CN105205699A - User label and hotel label matching method and device based on hotel comments - Google Patents

Abstract

The invention discloses a method and device for matching user tags and hotel tags based on hotel reviews. The method of the invention includes: preparing a database of emotional sentence templates for the hotel industry; preparing final hotel tags for at least three hotels; At least two user reviews of the hotel or different hotels; compare the emotional sentence with the emotional sentence template, filter out the matching emotional sentence and identify it as different dimensions, and then form a user tag set for a specific user with all the identified dimensions ;Calculate the weight of each user tag separately, wherein the higher the frequency of occurrence in all user reviews of a specific user and the lower the frequency of appearance in all user reviews of all users for all hotels, the higher the weight of the user tag; select The user label with higher weight is used as the end-user label of the specific user; and the hotel with a high matching rate between the final hotel label and the specific user's end-user label is recommended to the specific user.

Description

Method and device for matching user tags and hotel tags based on hotel reviews

technical field

The invention relates to an Internet information processing method, in particular to a method and device for generating user portraits.

Background technique

The changing times will inevitably bring about many social changes. After the Internet gradually enters the era of big data, it will inevitably bring a series of changes and reshaping to the behavior of enterprises and consumers. The fast and unbreakable rhythm of the Internet has disrupted the logic of the original business evolution, making business participants have to face unprecedented changes and adapt to the changes of the times at an accelerated pace. How to use big data to mine potential business value, and how to actually apply big data technology in enterprises. With the discussion and innovation of big data applications, personalized technology has become an important landing point. Compared with traditional offline member management, questionnaire survey, and shopping basket analysis, big data enables enterprises to obtain more extensive feedback information from users conveniently through the Internet for the first time, in order to further accurately and quickly analyze user behavior habits and consumption habits and other important business information, providing a sufficient data base. With the gradual deepening of understanding of people, the concept of "user portrait" came into being. It perfectly abstracts the whole picture of a user's information, which can be regarded as the foundation of enterprise application of big data.

A user portrait is a virtual representative of a real user, a virtual user based on a deep understanding of real data. After collecting and analyzing the main information data of consumers such as social attributes, living habits, consumption behavior, and differences in opinions, the company distinguishes them into different types, and then extracts typical characteristics from each type, and gives them a name and a photo. , some demographic elements, scenes and other descriptions form a user portrait, which is the overall business picture of the user, and can be regarded as the basic way for enterprises to apply big data technology. User portraits provide a sufficient information base for companies, and can help companies quickly find accurate user groups and broader feedback information such as user needs.

Big data processing is inseparable from computer operations. User portraits can be represented by a set of tags. Tags are symbolic representations of certain user characteristics. User information tagging provides a convenient way for computers to programmatically process and People-related information, and even "understand" people through algorithms and models.

A label is usually a pre-defined highly refined feature identification, such as age group label: 25 to 35 years old, region label: Beijing, the label presents two important features: (1) Semantic, people can easily understand each Label meaning, which also makes the user portrait model have practical significance, and can better meet business needs, such as judging user preferences; (2) short text, each label usually only expresses one meaning, and the label itself does not need to do too much text Analysis and other preprocessing work, which facilitates the use of machines to extract standardized information.

User portrait tags specifically include two aspects: tags and their weights. Tags represent content, and users have interests, preferences, needs, etc. in this content. Weight, which represents the index, the user's interest and preference index, may also represent the user's demand, which can be simply understood as confidence.

The focus of user portrait work is to "label" users, and a label is usually a highly refined feature identifier that is artificially specified, such as age, gender, region, user preference, etc. Finally, when all the labels of the user are combined, it is basically A three-dimensional "portrait" of the user can be drawn.

Specifically, when profiling users, two steps are required: collecting data and analyzing tags.

First of all, collect all the relevant data of the user and divide the user data into two categories: static information data and dynamic information data. Static data refers to the relatively stable information of the user, such as gender, age, region, occupation, etc. Constantly changing behavioral information, such as browsing the web, searching for products, posting comments, contact channels, etc.

Secondly, by analyzing the data, affix corresponding tags and indexes to the users. The tags represent the user's interest, preference, and demand for the content, and the index represents the user's degree of interest, demand, and purchase probability.

For example, a method for obtaining a complete user portrait disclosed in Chinese Patent Application Publication No. 104750731A includes: obtaining an incomplete user portrait matrix, and randomly generating a user parameter matrix P and a label matrix Q; calculating the first part of the user portrait error, and updating the user A parameter matrix and a label parameter matrix, wherein the first change difference of the selected first part of users is greater than the first change difference of the first remaining users, and the first remaining users are users other than the first part of users among the plurality of users , the first change difference is the difference between the first predicted value updated by the user at the r-1th time and the first predicted value updated by the user at the r-2th time; update the user parameter matrix P and label parameters at the Rth time After matrix Q, according to the result of matrix decomposition, a complete user portrait matrix is obtained.

Another example is a method for updating tag thesaurus of a portrait system disclosed in Chinese Patent Application Publication No. 104268292A, which includes: acquiring user portrait data, the portrait data including tags used to describe the user and published by the user Original text; when the ratio of the number of labels to the number of original texts is less than the preset first threshold, word segmentation is performed on all original texts published by the user to obtain multiple label candidates, and the label candidates are sent To the recommendation system; the recommendation system calculates the vector distance between each label candidate word and each word in the preset word vector model file, and adds the label candidate words whose vector distance is greater than the preset second threshold to the label lexicon, Deleting no label candidate words whose vector distance is greater than the second threshold.

Another example is a microblog tag-based crowd portrait system and method disclosed in Chinese Patent Application Publication No. 103577549A, which includes two modules: microblog tag recommendation and tag topic clustering. The first module uses a three-step Tag recommendation algorithm. The first step is the recommendation of homogeneous tags, the second step is the expansion of co-occurrence tags; the third step is to establish a semantic network based on the Chinese knowledge graph, and use the network topology characteristics to measure the semantic similarity between tags, so that Remove tags with the same or similar semantics to ensure the refinement of tags used to describe users.

However, none of the application fields of the user portrait technology disclosed in the above three patent documents belong to the hotel industry involved in the present invention.

In the hotel industry, the current research and application of user portrait labeling analysis mainly focus on data such as user attributes and user behaviors. User attribute data includes age, gender, region, etc. Access history, click history, consumption history and other data, research and application based on review data are less. The main problem in this regard is that it is difficult to analyze and understand review texts. It is necessary to use technologies such as natural language processing to convert unstructured data into structured data, so that common user tag analysis algorithms can be applied.

Therefore, providing a method for matching user tags and hotel tags based on hotel reviews has become an urgent problem in the industry.

Contents of the invention

The purpose of the present invention is to provide a method and device for matching user tags and hotel tags based on hotel reviews, which model hotels and users through tags, so as to better establish associations between hotels and users.

Common user comment analysis methods are based on structured data, such as user attribute data, including age, gender, region, etc., or user behavior data including user visit history, click history, and consumption history on official websites or mobile applications. The present invention aims at less research and application of hotel review data, not only can analyze whether the user's evaluation of the hotel is positive or negative, but also can dig out the dimensions, based on which the hotel and user tags can be constructed.

The present invention first obtains online review data from major mainstream review (Online Travel Agent, OTA) websites by focusing on crawlers. Then, for large-scale reviews, the hotel industry emotional lexicon and domain knowledge base are sorted out in an automatic/semi-automatic manner. Finally, for each sentence in the review, analyze natural language processing technologies such as word segmentation, part-of-speech tagging, and phrase structure syntax analysis. On this basis, keywords or key sentence patterns are extracted as features, and sentiment classification is realized through the maximum entropy classifier . For sentences expressing emotions, further dimensionality is obtained based on domain keywords and knowledge base reasoning. Each dimension reflects a perspective from which people observe, understand and describe a hotel or user.

The present invention describes in detail the focus of attention of both hotels and users in the hotel industry through dimensions, and uses this as a label set. The user tags reflect what the user cares about, while the hotel tags reflect what the hotel is good at. Take the scenario of recommending hotels to users as an example. When the tags that users care about are more similar to the tags that the hotel is good at, or the degree of matching is higher, the more suitable to recommend to users. With the set of labels, the next step is to calculate the label weight for all reviews of a user or all reviews of a hotel. The weight calculation is mainly based on how often tags appear in reviews. The difference between hotel tags and user tags is that in order to reflect the hotel's expertise in a certain aspect, it is necessary to consider the emotional polarity of the tag corresponding to the review point. On a certain label, the more good reviews there are, the better the hotel is considered to be good at this aspect and the better it is doing.

The dimension referred to in the present invention refers to the sentence emotion type that can express the evaluation of a certain aspect of the hotel, such as the hygiene level of the hotel, the convenience of transportation, the surrounding environment index, the size of the room space, etc. It can specifically include several dimensions, For example, dimension 1 indicates that the hygiene level is grade A; ...dimension 12 indicates that the transportation convenience is grade B; ...dimension 53 indicates that the surrounding environment index is grade C; ...dimension 104 indicates that the room size is grade D, etc.

The different attributes of vocabulary referred to in the present invention refer to the classification of vocabulary into attributes such as evaluation object words, evaluation attribute words, and emotional words.

According to one aspect of the present invention, a kind of user tag and hotel tag matching method based on hotel comments are provided, comprising: (1), preparing the hotel industry emotional sentence template library, the hotel industry emotional sentence template library includes at least 100 emotional sentence templates; (2), prepare the final hotel labels of at least three hotels; (3), obtain at least two user reviews of the same hotel or different hotels from a specific user from the Internet; (4), combine the emotional sentences of all user reviews of a specific user Compare with at least 100 emotional sentence templates one by one, filter out the emotional sentences that match at least 100 emotional sentence templates, and identify the screened emotional sentences into different dimensions according to the type of emotion expressed, and then use the All dimensions identified form a user label set of a specific user; (5), respectively calculate the weight of each user label in the user label set of a specific user, wherein the higher the frequency of occurrence in all user comments of a specific user and the higher the The lower the frequency that all users appear in all user reviews of all hotels, the higher the weight of the user tag; (6), select the user tag with a weight greater than the first set threshold from the user tag set of the specific user as the final result of the specific user user tags; and (7), at least recommending to the specific user the hotels whose matching rate between the final hotel tag and the specific user's end user tag is among the top three.

Wherein, according to the specific conditions of use, preparing the final hotel labels of at least three hotels may be at least 10, at least 100 or at least 500 final hotel labels.

Optionally, the review data can be obtained from the review website through other devices or manually for backup.

Optionally, the hotel industry semantic dictionary can be sorted out in advance by other devices or manually for future use.

Optionally, the hotel industry emotional sentence template library can be sorted out in advance by other devices or manually for future use.

Optionally, the seed semantic dictionary can be sorted out in advance through other devices or manually for future use.

Optionally, preparing the final hotel labels of at least three hotels in step (2) includes: (2.1), obtaining user reviews for at least three hotels from the Internet, wherein each hotel includes user reviews for at least three users (2.2), comparing the emotional sentences of all user reviews of a specific hotel with at least 100 emotional sentence templates one by one, screening out the emotional sentences matching at least 100 emotional sentence templates, and using the selected emotional sentences Sentences are identified as different dimensions according to the type of emotion expressed, and then form a hotel tag set of a specific hotel with all the identified dimensions; (2.3), respectively calculate the weight of each hotel tag in the hotel tag set of a specific hotel, where , the higher the frequency of appearing in all user reviews of the same hotel and the lower the frequency of appearing in all user reviews of all hotels, the higher the weight of the hotel label; (2.4), select the weight greater than the first 2. The hotel label of the threshold value is used as the final hotel label of the specific hotel; and (2.5), repeating steps (2.2)-(2.4) until the final hotel labels of all hotels are obtained.

Optionally, preparing the hotel industry emotion statement template library in step (1) may include selecting at least 100 emotion statements as emotion statement templates according to the frequency of occurrence of the statements in at least 10,000 hotel user reviews obtained from the Internet.

Optionally, it further includes selecting at least 1,000 commonly used words in the hotel industry from at least 10,000 hotel user reviews according to the frequency of occurrence of the words to construct a semantic dictionary for the hotel industry.

Optionally, in step (1), before preparing the hotel industry emotional sentence template storehouse, further comprise the step of constructing hotel industry semantic dictionary, in step (4), the emotional sentence of all user comments of specific user is compared one by one with at least 100 Comparing the emotional sentence templates includes: (4.1), cutting the specific emotional sentence into a number of corresponding hotel industry vocabulary in the semantic dictionary of the hotel industry; (4.2), according to the different attributes of each vocabulary in the specific emotional sentence, respectively Compare with at least 100 emotional sentence templates, thereby determine whether match with any emotional sentence template in at least 100 emotional sentence templates; And (4.3), repeat steps (4.1)-(4.2), until screening out and All sentiment statements that match at least 100 sentiment statement templates.

Optionally, in step (2.2), the emotional sentences for all user reviews of specific hotels are compared with at least 100 emotional sentence templates one by one and include: (2.2.1), specific emotional sentences are segmented into the hotel industry semantic dictionary (2.2.2), compare with at least 100 emotional sentence templates respectively according to the different attributes of each vocabulary in the specific emotional sentence, thereby determine and at least 100 emotional sentence templates Whether any emotional statement template matches; and (2.2.3), repeat steps (2.2.1)-(2.2.2), until all emotional statements matching at least 100 emotional statement templates are screened out.

Optionally, in step (3), the focused crawler may be used to obtain user reviews from review websites.

Optionally, in step (1), preparing the emotional sentence template library for the hotel industry may extract sentence pattern templates through a bootstrapping method based on user comments, thereby obtaining the emotional sentence template library for the hotel industry.

Optionally, the steps of preparing the hotel industry emotional sentence template library and building the hotel industry semantic dictionary include: (1.1), obtaining comment data, forming a seed dictionary by sorting out the words of each emotional element; Word segmentation processing, and then determine its semantic class word by word and replace it with the semantic class label; (1.3), segment the comment data after the label replacement, and generate a template according to the name of each semantic class and the specific words contained in each semantic class; (1.4 ), apply the template to the comment data after the replacement of the semantic class label, to extract the semantic words of each semantic class; (1.5), score each template according to the importance, generalization and accuracy of the template; (1.6) , select the part template with the highest score, calculate the score of the semantic words extracted by each template according to the selected template and scoring, and then select the part semantic word with the highest score to expand the semantic dictionary; and (1.7), step (1.2) Go to step (1.6) iteratively, until the selected semantic words are incorrect, the iteration is terminated, and the final hotel industry semantic dictionary is obtained, and the template library of hotel industry emotion sentences is formed by each template.

Optionally, step (1.1) obtains online review data from review websites by focusing crawlers, and manually checks a small number of reviews to sort out words of each semantic category to form a seed dictionary.

Optionally, step (1.2) first adopts the maximum matching word segmentation method based on the dictionary to perform word segmentation, and then uses the word segmentation method of sequence labeling to obtain correct word segmentation results for the ambiguous part of the word segmentation; The segmentation problem is transformed into a character classification problem. Each character is assigned a different position category mark according to its different position in the word, and the sentence segmentation method is determined based on such a mark sequence.

Optionally, different position categories are labeled, including initials, midwords, endings, and single-word words, and a conditional random field model is used to implement the sequence labeling task.

Optionally, the semantic classes in step (1.2) include evaluation object words, evaluation attribute words, emotional words, degree adverbs, common adverbs, negative words, and insertion words.

Optionally, step (1.3) performs sentence segmentation according to the three punctuation marks ".", "!", and "?", and limits the minimum length of the template to 3 words and the maximum length to 7 words.

Optionally, when step (1.4) extracts semantic words of each semantic class, when the difference between the template corresponding to a certain review segment and the template obtained in step (1.3) is only one word, the word is used as an instance word of the corresponding semantic class.

Optionally, the method for scoring each template in step (1.5) is:

1) The formula for calculating the score S(pat _i ) for the importance and promotion of the template is as follows:

Among them, |pat _i | is the length of the template pat _i , calculated by the number of words, f(pat _i ) indicates the frequency of the template pat _i , and C(pat _i ) indicates the template set of the nested pat _i ;

2) The formula for calculating the template accuracy score P(pat _i ) is as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ,</span>$

Wherein, T(pat _i ) represents the set of semantic words extracted from template pat _i , f(t) represents the frequency of semantic word t, and SemLex is a seed semantic dictionary;

3) Using the Sigmoid function Normalize S(pat _i ) to (0,1), and then integrate the two scores to get F(pat _i ), the calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>$

Where α is the weight of the importance and generalization score S(pat _i ), and the value range is [0,1].

Optionally, the partial templates with the highest score in step (1.6) are the top 5-10% templates with the highest score, and the partial semantic words with the highest score are the top 5-10% semantic words with the highest score.

Optionally, after step (1.7), manually determine the polarity of the emotional word in the semantic dictionary, and the collocation polarity of the emotional word, the evaluation object word, and the evaluation attribute word; The review fragments are used as the basis for judgment.

Optionally, the step of performing sentiment analysis on comments in the present invention includes: obtaining comment data, and standardizing it; performing word segmentation processing on sentences of comment data after normalization; performing element analysis on sentences after word segmentation, and identifying All kinds of words that affect the detection and analysis of the text’s emotional tendency; perform sentence template matching on the comment data after element analysis according to the sentence template library; determine the antecedent corresponding to the pronoun in the sentence of the comment data, and restore the omitted subject; Sentences with evaluation object words, evaluation attribute words or emotional words are used as candidate emotional sentences, and the maximum entropy model is used to judge the sentence polarity of candidate emotional sentences to obtain the emotional orientation of the sentence.

Optionally, the standardization process is to adopt a rule-based method to process spelling errors in the review text, and the rule is a mapping from "a word string or a word string that contains a typo" to "the corresponding correct word string or a word string"; Obtained through two methods: one is based on existing knowledge, that is, common spelling mistakes summarized by predecessors; the other is to extract similar words or words based on the context of each word or word, and determine the correct string or word through manual verification string.

Optionally, at first adopt the dictionary-based maximum matching word segmentation method to carry out word segmentation, then use the word segmentation method of sequence labeling to obtain correct word segmentation results for the ambiguous part of word segmentation; the word segmentation method of said sequence labeling converts the segmentation problem of words into In the word classification problem, each word is assigned a different position category mark according to its different position in the word, and the segmentation method of the sentence is determined based on such a mark sequence.

Optionally, different position categories are labeled, including initials, midwords, endings, and single-word words, and a conditional random field model is used to implement the sequence labeling task.

Optionally, the elements include evaluation object words, evaluation attribute words, emotional words, degree adverbs, common adverbs, negative words, interjections, and words about cities and scenic spots in the review data. After the elements in the sentence are identified , labeled with the corresponding category label.

Optionally, sentence pattern templates are extracted through a comment-based bootstrapping method, thereby establishing a sentence pattern template library.

Optionally, if there are no evaluation object words or evaluation attribute words in the current sentence, then select the last mentioned evaluation object or evaluation attribute words in the previous sentence to introduce into the current sentence; if there are only evaluation attribute words in the current sentence, then when the previous sentence appears Introduce it into the current sentence when evaluating an object.

Optionally, the maximum entropy model predicts different emotional categories and estimates their probabilities by establishing a conditional probability model. The emotional categories include -1, 0, and 1, which represent negative reviews, no emotions, and favorable reviews, respectively.

According to another aspect of the present invention, a kind of user tag and hotel tag matching device based on hotel reviews is provided, comprising: a hotel industry emotional sentence template library generating module, the hotel industry emotional sentence template library including at least 100 emotional sentence templates; A tag generation module, which is used to generate the final hotel tags of at least three hotels; a user review acquisition module, which obtains at least two user reviews of a specific user for the same hotel or different hotels from the Internet; a user tag set generation module, which uses specific All user commented emotional sentences are compared with at least 100 emotional sentence templates one by one, and the emotional sentences matching at least 100 emotional sentence templates are screened out, and the selected emotional sentences are identified according to the type of emotion expressed are different dimensions, and then form a specific user’s user tag set with all the identified dimensions; the user tag weight calculation module calculates the weight of each user tag in the specific user’s user tag set respectively, wherein, in the specific user’s The higher the frequency in all user reviews and the lower the frequency in all user reviews for all hotels, the higher the weight of the user tag; the end user tag generation module, which selects the weight from the user tag set of a specific user A user label greater than the first set threshold is used as the end user label of the specific user; and a hotel recommendation module, which at least recommends to the specific user the hotels whose matching rate between the final hotel label and the specific user's end user label is in the top three.

Optionally, the final hotel label generation module can obtain user reviews for at least three hotels from the Internet through the user review acquisition module, wherein each hotel includes user reviews of at least three users; the final hotel label generation module can also include : The hotel tag collection generation sub-module, which compares the emotional sentences of all user reviews of a specific hotel with at least 100 emotional sentence templates one by one, screens out the emotional sentences matching at least 100 emotional sentence templates, and The screened emotional sentences are identified as different dimensions according to the type of emotion expressed, and then form a hotel label set of a specific hotel with all the identified dimensions; and the hotel label weight calculation sub-module, which calculates the hotel label set of a specific hotel respectively. The weight of each hotel tag in , where the higher the frequency of occurrence in all user reviews for the same hotel and the lower the frequency of occurrence in all user reviews for all hotels, the higher the weight of the hotel tag; where, the final hotel The label generation module selects the hotel label with a weight greater than the second set threshold from the hotel label set as the final hotel label of the specific hotel.

Optionally, the hotel industry emotion statement template library generation module can obtain at least 10,000 hotel user comments from the Internet through the user comment acquisition module and select at least 100 emotion statements as emotion statement templates according to the frequency of the statements.

Optionally, it may further include a hotel industry semantic dictionary generation module, which screens out at least 1000 commonly used words in the hotel industry from at least 10,000 hotel user reviews according to the frequency of occurrence of words to construct a hotel industry semantic dictionary.

Optionally, the first set threshold or the second set threshold can be arbitrarily selected within the range of 0-1. For example, the first set threshold is selected as 0.5, and the second set threshold is selected as 0.3.

As an alternative, the present invention can adopt a method based on Bootstrapping for the construction of the hotel industry semantic dictionary and sentence template library.

Bootstrapping, that is, self-expanding or bootstrapping, is a semi-supervised machine learning method that can be used to simultaneously extract semantic dictionaries and templates. The idea of this approach is based on the observation that extracted templates can be used to extract new instances, which in turn can be used to extract new templates. The advantage of this approach is that no labeled training corpus is required and only a few seeds are required. Firstly, the initialized seed words are obtained through manual intervention, the templates are obtained by using the seed words, and then the seed words are obtained through the templates, and so on. In each round of iteration, new annotation data will be generated, the optimal word will be added to the corresponding semantic dictionary, and the optimal template will also be added to the template library, and these new annotation data will be used to re-learn the model. In this way, new data can be generated, and so on, until the final convergence is over, so as to obtain more seed words and templates. This is the most basic Bootstrapping algorithm (or process).

The semantic categories of the semantic dictionary include evaluation object words, evaluation attribute words, emotional words, degree adverbs, common adverbs, negative words, insertion words, etc. Each semantic category includes several words, and the template is composed of semantic category names or specific words. sequence.

The following are the specific implementation steps:

Step 1: Data preparation. Obtain online review data from mainstream review sites such as Ctrip through focused crawlers.

Step 2: Seed dictionary construction. Manually check a small number of comments (such as 500), sort out the words of each semantic category, and record the semantic dictionary as SemLex.

Step 3: Click on the scoring words. Chinese word segmentation is the basic step of Chinese natural language processing, and the word segmentation of the present invention adopts the method of fusion of dictionary word segmentation and statistical word segmentation. Firstly, the maximum matching word segmentation method based on the dictionary is adopted, and then the word segmentation method of sequence labeling is used for the ambiguous part of the word segmentation.

Based on the dictionary-based maximum matching word segmentation method, given a dictionary, for the sequence of Chinese characters to be segmented, search for the longest matching dictionary words in turn, and those without a match will be treated as single-character words until the Chinese character sequence is processed. According to the different scanning directions of the Chinese character sequence, the method can be divided into: forward maximum matching (matching from left to right) and reverse maximum matching (matching from right to left). For example, for the sequence "when atoms combine into molecules", the forward maximum matching result is "when | ". Obviously, neither the forward maximum matching nor the reverse maximum matching can handle the severance ambiguity well. Forward maximum matching and reverse maximum matching can also be combined to form two-way maximum matching. In two-way matching, where the forward and reverse matching are inconsistent is often a potential ambiguity. If there is ambiguity, it is often necessary to confirm the word segmentation result according to the specific context. The supervised sequence labeling method can fully mine the rich features of the context, so the present invention introduces the sequence labeling method to eliminate the ambiguity when there is ambiguity. This method converts the word segmentation problem into a character classification problem, and each character is assigned a different position category mark according to its different position in the word, such as the beginning of the word, the middle of the word, the end of the word, and a single word. Based on such a sequence of tokens, it is easy to determine how a sentence should be segmented. Among them, B (Begin), M (Middle), E (End), and S (Single) represent the beginning of a word, the middle of a word, the end of a word, and a single word, respectively. With the tag sequence of words, the word sequence conforming to the regular expression "S" or "B(M)*E" represents a word, so that sentence segmentation can be easily completed. In order to realize the task of sequence labeling, the invention adopts the conditional random field model (Conditional Random Fields, CRF), which is widely used in natural language processing and has achieved great success. The specific features include: the previous character, the current character, the next character, the previous character and the current character, the current character and the next character. The conditional random field model uses these extracted features to predict the category label of each word.

The dictionary of the maximum matching method and the training and learning corpus of the supervised conditional random field model all come from the 100,000 hotel reviews manually marked by the present invention.

Step 4: Semantic class label replacement. Determine the semantic class of the comment after word segmentation word by word and replace it with the semantic class label, such as "Restaurant|的|price|very|high", replace it with "Obj|的|Attr|Dgr|Sent", for the start and end of the comment "Start" and "End" tags are added to the positions, and punctuation marks other than ".", "!", and "?" in the comments are also replaced with "Punc" tags.

Step 5: Template generation. According to the three punctuation marks ".", "!", "?", the minimum length of the template is 3 words, and the maximum length is 7 words. Scan the comments after tag replacement to generate a template.

Step 6: Score templates. The present invention scores from two aspects. On the one hand, the importance and popularization of the template is measured by the frequency, and on the other hand, the accuracy of the template is measured by the hit rate in the semantic dictionary.

The calculation formula of pat _i importance and generalization score S(pat _i ) is as follows:

Among them, |pat _i | is the length of the template pat _i , calculated by the number of words, f(pat _i ) indicates the frequency of the template pat _i , and C(pat _i ) indicates the template set of the nested pat _i .

The calculation formula of pat _i accuracy score P(pat _i ) is as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; m</span>}\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span>}$

Among them, T(pat _i ) represents the set of semantic words extracted from template pat _i , and f(t) represents the frequency of semantic words.

Using the Sigmoid function Normalize S(pat _i ) to (0,1), and then integrate the two scores to get F(pat _i ), the calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; e</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; S</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; log</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$

α=0.4, the present invention pays more attention to the accuracy of the template.

Step 7: Template selection. Select the top 5% with the highest score according to F(pat _i ).

Step 8: Semantic word extraction. Apply the selected template to the comment after replacing the semantic class label. When there is only one word difference between a review segment and the selected template, this word is used as an instance word of the corresponding semantic class.

Step 9: Score semantic words.

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; pat</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>$

Step 10: Semantic Dictionary Expansion. Pick the top 5% with the highest scores.

Steps 4 to 10 are performed iteratively. The iteration termination condition. Terminate when the selected semantic words are obviously incorrect.

Step 11: Polarity determination. The polarity of emotional words, as well as the collocation polarity of emotional words, evaluation object words, and evaluation attribute words are done manually. In the manual determination process, the comment segment corresponding to the template it belongs to is used as the basis for determination.

The results show that the present invention has achieved good performance in both accuracy and recall. Generate high-quality semantic dictionaries and sentence template libraries.

As another alternative, the emotional sentence template construction and sentence comparison and analysis method of the present invention are as follows.

The present invention first obtains online review data from major mainstream review websites by focusing crawlers. Then, for large-scale reviews, the semantic dictionary and sentence library are sorted out in a semi-automatic manner. Finally, for each sentence in the review, word segmentation and other processing and analysis are performed, and on this basis, keywords or key sentence patterns are extracted as features, and sentiment classification is realized through the maximum entropy classifier. Including the following steps:

Step 1: Text normalization.

Internet review texts are often subject to spelling errors, and we take a rule-based approach to these issues. These rules are the mapping of "strings or word strings containing typos" to "corresponding correct word strings or word strings". This rule is obtained through two methods: one is based on existing knowledge, that is, common spelling mistakes summarized by predecessors; the other is to extract similar words or words based on the context of each word or word, and manually check and confirm. This method is simple and effective. The performance of this module of the system depends on the number of spelling error correction rules, which can be continuously summarized and enriched during the operation and maintenance of the system.

There is also the problem of full-width punctuation marks in Chinese. According to the mapping relationship between full-width and half-width symbols, punctuation marks are uniformly marked as half-width symbols.

Step 2: Click on the scoring words.

Point rating words. Chinese word segmentation is the basic step of Chinese natural language processing, and the word segmentation of the present invention adopts the method of fusion of dictionary word segmentation and statistical word segmentation. Firstly, the maximum matching word segmentation method based on the dictionary is adopted, and then the word segmentation method of sequence labeling is used for the ambiguous part of the word segmentation.

Based on the dictionary-based maximum matching word segmentation method, given a dictionary, for the sequence of Chinese characters to be segmented, search for the longest matching dictionary words in turn, and those without a match will be treated as single-character words until the Chinese character sequence is processed. According to the different scanning directions of the Chinese character sequence, the method can be divided into: forward maximum matching (matching from left to right) and reverse maximum matching (matching from right to left). For example, for the sequence "when atoms combine into molecules", the forward maximum matching result is "when | ". Obviously, neither the forward maximum matching nor the reverse maximum matching can handle the severance ambiguity well. Forward maximum matching and reverse maximum matching can also be combined to form two-way maximum matching. In two-way matching, where the forward and reverse matching are inconsistent is often a potential ambiguity. If there is ambiguity, it is often necessary to confirm the word segmentation result according to the specific context. The supervised sequence labeling method can fully mine the rich features of the context, so the present invention introduces the sequence labeling method to eliminate the ambiguity when there is ambiguity. This method converts the word segmentation problem into a character classification problem, and each character is assigned a different position category mark according to its different position in the word, such as the beginning of the word, the middle of the word, the end of the word, and a single word. Based on such a sequence of tokens, it is easy to determine how a sentence should be segmented. Among them, B (Begin), M (Middle), E (End), and S (Single) represent the beginning of a word, the middle of a word, the end of a word, and a single word, respectively. With the tag sequence of words, the word sequence conforming to the regular expression "S" or "B(M)*E" represents a word, so that sentence segmentation can be easily completed. In order to realize the task of sequence labeling, the invention adopts the conditional random field model (Conditional Random Fields, CRF), which is widely used in natural language processing and has achieved great success. The specific features include: the previous character, the current character, the next character, the previous character and the current character, the current character and the next character. The conditional random field model uses these extracted features to predict the category label of each word.

The dictionary of the maximum matching method and the training and learning corpus of the supervised conditional random field model all come from the 100,000 hotel reviews manually marked by the present invention.

Step 3: Element analysis.

Elements refer to the important factors that affect text sentiment analysis, including the above-mentioned emotional information elements, such as evaluation object words, evaluation attribute words, emotional words, degree adverbs, common adverbs, negative words, insertion words, etc. Including words of multiple categories such as cities and scenic spots. Element analysis is to identify the elements in the sentence and mark their corresponding category labels.

Step 4: Sentence matching.

After analyzing the elements of the sentence, the semantic classification form of the sentence is obtained, that is, the sentence pattern. The sentence pattern reflects the common context of the words or elements in it, so it has a certain disambiguation ability. In the sentence pattern matching process, the existing sentence pattern library plays a key role, which reflects the common sentence patterns expressing emotion in the domain. The sentence pattern database is the core resource of the present invention, which reflects the common sentence patterns of emotional expression in comments. The present invention extracts abstract sentences through a comment-based bootstrapping (Bootstrapping) method.

Step 5: Referential resolution.

Reference and omission are common linguistic phenomena. Coreference often means coreference, that is, both expressions refer to the same object. There are many types of reference, and we mainly focus on the situation where personal pronouns and demonstrative pronouns are used as reference pronouns. Ellipsis can be regarded as the case of zero referential pronouns, so we regard both referential and ellipsis as "sufficiency" in a broad sense, and referential resolution refers to finding the antecedent corresponding to the referent, or recovering the omitted subject. If there is no evaluation object word or evaluation attribute word in the current sentence, select the evaluation object or evaluation attribute word mentioned last in the previous sentence to introduce into the current sentence. If there are only evaluation attribute words in the current sentence, when the evaluation object appears in the previous sentence, it will be introduced into the current sentence.

Step 6: Sentiment Analysis.

Sentences with evaluation object words, evaluation attribute words or emotional words appear as candidate emotional sentences. For candidate emotional sentences, the Maximum Entropy (Maximum Entropy) model is used to integrate rich context features to judge the polarity of the sentence and obtain the emotional orientation of the sentence. In classification tasks, discriminative models tend to outperform generative models. The production model estimates the joint probability distribution, which is used in machine learning to directly model the data, or use Bayesian rule as an intermediate step to obtain the conditional probability. Discriminative models, on the other hand, directly model conditional probabilities, making the training and prediction of the model consistent and thus better distinguishing between classes. Among the discriminative models, the maximum entropy model is widely used in the field of natural processing. For the classification problem of given context information x ∈ X to predict category y ∈ Y, the maximum entropy model establishes a conditional probability model P(y|x) to predict different categories y ∈ Y and estimate its probability. The categories include -1 (bad review), 0 (no emotion), and 1 (good review). Features include evaluation object words, evaluation attribute words, emotional words, and their collocations, as well as features such as negative words and sentence patterns.

The beneficial effects of the present invention are: the solution of the present invention can effectively use hotel review data to form user portraits, and recommend hotels that best meet user needs to specific users according to the user portraits, which can significantly save users the time spent searching for hotels on the Internet and energy, and can also help the hotel discover/overcome its own deficiencies and further improve/optimize its own characteristics.

Description of drawings

FIG. 1 shows a schematic flow chart of the method for matching user tags and hotel tags based on hotel reviews in the present invention.

Detailed ways

The present invention will be described in further detail below with reference to the drawings and examples, but these descriptions do not limit the present invention in any form. Unless defined otherwise, all scientific and technical terms used herein have the meaning commonly understood by one of ordinary skill in the art to which this invention belongs and related technologies.

Referring to FIG. 1 , according to a non-limiting embodiment of the present invention, a method for matching user tags and hotel tags based on hotel reviews is provided, which specifically includes the following steps.

In step S1, about 50,000 hotel user reviews are obtained from the Internet, and about 5,000 commonly used words in the hotel industry are selected from them according to the frequency of occurrence of words to construct a semantic dictionary for the hotel industry.

In step S2, prepare the emotional sentence template database for the hotel industry, including selecting about 500 emotional sentences from about 50,000 hotel user reviews obtained from the Internet according to the frequency of the sentences as emotional sentence templates.

In step S3, the final hotel labels of about 200 hotels are prepared, specifically including: selecting user reviews for about 200 hotels from the about 50,000 hotels obtained above, wherein each hotel includes about 100 user comments User comments: compare the emotional sentences of all user comments for a specific hotel with about 500 emotional sentence templates one by one, filter out the emotional sentences that match about 500 emotional sentence templates, and use the selected emotional sentences according to The type of emotion expressed is identified as different dimensions, and then all the identified dimensions are used to form a hotel label set for a specific hotel. For example, the hotel label set of No. 1 Hotel includes: Dimension 1 (the hygiene level is A-level), dimension 11 ( Transportation convenience is grade A), dimension 51 (surrounding environment index is grade A), dimension 101 (room space size is grade A), etc.; the hotel label set of No. 2 Hotel includes: dimension 2 (level of sanitation is grade B), Dimension 12 (traffic convenience is grade B), dimension 52 (surrounding environment index is grade B), dimension 102 (room space size is grade B), etc.; the hotel label set of No. 3 Hotel includes: Dimension 3 (hygienic grade is C level), dimension 13 (level C for transportation convenience), dimension 53 (level C for surrounding environment index), dimension 103 (level C for room space size), etc.; calculate each hotel in the hotel label set of a specific hotel separately The weight of the label, wherein the higher the frequency of appearing in all user reviews of the same hotel and the lower the frequency of appearing in all user reviews of all hotels, the higher the weight of the hotel label; the weight of selecting from the hotel label set is greater than The hotel label of the second set threshold is used as the final hotel label of the specific hotel, wherein the second set threshold is selected as 0.4. Repeat this step until you have the final hotel labels for all hotels. Among them, the process of comparing the emotional sentences of all user reviews for a specific hotel with about 500 emotional sentence templates may specifically include: segmenting the specific emotional sentence into a number of common words in the hotel industry corresponding to the semantic dictionary of the hotel industry; Compare with 500 emotional statement templates respectively according to the different attribute of each vocabulary in specific emotional statement, thus determine whether match with any one emotional statement template in 500 emotional statement templates; And repeat this process until screening out and 500 All sentiment sentences matching a sentiment sentence template.

In step S4, three user comments on three hotels by a specific user are obtained from the Internet.

In step S5, compare the emotional sentences of all user reviews of the specific user with about 500 emotional sentence templates one by one, filter out the emotional sentences matching the about 500 emotional sentence templates, and use the filtered emotional sentences Identify different dimensions according to the type of emotion expressed, and then use all the identified dimensions to form a user label set for a specific user. For example, the user label set for a specific customer includes: Dimension 1 (the hygiene level is A), dimension 12 ( Transportation convenience is grade B), dimension 51 (surrounding environment index is grade A), dimension 103 (room space size is grade C), etc. Among them, the process of comparing the emotional sentences of all user reviews of a specific user with about 500 emotional sentence templates includes: dividing the specific emotional sentence into a number of common words in the hotel industry corresponding to the semantic dictionary of the hotel industry; The different attributes of each vocabulary in the emotional sentence are compared with 500 emotional sentence templates respectively, thereby determine whether match with any emotional sentence template in the 500 emotional sentence templates; And repeat this process until screening out and 500 emotional sentence templates All sentiment statements that match the statement template.

In step S6, the weight of each user tag in the user tag set of a specific user is calculated separately, wherein the higher the frequency of occurrence in all user reviews of a specific user and the higher the frequency of appearance in all user reviews of all users for all hotels The lower the frequency of , the higher the user tag weight.

In step S7, a user tag whose weight is greater than a first set threshold is selected from the user tag set of the specific user as the final user tag of the specific user, wherein the first set threshold is selected as 0.6.

In step S8, the hotel with the highest matching rate between the final hotel tag and the specific user's end user tag is recommended to the specific user, for example, in this non-limiting embodiment, Hotel No. 1 is recommended to the specific user.

According to another non-limiting embodiment of the present invention, a kind of user tag and hotel tag matching device based on hotel reviews is provided, including: a hotel industry emotional sentence template library generation module, and the hotel industry emotional sentence template library includes 1000 emotional sentences Template; the final hotel label generation module, which is used to generate the final hotel labels of 500 hotels; the user comment acquisition module, which obtains five user comments from the Internet for different hotels by a specific user; the user label collection generation module, which uses a specific user The emotional sentences commented by all users are compared with 1000 emotional sentence templates one by one, and the emotional sentences matching the 1000 emotional sentence templates are screened out, and the screened emotional sentences are identified as different according to the type of emotion expressed. Dimensions, and then form a user tag set of a specific user with all the identified dimensions; a user tag weight calculation module, which calculates the weight of each user tag in the user tag set of a specific user, wherein, in all user reviews of a specific user The higher the frequency of occurrence in and the lower the frequency of occurrence in all user reviews of all users for all hotels, the higher the user tag weight; the end user tag generation module, which selects a weight greater than the first from the user tag set of a specific user The user tag with a threshold value is used as the end user tag of the specific user; and a hotel recommendation module, which recommends the top ten hotels whose matching rate between the final hotel tag and the specific user's end user tag is in the top ten to the specific user.

The final hotel tag generation module obtains user reviews for 500 hotels from the Internet through the user review acquisition module, wherein each hotel includes user reviews of 200 users; the final hotel tag generation module also includes: a hotel tag set generation sub-module, It compares the emotional sentences of all user reviews of a specific hotel with 1000 emotional sentence templates one by one, screens out the emotional sentences that match the 1000 emotional sentence templates, and uses the selected emotional sentences according to the expressed emotions. The type is identified as different dimensions, and then a hotel tag set of a specific hotel is formed with all the identified dimensions; and a hotel tag weight calculation sub-module, which calculates the weight of each hotel tag in the hotel tag set of a specific hotel, wherein, The higher the frequency in all user reviews for the same hotel and the lower the frequency in all user reviews for all hotels, the higher the weight of the hotel tag; wherein, the final hotel tag generation module selects the weight from the hotel tag set Hotel tags greater than the second set threshold are used as final hotel tags for the specific hotel.

The hotel industry emotional statement template library generation module obtains 100,000 hotel user comments from the Internet through the user comment acquisition module and selects 1000 emotional statements as emotional statement templates according to the frequency of the statement.

The device of the present invention further includes a hotel industry semantic dictionary generating module, which screens out 10,000 commonly used words in the hotel industry from 100,000 hotel user reviews according to the frequency of occurrence of the words to construct a hotel industry semantic dictionary.

The present invention will be further described in detail below in conjunction with specific examples, but the examples should not be construed as limiting the protection scope of the present invention.

A method for matching user tags and hotel tags based on hotel reviews, comprising the steps of:

Step 1: Obtain online review data from mainstream review sites such as Ctrip through the focused crawler;

Step 2: Filter spam comments, spam comments include meaningless sentences;

Step 3: Build a hotel industry semantic dictionary and sentence template library;

Step 4: Perform sentiment analysis on reviews.

Step 5: Tag Analysis.

For each sentence that expresses emotion in the comments, dig out the views expressed, and express them through tags.

Step 6: Aggregate comment fragments according to tags, and calculate the weights of different tags of different users according to the TF-IDF algorithm. TF-IDF (Term Frequency-Inverse Document Frequency) is a statistical method used to evaluate the importance of words to documents, and is widely used in information retrieval, text feature selection and calculation. The main idea of TF-IDF is: if a word appears frequently in a document and rarely appears in other documents, it is considered that the word has a good category discrimination ability and is suitable for characterizing the document.

TF-IDF is actually the product of TF and IDF. TF stands for Term Frequency, which is the frequency of a given word appearing in a document. It is the normalization of the frequency of words to prevent bias towards documents with many words. Calculated as follows:

${\mathrm{<span\; class="notranslate">\; tf</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; no</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}}$

Among them, tf _{i, j} represents the frequency of word i in document j, _{ni, j} represents the frequency of word i in document j, and Σ _k n _{k, j} represents the sum of the frequencies of all words in the document.

IDF stands for Inverse Document Frequency (InverseDocumentFrequency), which is a measure of the general importance of a word. The calculation formula is as follows:

${\mathrm{<span\; class="notranslate">\; idf</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; log</span>}\frac{<span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \{</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; :</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \}</span><span\; class="notranslate">\; |</span>}$

Among them, idf _i represents the inverse document frequency of term i in the corpus, |D| represents the total number of documents in the corpus, and |{j:t _i ∈ d _j }| represents the number of documents containing term i. If the word is not in the corpus, the denominator will be zero, so in general the denominator uses |{j:t _i ∈ d _j }|+1.

With TF and IDF, and then calculate TFIDF, the calculation formula is as follows:

tfidf _{i, j} = tf _{i, j} × idf _i

A high-frequency term within a particular document, and a low document frequency of that term in the entire document collection, can produce a highly weighted TF-IDF. Therefore, TF-IDF tends to filter out common words and keep important words.

Step 7: For different hotels and different users, select according to their TF-IDF and preset thresholds, so as to obtain the final hotel labels and user labels.

Although the preferred embodiments of the present invention have been described in detail herein, it is to be understood that the present invention is not limited to the specific constructions described and shown in detail herein, and can be modified by those skilled in the art without departing from the spirit and scope of the present invention. Other modifications and variations will occur to the skilled artisan.

Claims (10)

1. A method for matching user tags and hotel tags based on hotel reviews, comprising: (1), prepare hotel industry emotional sentence template library, described hotel industry emotional sentence template library includes at least 100 emotional sentence templates; (2) Prepare final hotel labels for at least three hotels; (3) Obtain at least two user comments from a specific user on the same hotel or different hotels from the Internet; (4), compare the emotional sentences of all user reviews of the specific user with the at least 100 emotional sentence templates one by one, filter out the emotional sentences matching the at least 100 emotional sentence templates, and use the The filtered emotional sentences are identified as different dimensions according to the type of emotion expressed, and then form the user label set of the specific user with all the identified dimensions; (5), respectively calculate the weight of each user tag in the user tag set of the specific user, wherein the higher the frequency of occurrence in all user comments of the specific user and all users for all users of all hotels The lower the frequency of comments, the higher the weight of user tags; (6), selecting a user tag whose weight is greater than a first set threshold from the user tag set of the specific user as the end user tag of the specific user; and (7) At least recommending to the specific user the hotels whose matching ratios of the final hotel tags and the specific user's final user tags are among the top three. 2. the user label and hotel label matching method based on hotel reviews as claimed in claim 1, is characterized in that, prepares the final hotel label of at least three hotels in the described step (2) and comprises: (2.1) Obtain user reviews for at least three hotels from the Internet, including user reviews for at least three users for each hotel; (2.2), compare the emotional sentences for all user reviews of specific hotels one by one with the at least 100 emotional sentence templates, filter out the matching emotional sentences with the at least 100 emotional sentence templates, and filter the The emotion sentence that goes out is identified as different dimensions according to the emotion type expressed, forms the hotel label collection of described specific hotel with all dimensions identified again; (2.3), respectively calculate the weight of each hotel tag in the hotel tag set of the specific hotel, wherein the higher the frequency of occurrence in all user reviews for the same hotel and appear in all user reviews for all hotels The lower the frequency, the higher the weight of the hotel label; (2.4), select the hotel label with weight greater than the second set threshold from the hotel label set as the final hotel label of the specific hotel; and (2.5), repeat steps (2.2)-(2.4) until the final hotel labels of all hotels are obtained. 3. the user label and hotel label matching method based on hotel reviews as claimed in claim 2, it is characterized in that, in described step (1), before preparing hotel industry emotional statement template storehouse, further comprise building hotel industry semantics The step of dictionary, in described step (4), the emotional sentence of all user reviews of described specific user is compared with described at least 100 emotional sentence templates one by one and comprises: (4.1), the specific emotional sentence is divided into some corresponding common words in the hotel industry in the semantic dictionary of the hotel industry; (4.2), compare respectively with described at least 100 emotional sentence templates according to the different attributes of each vocabulary in the specific emotional sentence, thereby determine whether to match with any one emotional sentence template in the described at least 100 emotional sentence templates ;as well as (4.3), steps (4.1)-(4.2) are repeated until all emotional sentences matching the at least 100 emotional sentence templates are screened out. 4. the user label and hotel label matching method based on hotel reviews as claimed in claim 3, is characterized in that, in described step (2.2), will be aimed at the emotion sentence of all user reviews of specific hotel one by one and described at least 100 The emotional statement templates for comparison include: (2.2.1), the specific emotional sentence is divided into several hotel industry commonly used words corresponding to the hotel industry semantic dictionary; (2.2.2), compare with described at least 100 emotional sentence templates respectively according to the different attribute of each vocabulary in specific emotional sentence, thereby determine whether any emotional sentence template in described at least 100 emotional sentence templates match; and (2.2.3), repeating steps (2.2.1)-(2.2.2), until filtering out all emotional sentences matching the at least 100 emotional sentence templates. 5. The method for matching user tags and hotel tags based on hotel reviews as claimed in claim 4, wherein in said step (3), the user reviews are obtained from review sites by focusing crawlers. 6. the user label and the hotel label matching method based on hotel comments as claimed in claim 5, it is characterized in that, preparing the hotel industry emotion sentence template storehouse in the described step (1) is to extract sentence by the bootstrapping method based on user comments templates, so as to obtain the template library of emotional sentences in the hotel industry. 7. the user label and hotel label matching method based on hotel reviews as claimed in claim 6, is characterized in that, the step of preparing described hotel industry emotional sentence template storehouse and building described hotel industry semantic dictionary comprises: (1.1), obtain comment data, form the seed dictionary by sorting out the words of each emotional element; (1.2), word segmentation processing is performed on the sentence of the comment data, and then its semantic class is determined word by word and replaced with a semantic class label; (1.3), segment the comment data after label replacement, and generate templates according to the names of each semantic class and the specific words contained in each semantic class; (1.4), the template is applied to the comment data after the semantic class label is replaced, to extract the semantic words of each semantic class; (1.5) Score each template according to its importance, promotion and accuracy; (1.6), select the partial template with the highest score, calculate the score of the semantic words extracted by each template according to the selected template and its scoring, and then select the partial semantic word with the highest score to expand the semantic dictionary; and (1.7), step (1.2) to step (1.6) are carried out iteratively, until the selected semantic words are incorrect, the iteration is terminated, and the final hotel industry semantic dictionary is obtained, and the hotel industry emotional sentence template library is formed by each template. 8. The user label and hotel label matching method based on hotel reviews as claimed in claim 7, wherein the part templates with the highest score described in the step (1.6) are the top 5-10% templates with the highest score, so The part of semantic words with the highest score is the top 5-10% semantic words with the highest score. 9. A device for matching user tags and hotel tags based on hotel reviews, comprising: The hotel industry emotional sentence template library generation module, the hotel industry emotional sentence template library includes at least 100 emotional sentence templates; a final hotel label generation module for generating final hotel labels for at least three hotels; A user review acquisition module, which obtains at least two user reviews of a specific user for the same hotel or different hotels from the Internet; A user label set generation module, which compares the emotional sentences of all user comments of the specific user with the at least 100 emotional sentence templates one by one, and filters out the emotional sentences matching the at least 100 emotional sentence templates, And identify the emotional sentences that are screened out as different dimensions according to the expressed emotion type, and then form the user tag set of the specific user with all the identified dimensions; User tag weight calculation module, which respectively calculates the weight of each user tag in the user tag set of the specific user, wherein the higher the frequency of occurrence in all user comments of the specific user and the higher the frequency of occurrence in all user reviews for all hotels The lower the frequency of occurrence in all user reviews of , the higher the weight of the user tag; An end-user tag generation module, which selects a user tag whose weight is greater than a first set threshold from the user tag set of the specific user as the end-user tag of the specific user; and The hotel recommending module at least recommends to the specific user the hotels whose matching ratios of final hotel tags and the specific user's final user tags are among the top three. 10. The user tag and hotel tag matching device based on hotel reviews as claimed in claim 9, wherein said final hotel tag generation module obtains user reviews for at least three hotels from the Internet through said user review acquisition module. Reviews, which include, for each property, user reviews of at least three users; Described final hotel label generating module also includes: The hotel label set generates a submodule, which compares the emotional sentences for all user comments of a specific hotel with the at least 100 emotional sentence templates one by one, and screens out the matching emotional sentences with the at least 100 emotional sentence templates, And identify the filtered emotion sentences as different dimensions according to the type of emotion expressed, and then form the hotel tag set of the specific hotel with all the identified dimensions; and The hotel tag weight calculation sub-module, which calculates the weight of each hotel tag in the hotel tag set of the specific hotel respectively, wherein, the higher the frequency of occurrence in all user comments for the same hotel and the higher the frequency for all hotel tags The lower the frequency of user reviews, the higher the weight of the hotel label; Wherein, the final hotel label generating module selects a hotel label with a weight greater than a second set threshold from the hotel label set as the final hotel label of the specific hotel.