KR20120071194A - Apparatus of recommending contents using user reviews and method thereof - Google Patents

Abstract

PURPOSE: A content recommendation apparatus using user reviews and method thereof are provided to recommend content according to the preference of a user by providing similar content with content which the user is watching. CONSTITUTION: A review searching unit(200) searches one or more review text corresponding to each content based on metadata. A review analysis unit(210) extracts core words through a natural language analysis techniques for the searched review text. The review analysis unit stores the extracted core words in a review analysis DB(DataBase)(240). A review analysis database construction apparatus stores the core data as a feature vector type.

Description

Apparatus of recommending contents using user reviews and method approximately}

The present invention relates to an apparatus for recommending content and a method thereof, and more particularly, to an apparatus and method for recommending content having characteristics similar to content currently being viewed using a user review to a user.

Conventional Video on Demand (VOD) search is a method of searching for other content having a common keyword by comparing metadata (eg, genre, actor, synopsis, title, production year, etc.) of each content. The number of IPTV-based VOD content reaches about 90,000, and this number will increase rapidly due to the impact of open IPTV. Metadata is useful when searching for a specific content in a large amount of content, but a user who wants to find a movie or video to watch should spend a lot of time searching based on the metadata. For example, 'A actor appearing movie' content can be searched by inputting 'A actor' into the actor, but there is a disadvantage in that a user must remember keywords that can distinguish each content from other contents. However, for most movies, the user cannot remember these characteristics, so the keyword-based search using conventional metadata is limited.

In addition, there is a conventional Collaborative Filtering-based recommendation method that recommends a VOD list without searching for a user. This recommendation method analyzes the VOD purchase pattern of the user and compares it with other users with similar purchase patterns, and recommends VOD content that the current user has not watched. However, since the conventional recommendation method is mainly based on the past purchase pattern, it is highly likely that the recommended list is not a movie that the viewer wants to watch at present, and there is a disadvantage in that there is a accumulated past purchase pattern.

The technical problem to be achieved by the present invention is to recommend content that has characteristics similar to the content currently being viewed by the user, thereby reducing the time spent by the user to find the next content to be viewed and increasing the likelihood of purchase by recommending content that matches the user's preference. The present invention provides a device for recommending content using a user review and a method thereof.

In accordance with one aspect of the present invention, there is provided a review DB building device, including: a review searcher searching for at least one review text corresponding to each content based on metadata; And a review analyzing unit extracting the searched review texts through natural language analysis techniques and storing them in a review analysis database.

In accordance with an aspect of the present invention, there is provided a method for constructing a review analysis database according to an embodiment of the present invention. The method for constructing a review analysis database by a content recommendation apparatus includes: at least one corresponding to each content based on metadata; Searching and downloading the review text from the review site; And a review analysis step of extracting and storing the core word through the downloaded review text through a natural language analysis technique.

In order to achieve the above technical problem, an embodiment of a content recommendation apparatus according to the present invention includes a review analysis database for storing review data of each content; A content searching unit searching for contents having review data similar to review data of the content currently being viewed in the review analysis database; And a content recommending unit recommending the searched contents to the user.

In order to achieve the above technical problem, an embodiment of a content recommendation method according to the present invention includes a content having review data similar to review data of a content currently being viewed in a review analysis database storing review data of each content. Searching; And recommending the retrieved content to the user.

According to the present invention, by recommending content having characteristics similar to the content currently being viewed by the user, the user can reduce the time spent searching for other contents, and can increase the likelihood of purchase by recommending content that matches the user's preference. . In addition, the recommendation method using a review has an advantage of recommending similar contents even without a user's long-term purchase history pattern.

1 is a diagram illustrating an example of a process of databaseting a user review according to the present invention;
2 is a view showing the configuration of an example of a review analysis database building apparatus according to the present invention,
3 is a flowchart illustrating an example of a method of building a review analysis database according to the present invention;
4 is a view showing an example of a content recommendation process using a user review according to the present invention;
5 is a diagram illustrating an example of a configuration of a content recommendation apparatus according to the present invention;
6 is a flowchart illustrating an example of a content recommendation method according to the present invention.

Hereinafter, an apparatus and method for recommending content using a user review according to the present invention will be described in detail with reference to the accompanying drawings.

Synopsis is mainly used for movies when comparing similarities between contents, but synopsis includes specific story lines or abstract metaphors for the purpose of indicating what the movie is about. Since synopsis is distributed to film producers in a batch, it is difficult to find other movies with synopsis of the same contents, but it is possible to search for movies with similar viewing feelings or reviews.

For example, the synopsis of the Wine Miracle (2008, Bottle Shock) movie is as follows. "A vineyard in California dreaming of the best wines to stay in history, Jim, a proud farmer, and a spoiled son," Bo. " One day, a lively and beautiful lady, Sam, dreams of becoming a wine master and visits the farm. "Bo" dreams of the last hope, but the white wine, which should be golden, turns to brown just before entry. "

Here's a review of a `` Wine Miracle '' movie downloaded from a review site on the Internet: "Awesome movie, a wine miracle movie, a European American movie, a calm movie, a cute movie".

In other words, the review is a short text composed of a few words that describes the user's view of various contents and the like. Thus, a review is a general review of the movie rather than a plot when the content is a movie.

Therefore, it is difficult to search for another movie with a plot similar to the one described in Synopsis, but a review can search for movies with similar inspiration. For example, the "Wine Miracle" movie is classified as a drama when referring to metadata in the VOD DB, but when searching based on metadata only, it is a sad or cool impression like this drama among other dramas. I don't know at all whether it is unfortunate. However, using reviews from other users is more practical because you can search for 'Vitus' or 'Amazing Grace' movies with similar content.

The present invention mainly consists of constructing a review DB in order to use a user review, and a configuration of searching for and recommending content using the constructed review DB. Let's look at them in turn.

1 is a diagram illustrating an example of a process of databaseting a user review according to the present invention.

Referring to FIG. 1, since there is a possibility that other content (VOD, etc.) of the same title may exist, the metadata (eg, title, actor, producer, etc.) for the corresponding content is retrieved through the VOD DB 100. Based on this, the movie review site 110 searches for and downloads the corresponding content. For example, when searching for reviews, a desired review may be downloaded in order of title, production year, producer, and actor. Here, the movie review site for searching for reviews may be various sites where user reviews of various contents exist on the Internet. When a review for the corresponding content is downloaded, a unique identifier may be assigned to each content to facilitate searching, and then the review may be searched by the identifier. Of course, a review analysis database 120 may be constructed to search for related content as a review.

The downloaded reviews are analyzed using natural language processing and statistical analysis. In general, natural language processing techniques are used to extract key words from reviews, and statistical-based analysis methods are used to convert extracted key words into feature vectors for efficient comparison of text. Key words extracted from reviews are words used to summarize the contents of multi-word text or compare it with other text.

The keyword-based text processing method determines key words by comparing the frequency of each word. Therefore, the most frequent words among repeated words are used when indexing DB of text. Indexing is essential for developing text-based information retrieval techniques, and efficient indexing is required to provide fast and accurate retrieval services. For example, when a review of a movie called "Wine Miracle" is called "Awesome Movies, Softening of Wine Miracles, European American Movies, Gentle Movies, Cute Movies", "Movies" is the most mentioned, Key words are not enough. On the contrary, words such as 'cool', 'miracle of wine' or 'tranquil' such as modifier are suitable as key words to be stored in review DB.

In addition, keyword-based text processing is not suitable for comparing short-length text because it can only be analyzed at the surface level of the text. Unlike web documents, short documents like reviews are made up of a few keywords, so it's unlikely that two texts contain the same keywords. For example, 'wonderful imagination' and 'excellent creativity' are reviews of the same content, but in keyword-based processing methods, there is no common keyword and thus cannot be compared. However, using natural language processing techniques, the analysis is based on the meaning of keywords, so that 'imagination' and 'creativeness' are synonymous words and searched for the same contents. When searching for synonyms, the same word can represent multiple meanings, so we have to analyze what the particular word is used in the current text. Therefore, synonym extraction uses natural language processing that can be deep-level. Another disadvantage of the keyword base is that it can't determine whether two reviews are the same or opposite. For example, 'ingenious imagination' and 'low imagination' are opposite reviews, but searching with the keyword imagination extracts the two reviews in the same way.

Therefore, the present invention uses a natural language processing technique rather than a keyword-based text search. Natural language processing techniques automate language activities between people so that computers can access them in more natural ways when interacting with them. Currently, natural language processing techniques have been developed to process such fractions with high accuracy and high speed, and are applied to bio or web document analysis. For example, the computer not only recognizes the voice, but also enables the user to interact with the person with a cognitive context. In order to use such a conversational ability in real life, a large amount of language resources and complex processing techniques are required, and the probability of error occurrence is high. Analyze through the process.

Step 1: tokenization

Divide a word into tokens, which are processing units. For example, in the text 'Ingenious Imagination', separate it into two tokens: 'Ingenious' and 'Imagination'. Tokenization divides each token using spaces between words or special characters (eg,-,?).

Step 2: Syntactic analysis

Determine the parts of speech of each tokenized word. For example, determine the parts of speech of each tokenized word, such as 'incidental / adj (adjective)' or 'imagination / nn (noun)'. Rather than relying solely on grammar rules, this parsing builds up parsing formats for commonly used sentences by analyzing a large amount of text in advance, and then parses the structure that is most similar to the text currently being analyzed by a statistical method. Determine the format and parse the syntax.

Step 3: Typed Denpendency Analysis

When parts of tokenized words are determined, the relationship between the words is analyzed. For example, 'quirky' and 'imagination' analyzed above are qualifier relations in which 'quirky' words modify 'imagination'. Therefore, we use amod, which means that it is a modifier relationship, to analyze the word relationship with amod ('imagination', 'wonder'). This word relationship analysis extracts the relationship between words in a sentence using predetermined grammatic relations. According to an embodiment, dozens of word relationships may be defined in advance.

For example, the review of 'Funny Movie' is first divided into 'Token' and 'Film' in Step 1, and the parts of speech are determined in Step 2, and then the word relationship analysis process is performed. It can be analyzed as neg (movie, not funny ') using neg, which means that an expression of negation is included.

Review analysis is used to extract morphological conversions (eg imagination => imagination), synonym mapping (eg imagination vs creativity), and extract key words (eg 'quite imagination' => imagination). In particular, morphology conversion and synonym mapping are useful for converting various reviews into a given format.

FIG. 2 is a diagram illustrating an example of a configuration of an apparatus for constructing a review analysis DB according to the present invention for forming a review database of FIG. 1, and FIG. 3 is a flowchart illustrating an example of a method for constructing a review analysis DB. Hereinafter, a description will be given with reference to FIGS. 2 and 3.

First, referring to FIG. 2, the apparatus for constructing a review analysis DB includes a review search unit 200 connected to a review site 230, a review analyzer 210, a VOD DB 220, and a review analysis DB 240.

The VOD DB 220 is a database that stores various contents and metadata thereof (title, actor, producer, producer, synopsis, production year, genre, etc.), and the review analysis DB 240 stores the analyzed reviews. Database.

The review search unit 200 searches for and downloads reviews of various contents to the review site 230 (S300). In this case, since a plurality of contents of the same title may exist, the search search unit 200 first searches the metadata of the corresponding content through the VOD DB 220 and then reviews the site based on the metadata instead of simply searching for the content title alone. Search for and download the review of the corresponding content (230).

The review analyzing unit 210 analyzes the reviews of each content downloaded through the review searching unit 200 through the natural language processing technique described in FIG. 1 and extracts key words (S310). That is, the review analyzing unit 210 separates the review text of each content by each token unit, determines a part-of-speech, analyzes the relationship of each word, extracts key words, and stores them in the review analysis DB 240. In this case, the review analyzer 210 may store the review in a predetermined format through synonym mapping and morphology conversion of core words analyzed using a dictionary or a synonym DB as necessary.

More than one review may be searched and downloaded in one content, and the review analyzer 210 may give more weight to reviews that are more different from other content. This weighting uses statistical analysis. That is, the review analysis unit 210 extracts key words for the analyzed review and checks the frequency in the entire review analysis DB 240 so that words with high frequency are given low weights and low words with high weights to give other Allow content and differentiation to weight more reviews.

The review analysis unit 210 stores the features of each review identified through review analysis, key word extraction, weighting, etc. in the form of a feature vector to facilitate later search and analysis of similar reviews. Of course, the VOD DB 220 is also preferably stored in the form of a feature vector. The feature vector is a vector expressing n features representing a specific object in n dimensions and is widely used for comparing features of an object, which is outside the scope of the present invention, and thus a detailed description thereof will be omitted.

4 is a diagram illustrating an example of a content recommendation process using a user review according to the present invention.

Referring to FIG. 4, a user searches a VOD DB 410 based on metadata 400 of content currently being viewed, searches for content having similar characteristics, and also based on a review 420 of content currently being viewed. By searching the review analysis DB (430) to search for content with similar features. The similarity is determined by integrating the search results based on metadata and reviews, but the similarity of the two search results is integrated by giving a higher weight to the search results based on the reviews. In addition, the content having the integrated similarity between the two search results or more is set as the recommended content 440. As such, when a user recommends content similar to the content being viewed, statistically, 25% of users purchase the recommended content within 24 hours.

FIG. 5 is a diagram illustrating an example of an apparatus for a content recommendation process illustrated in FIG. 4, and FIG. 6 is a flowchart illustrating an example of a method thereof. Hereinafter, a description will be given with reference to FIGS. 5 and 6.

Referring to FIG. 5, the content recommendation apparatus includes a review analysis DB 500, a VOD DB 510, a content search unit 520, and a content recommendation unit 530.

The review analysis DB 500 is a database that analyzes reviews for each content through the review analysis process described with reference to FIGS. 1 to 3, extracts and stores key words, and analyzes the results in a feature vector form using a statistical analysis method. Save it. The VOD DB 510 stores metadata about various contents in the form of a feature vector.

The content retrieval unit 520 grasps metadata and reviews about the content currently being viewed by the user, and then searches for the VOD DB 510 and the review analysis DB 500 based on this to identify similar contents (S600 and S610). )). The content search unit 520 may search for metadata and reviews about the content currently being viewed in the VOD DB 500 and the review analysis DB 510 or download the data from a real-time review site to search and analyze the reviews. The content retrieval unit 520 uses the VOD DB 510 based on the feature vector for the metadata and the review of the content currently being viewed in which the VOD DB 510 and the review analysis DB 500 store data in the form of a feature vector. And similar content is searched in comparison with the feature vector of the review analysis DB 500.

The content recommender 530 integrates the similarity results searched by the content searcher 520 based on the metadata and the reviews, respectively, and gives the weighted results to the searched results based on the review, thereby integrating the two results. S620). The content recommendation unit 530 provides the recommended content when the integrated similarity of the search results of the metadata and the review is a predetermined threshold value (S630). In this case, when a plurality of contents having a predetermined threshold or more exist, a list of recommended contents is created in order of high similarity and provided to the user.

The present invention can also be embodied as computer-readable codes on a computer-readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disks, optical data storage devices, and the like. The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

So far I looked at the center of the preferred embodiment for the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the appended claims rather than by the foregoing description, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

Claims (14)

A review search unit searching for at least one review text corresponding to each content based on metadata; And
And a review analyzing unit extracting the searched review texts through a natural language analysis technique and storing the searched words in a review analysis database. The method of claim 1, wherein the review analysis unit,
A review analysis database construction apparatus comprising: separating review text for each word as a processing unit, determining a part-of-speech for each word, analyzing the relationship between each word, and extracting the key words and storing them in the form of a feature vector. The method of claim 2, wherein the review analysis unit,
And a parsing format constructed by analyzing a large amount of text, and parsing the review text, and analyzing a relationship between words using a grammatical relationship. The method of claim 1, wherein the review analysis unit,
A review analysis database building apparatus, characterized in that a low weight is assigned to a high frequency key word and a high weight is given to a low frequency key word based on the frequency of the key words in the review analysis database. In the method for the content recommendation device to build a review analysis database,
Searching for and downloading at least one review text corresponding to each content based on the metadata from the review site;
And a review analysis step of extracting and storing the key words through the natural language analysis technique of the downloaded review text. The method of claim 5, wherein the review analysis step,
Separating the downloaded review text by each word through a natural language analysis technique, determining a part-of-speech of each word, analyzing the relationship between each word, and extracting the core words and storing them in the form of a feature vector; How to build an analytics database. The method of claim 6, wherein the review analysis step,
Parsing the review text using a parsing format constructed by analyzing a large amount of text, and analyzing a relationship between words using a grammatical relationship. The method of claim 5, wherein the review analysis step,
And assigning low weights to key words with high frequency and high weights to key words with low frequency based on the frequency of key words in the review analysis database. A review analysis database for storing review data of each content;
A content searching unit searching for contents having review data similar to review data of the content currently being viewed in the review analysis database; And
And a content recommending unit for recommending the searched contents to the user. The method of claim 9,
The content searching unit compares the similarity between the review data for the content currently being viewed in the review analysis database and the review data for the other contents through a natural language processing technique, and searches for the content having the similarity level more than a predetermined value. Content recommendation device. The method of claim 9,
It further includes a VOD database that stores metadata about each content,
The content search unit searches for the similar content by searching the review analysis database and the VOD database based on review data and metadata for the content currently being viewed, but the search results through the review analysis database are provided through the VOD database. Content recommending device, characterized in that the weight higher than the search results. Searching for a content having review data similar to the review data of the content currently being viewed in a review analysis database storing review data of each content; And
Recommending the retrieved content to a user. The method of claim 12, wherein the searching comprises:
And comparing the similarity of the review data for the content currently being viewed in the review analysis database with the review data for the other contents through a natural language processing technique, and searching for the content having the similarity equal to or more than a predetermined value. How to recommend content. The method of claim 12, wherein the searching comprises:
A first search step of searching for similar content by comparing the review data of the content currently being viewed and the review data of the review analysis database;
A second search step of searching for similar content by comparing metadata about the content currently being viewed and metadata about each content stored in the VOD database;
Adding two search results after assigning a weight to a result of the first search step higher than a result of the second search step; And
And extracting content whose summed result is equal to or greater than a predetermined threshold value.

Images