimensions[i])

[0413] row_results.extend(membership_results)

[0414] all_results.append(row_results)

[0415] result_matrix = np.array(all_results)

[0416] print("Result matrix:")

[0417] print(result_matrix)

[0418] vector_to_dot = np.array([[0.30878385+0.j, 0.14794154+0.j,0.20889876+0.j, 0.12941033+0.j, 0.04869789+0.j, 0.09734021+0.j, 0.03679227+0.j,0.02213515+0.j]])

[0419] dot_product = np.dot(vector_to_dot, result_matrix)

[0420] print("Dot product result of matrix:")

[0421] print(dot_product)

[0422] dot_product_real = dot_product.real

[0423] print("Real part result of dot product of matrix:")

[0424] print(dot_product_real)

[0425] dot_product_2d = np.squeeze(dot_product_real, axis=0)

[0426] result_df = pd.DataFrame(dot_product_2d, columns=['small size', 'intermediate size', 'large size'])

[0427] An application embodiment of the present invention and its verification process are given below:

[0428] The inventor team applied the method of the present invention to rate 21,693 reviews collected from Taobao (the data is public data from Taobao, and the source is legal and compliant). There were 6,565 reviews with general evaluation results, accounting for 30.2%; 13,425 reviews with good evaluation results, accounting for 61.8%; and 1,702 reviews with excellent evaluation results, accounting for 7.8%. It showed a distribution with small ends and a large middle, which is a normal distribution and conforms to the general rules of online reviews. In order to verify the accuracy of the present invention, manual labeling was first adopted. 10 undergraduates and graduate students with rich online shopping experience were invited as experts in the field. 5,000 reviews were randomly selected and the quality of the reviews was scored from 0 to 10. 0 represents the lowest quality and 10 represents the worst quality. The scores are based on the good or bad first impression of the individual. After scoring and statistics, the final results are as follows: Figure 4 shown.

[0429] like Figure 4 , the overall result is a left-skewed distribution, that is, in the evaluation results, there are more low values ​​than high values, which means that the evaluators generally believe that the quality of this batch of comments is not high. This was also confirmed by the process consultation of the evaluators. On this basis, after comparing the score range and the proportion of comments, the comments with a score between 0 and 1.5 are divided into low quality, 1.5 to 8 are medium quality, and 8 or above are high quality. The experiment assumes that manual evaluation is the closest to the truth, and then the quality results calculated by the present invention are used as the degree of conformity with the manual evaluation results as an evaluation index to measure the accuracy of the method. The calculation method is: among the comments of different quality levels obtained by statistical manual evaluation, the proportion of comments with the same quality level obtained by the model evaluation method in all comments. After calculation and comparison, the conformity of the present invention reached 64.9%, which is about 15% higher than the average value. These results show that the online comment quality evaluation model constructed by the present invention can accurately reflect the actual quality of the comments to a large extent and has a high reference value.

[0430] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as generally understood by those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0431] With the above-mentioned ideal embodiments of the present invention as inspiration, through the above-mentioned description content, relevant personnel can make various changes and modifications without departing from the scope of the technical idea of ​​the present invention.

Claims

1. A quantitative method for the quality of online reviews of e-commerce companies based on AHP-fuzzy comprehensive evaluation method, characterized in that: The following steps are involved: Step 1: Construct an online review quality evaluation framework, which includes several dimensions, each of which includes several indicators; the dimensions include credibility, usefulness, completeness, accuracy, freshness, user expertise, user follow-up review richness and merchant response acceptance; The indicators of each dimension are as follows: The indicators of credibility include objectivity and similarity; The indicators of usefulness include useful word coverage, user recognition, sentiment intensity and useful tags; The indicators of completeness include review length, number of pictures and product feature word coverage; Accuracy includes indicators such as grammatical accuracy and text readability; Freshness includes indicators such as the time interval between the current comment and the most recent comment, the time interval between placing an order and confirming receipt, the time interval between placing an order and publishing a comment, and the time interval between additional comments and the main comment. The indicators of user professionalism include anonymous mark, registered mark and member mark; The indicators of user review richness include review length, objectivity, emotional intensity, number of pictures and user recognition; The indicators of merchant response acceptance include response length, objectivity, emotional intensity and merchant response recognition; Step 2: Quantitatively calculate the dimensions and indicators, use natural language processing and mathematical statistics methods to calculate the quantitative value of each indicator, and standardize the indicator values ​​in a targeted manner so that the value range of each indicator falls within the interval [0~1] and its value is proportional to the quality, so as to unify the measurement indicators and the quality of the comments; the specific method for quantitative calculation of each indicator is: (1) The quantitative calculation method of the credibility is: ① Objectivity: Use TextBlob to load the comment text and call its subjectivity property. The output value range is [0~1], where 0 represents complete objectivity and 1 represents complete subjectivity. Subtract the subjectivity value from 1. The larger the value, the more objective it is. ②Similarity: Use SpaCy to load the text and call the similarity method to get the similarity, which ranges from [0 to 1]; calculate the median similarity between the comment text and other comments in the same set, and then subtract the similarity value from 1. The result is proportional to the high quality; (2) The quantitative calculation method of the usefulness is: ① Useful word coverage: Build a useful word library and calculate the ratio of useful words in the comments, that is, the number of useful words divided by the length of the comment; ② User recognition: Build a vocabulary of positive and negative words, and calculate the absolute value of the difference between the coverage rate of positive words and the coverage rate of negative words; ③ Sentiment intensity: Use TextBlob to load the text, call the polarity attribute, and take the absolute value as the sentiment intensity, ranging from [-1 to 1]; ④ Useful Mark: Calculate the number of likes or useful marks, and use it as the final value after normalization; (3) The quantitative calculation method of the integrity is: ① Comment length: Use regular expressions to remove punctuation and count the number of words; Set the length interval and convert the length value to the [0~1] interval. The larger the value, the higher the quality. ②Number of images: Use regular expressions to identify the number of img tags in HTML; Divide the interval and convert the number of pictures into the interval [0~1]. The larger the value, the higher the quality. ③ Product feature word coverage: Build a feature word library and calculate the ratio of feature words in reviews; the calculation method is the same as the useful word coverage; (4) The quantitative calculation method of the accuracy is: ① Grammatical accuracy: Use LanguageTool to detect grammatical errors and count the number of errors; take the reciprocal as the accuracy value. If there are no grammatical errors, the value is 1; ② Text readability: The readability is evaluated by combining the Fog Index (FOG), the Coleman-Liu Index (CLI), and the Automatic Readability Index (ARI), and the arithmetic mean of the three indexes is taken as the readability value; Use the formula for statistical analysis to convert the value into the [0~1] interval. The larger the value, the higher the quality. (5) The quantitative calculation method of the freshness is: Convert the time interval between the previous comment and the latest comment, the time interval between placing an order and confirming receipt, the time interval between placing an order and publishing a comment, and the time interval between additional comments and the main comment into seconds, calculate the absolute value of the difference in seconds, and then convert it into days as the interval value between the two times; if they are the same second, they are regarded as 0; the initial value is a real number greater than or equal to 0; take the reciprocal of the interval as the freshness, the smaller the interval, the larger the value, and the higher the quality; (6) The quantitative calculation method of user expertise is: According to the importance of user identities of anonymous tags, registered tags, and member tags, the importance weights of different identities are obtained, and then weighted summation and normalization are performed, and finally, corresponding scores are obtained; (7) The quantitative calculation method for the richness of user reviews is: For the content of user follow-up comments, we calculated them one by one according to the indicators of follow-up comment length, objectivity, emotional intensity, number of pictures and user recognition, using the calculation method of the completeness comment length indicator, the credibility objectivity indicator, the usefulness emotional intensity indicator, the completeness picture number indicator and the usefulness user recognition indicator; (8) The quantitative calculation method for the merchant response acceptance is: For the merchant's reply content, the length of the reply, objectivity, emotional intensity and merchant reply recognition indicators are used to calculate the completeness of the review length indicator, the objectivity indicator of credibility, the emotional intensity indicator of usefulness and the user recognition of usefulness. If there is no merchant reply, the dimension value is 0; Step 3: Use the fuzzy comprehensive evaluation method to first classify the review quality level, and then determine the weights of different indicators under the same dimension according to the characteristics of the indicators, so as to calculate the quantitative value of each dimension; Step 4: For the quantitative value of each dimension, use the hierarchical analysis method to determine the weights of different dimensions, and then use the fuzzy comprehensive evaluation method to calculate the quality value of the online reviews. Then, based on the correspondence between the value and the quality level, the quality level evaluation of the reviews is obtained.

2. The e-commerce online review quality quantification method based on AHP-fuzzy comprehensive evaluation method according to claim 1 is characterized in that: Assign a weight to each indicator, perform weighted average according to the weight of the indicator, and obtain the quantitative value of each dimension.

3. The e-commerce online review quality quantification method based on AHP-fuzzy comprehensive evaluation method according to claim 2 is characterized in that: The calculation process of the comment quality level in step 3 is: (1) Determine the evaluation factors and levels: The evaluation factors include 8 dimensions, denoted as V=[V1, V2, ..., V8], and the quantitative value of each dimension is denoted as Vi. The quality of comments is divided into three levels: low, medium, and high, corresponding to the evaluation set T=[low, medium, high]. The quality value of each comment is calculated through the fuzzy relationship matrix and weight of the evaluation factors; (2) Constructing the fuzzy relationship matrix: The fuzzy comprehensive evaluation method uses membership functions to construct a fuzzy relationship matrix to express the degree of compliance of each dimension with the quality grade; (3) Determination of evaluation factor weights based on the AHP method: The AHP method is used to calculate the weights of each dimension. The steps are as follows: ① Establish a hierarchical model: online review quality is the target layer, and the eight dimensions are the criterion layer; ② Construct a judgment matrix: Use the 1-9 scaling method to compare the dimensions of the criterion layer in pairs to form a judgment matrix; ③ Consistency test: Calculate the eigenvector and maximum eigenroot of the matrix and perform consistency test. After passing the test, the weight vector w is finally obtained, which is used as the weight of each dimension.

4. The e-commerce online review quality quantification method based on AHP-fuzzy comprehensive evaluation method according to claim 3 is characterized in that: The step 4 calculates the fuzzy comprehensive evaluation set for quality evaluation as follows: Multiply the fuzzy relationship matrix U by the eigenvector w to obtain the fuzzy comprehensive evaluation set S: in, , and Respectively represent The three quality level values ​​of low, medium and high in the fuzzy relationship matrix of the dimension are The first The weight value of each dimension, 1≤ ≤8; The fuzzy comprehensive evaluation set S has three values, corresponding to the three quality levels of low, medium and high. According to the rules of fuzzy comprehensive evaluation, the maximum value of these three values ​​is used as the quantitative representation of the quality of online reviews, and the quality level corresponding to the maximum value is the qualitative representation of the quality of online reviews.

5. The e-commerce online review quality quantification method based on AHP-fuzzy comprehensive evaluation method according to claim 3 is characterized in that: The trapezoidal function is selected as the membership function to construct the fuzzy relationship matrix. The trapezoidal function is a piecewise function divided into small, medium and large sizes. Each segment of the piecewise function represents the quality level of low, medium and high, as follows: The trapezoidal function formula for the small segment is: , The trapezoidal function formula of the intermediate segment is: , The trapezoidal function formula for the large segment is: , Among them, the independent variable That is, each dimension The quantitative value of ≤8, and the determination of the parameter values ​​of a, b, c, and d in each piecewise function needs to be adjusted according to the specific dimension value so that the adjusted dimension value falls between 0 and 1; and these three-segment divisions are the basic rules of the fuzzy comprehensive evaluation method. According to this rule, three values ​​will eventually be obtained, and the grade corresponding to the largest value is the corresponding quality; if the maximum value appears in a small size, then the quality of the review is low quality, if the maximum value appears in an intermediate size, then the quality of the review is medium quality, and if the maximum value appears in a large size, then the quality of the review is high quality.

Images