Disclosure of Invention
The embodiment of the application provides a song recommending method, a device, terminal equipment and a storage medium, which can improve the accuracy of recommending songs for new users and have high applicability.
In a first aspect, an embodiment of the present application provides a song recommendation method, including:
Obtaining a search object content text corresponding to a search record of a target user within a preset time period, and obtaining a content text feature vector corresponding to the search object content text, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the search object content text;
Obtaining a song text feature vector corresponding to each song included in a song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information;
And determining the text similarity between the content text of the search object and the song information of each song based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, and determining at least one song from the song library to recommend to the target user according to the text similarity.
With reference to the first aspect, in a possible implementation manner, the acquiring the search object content text corresponding to the search record of the target user in the preset time period includes:
Acquiring a log file corresponding to a search engine used by the target user, wherein the log file comprises a search record of the target user, and the search record comprises a search string used for searching by the target user and search time;
acquiring a search character string with the largest occurrence number in a search record of a target user in a preset time period, determining a plurality of search result content texts based on the search character string, and acquiring a webpage attribute feature vector corresponding to each search result content text;
Acquiring a search character text feature vector corresponding to the search character string, acquiring a user basic attribute feature vector of the target user, and generating a joint feature vector corresponding to each search result content text based on the search character text feature vector, the user basic attribute feature vector and a web page attribute feature vector corresponding to each search result content text;
obtaining a text ordering model, and inputting each joint feature vector into the text ordering model to obtain an ordering result of each search result content text output by the text ordering model, wherein the text ordering model is obtained by training according to a plurality of sample joint feature vectors corresponding to a plurality of sample texts and the ordering result of the plurality of sample texts;
and determining the first n search result content texts in the sorting results as search object content texts.
With reference to the first aspect, in a possible implementation manner, the search record includes a browsing record and a browsing time of the target user; the obtaining the search object content text corresponding to the search record of the target user within the preset time period includes:
and acquiring a Uniform Resource Locator (URL) included in the browsing record in a preset time period, and accessing the URL to acquire a corresponding content text as a search object content text.
With reference to the first aspect, in a possible implementation manner, the method further includes:
Word segmentation processing is carried out on the search object content text to obtain a plurality of words forming the search object content text;
Acquiring a preset word vector lookup table, wherein the word vector lookup table comprises a plurality of word vectors corresponding to a plurality of words, and one word corresponds to one word vector;
and determining word vectors corresponding to each word in the plurality of words composing the text of the search object content from the word vector lookup table.
With reference to the first aspect, in a possible implementation manner, the determining, based on the content text feature vector and a song text feature vector corresponding to each song included in the song library, a text similarity between the search object content text and song information of each song includes:
a text similarity classification model is obtained, the content text feature vector and the song text feature vector corresponding to any song are input into the text similarity classification model, wherein the text similarity classification model is obtained through training according to the content text feature vector corresponding to the sample search object content text, the song text feature vector corresponding to the sample song and the text similarity classification result identification;
And acquiring a text similarity classification result identifier output by the text similarity classification model, and determining the text similarity between the song text information corresponding to any song and the text of the search object content according to the text similarity classification result identifier.
With reference to the first aspect, in a possible implementation manner, the determining, based on the content text feature vector and a song text feature vector corresponding to each song included in the song library, a text similarity between the search object content text and song information of each song includes:
Calculating Euclidean distance between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song in the song library respectively;
And converting the Euclidean distance into a similarity value to serve as the text similarity between the text of the content of the search object and the song information of each song.
With reference to the first aspect, in a possible implementation manner, the determining, according to the text similarity, at least one song recommendation from the library to the user includes:
Acquiring a preset similarity threshold value, and acquiring song recommendation with text similarity not smaller than the preset similarity threshold value from the song library to the target user; or alternatively
And descending the text similarity, and obtaining k songs corresponding to the first k text similarity after descending the sequence to be recommended to the target user, wherein k is an integer greater than 0.
In a second aspect, an embodiment of the present application provides a song recommendation apparatus, including:
the first feature vector acquisition module is used for acquiring a search object content text corresponding to a search record of a target user in a preset time period and acquiring a content text feature vector corresponding to the search object content text, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the search object content text;
The second feature vector obtaining module is used for obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words forming song information;
And the song recommendation module is used for determining the text similarity between the content text of the search object and the song information of each song based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, and determining at least one song from the song library to recommend to the target user according to the text similarity.
With reference to the second aspect, in one possible implementation manner, the first feature vector obtaining module includes a first search object content text obtaining unit, a content text feature vector obtaining unit, where the first search object content text obtaining unit includes:
A log file obtaining subunit, configured to obtain a log file corresponding to a search engine used by the target user, where the log file includes a search record of the target user, and the search record includes a search string used by the target user for searching and a search time;
The search result content text acquisition subunit is used for acquiring a search character string with the largest occurrence number in a search record of a target user in a preset time period, determining a plurality of search result content texts based on the search character string, and acquiring a webpage attribute feature vector corresponding to each search result content text;
The joint feature vector obtaining subunit is used for obtaining the search character text feature vector corresponding to the search character string, obtaining the user basic attribute feature vector of the target user, and generating the joint feature vector corresponding to each search result content text based on the search character text feature vector, the user basic attribute feature vector and the webpage attribute feature vector corresponding to each search result content text;
the text sorting subunit is used for acquiring a text sorting model, inputting each joint feature vector into the text sorting model to obtain a sorting result of each search result content text output by the text sorting model, wherein the text sorting model is obtained by training according to a plurality of sample joint feature vectors corresponding to a plurality of sample texts and the sorting result of the plurality of sample texts;
and the sequencing result processing subunit is used for determining the first n search result content texts in the sequencing results as search object content texts.
With reference to the second aspect, in one possible implementation manner, the search record includes a browsing record of the target user, and the first feature vector acquisition module further includes a second search object content text acquisition unit, where the second search object content text acquisition unit is specifically configured to:
and acquiring a Uniform Resource Locator (URL) included in the browsing record in a preset time period, and accessing the URL to acquire a corresponding content text as a search object content text.
With reference to the second aspect, in a possible implementation manner, the apparatus further includes a word vector acquisition module, where the word vector acquisition template includes:
The word segmentation processing unit is used for carrying out word segmentation processing on the search object content text to obtain a plurality of words forming the search object content text;
The word vector query table acquisition unit is used for acquiring a preset word vector query table, wherein the word vector query table comprises a plurality of word vectors corresponding to a plurality of words, and one word corresponds to one word vector;
And the word vector determining unit is used for determining word vectors corresponding to each word in the plurality of words composing the text of the search object content from the word vector lookup table.
With reference to the second aspect, in one possible implementation manner, the song recommendation module includes a first text similarity determining unit, where the first text similarity determining unit includes:
The classifying model obtaining subunit is used for obtaining a text similarity classifying model, inputting the content text feature vector and the song text feature vector corresponding to any song into the text similarity classifying model, wherein the text similarity classifying model is obtained by training according to the content text feature vector corresponding to the sample searching object content text, the song text feature vector corresponding to the sample song and the text similarity classifying result identification;
and the result identification processing subunit is used for acquiring the text similarity classification result identification output by the text similarity classification model, and determining the text similarity between the song text information corresponding to any song and the text of the search object content according to the text similarity classification result identification.
With reference to the second aspect, in one possible implementation manner, the song recommendation module includes a second text similarity determining unit, where the second text similarity determining unit includes:
The distance determining unit is used for calculating Euclidean distances between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song in the song library respectively;
And the distance conversion unit is used for converting the Euclidean distance into a similarity value to be used as the text similarity between the text of the search object content and the song information of each song.
With reference to the second aspect, in one possible implementation manner, the song recommendation module includes a recommended song determining unit, where the recommended song determining unit is specifically configured to:
Acquiring a preset similarity threshold value, and acquiring song recommendation with text similarity not smaller than the preset similarity threshold value from the song library to the target user; or alternatively
And descending the text similarity, and obtaining k songs corresponding to the first k text similarity after descending the sequence to be recommended to the target user, wherein k is an integer greater than 0.
In a third aspect, an embodiment of the present application provides a terminal device, where the terminal device includes a processor and a memory, and the processor and the memory are connected to each other. The memory is configured to store a computer program supporting the terminal device to perform the method provided by the first aspect and/or any of the possible implementation manners of the first aspect, the computer program comprising program instructions, the processor being configured to invoke the program instructions to perform the method provided by the first aspect and/or any of the possible implementation manners of the first aspect.
In a fourth aspect, embodiments of the present application provide a computer readable storage medium storing a computer program comprising program instructions which, when executed by a processor, cause the processor to perform the method provided by the first aspect and/or any of the possible implementations of the first aspect.
In the embodiment of the application, the search record of the target user in the preset time period is obtained, so that the corresponding search object content text can be obtained according to the search record. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. Based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, the text similarity between the content text of the search object and the song information of each song can be determined, and then at least one song can be determined from the song library according to the text similarity and recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved.
Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
The song recommendation method provided by the embodiment of the application can be widely applied to the terminal equipment for song recommendation, wherein the terminal equipment can be hardware or software. When the terminal device is hardware, it may be various electronic devices having a display screen, including but not limited to a server, a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; when the terminal device is software, it may be installed in the above-listed electronic device, and is not limited thereto. For convenience of description, the following embodiments of the present application will be described by taking a smart phone as an example, and for convenience of description, the smart phone is simply referred to as a mobile phone. According to the method provided by the embodiment of the application, the search record of the target user in the preset time period is obtained, and the corresponding search object content text can be obtained according to the search record. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. Based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, the text similarity between the content text of the search object and the song information of each song can be determined, and then at least one song can be determined from the song library according to the text similarity and recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved.
The method and the related device according to the embodiments of the present application will be described in detail below with reference to fig. 1 to 8, respectively. The method provided by the embodiment of the application can comprise data processing stages for acquiring the content text of the search object, acquiring the content text feature vector, acquiring the song text feature vector, determining the text similarity based on the content text feature vector and the song text feature vector, recommending songs based on the text similarity and the like. The implementation of the above-mentioned individual data processing phases can be seen from the following implementation shown in fig. 1 and 5.
Referring to fig. 1, fig. 1 is a flow chart of a song recommendation method according to an embodiment of the present application. The method provided by the embodiment of the application can comprise the following steps 101 to 103:
101. and acquiring a content text of a search object corresponding to the search record of the target user in a preset time period, and acquiring a content text feature vector corresponding to the content text of the search object.
In some possible embodiments, the target user may enjoy different services based on various types of APP downloaded by downloading various Applications (APPs) in the handset. Currently, commonly used APPs mainly include shopping APP, entertainment APP, social APP, other types of APPs, and the like. The entertainment APP mainly comprises a music APP, a video APP, a game APP and a reading APP, other types of APP comprise a search engine APP and the like, and for convenience of description, the embodiment of the application mainly takes song recommendation for a target user in the music APP as an example for explanation. Here, the target user may be any directional object of music recommendation, for example, if the current music recommendation is for the new user a, the target user is the new user a at this time; if the music recommendation is for the new user B, the target user is the new user B. Wherein the new user may be a new registered user or a user without any usage behavior on the music class APP, etc. At present, the same user account can be used for logging in among many APPs or logging in through an authorized third party, so it is easy to understand that when different APPs are logged in by using the same user account or when the third party is authorized for logging in, the application program can obtain user data of the same user account on other application programs, such as user detailed information and the like, after logging in, and the application program is not limited herein. In the embodiment of the application, the target user is mainly taken as an example to log in the music APP and the search engine APP by using the same user account.
In some possible embodiments, by acquiring the content text of the search object corresponding to the search record of the target user within the preset time period, a content text feature vector corresponding to the content text of the search object may be acquired according to the content text of the search object, where the content text feature vector includes a plurality of word vectors corresponding to a plurality of words that compose the content text of the search object. Specifically, a log file corresponding to a search engine used by a target user can be obtained based on an application package name (i.e., a package name of a search engine class APP), wherein the log file includes a search record of the target user, an operation state of an application, and the like. Further, the search record includes a search string for searching by the target user, search time, and the like. It is to be understood that, by acquiring the search string with the largest occurrence number in the search record of the target user within the preset period, a plurality of search result content texts can be determined based on the search string. That is, by re-inputting the search string into the search engine, a plurality of search result content texts returned by the search engine can be obtained, and further, the web page attribute feature vector corresponding to each search result content text can be obtained. Generally, in a search scenario, the text of the search result content presented by the search engine to the user is typically a web page, and thus, the web page attribute feature vector may include, without limitation, a web page structure, a domain name, text information of the search result content text, and the like.
The search character text feature vector corresponding to the search character string and the user basic attribute feature vector of the target user are obtained, and the joint feature vector corresponding to each search result content text can be generated based on the search character text feature vector, the user basic attribute feature vector and the webpage attribute feature vector corresponding to each search result content text. The search character text feature vector is a vector corresponding to text information of the search character string, for example, after word segmentation is performed on the search character string, a plurality of word vectors corresponding to a plurality of words after word segmentation can be obtained, and the search character text feature vector is generated according to the plurality of word vectors corresponding to the plurality of words. The basic attribute feature vector of the user is a feature vector corresponding to a basic attribute of the target user, for example, the basic attribute of the target user includes age, gender, academic, post, residence city, and the like, which is not limited herein. The basic attribute feature vector of the target user can be obtained by digitizing and/or vectorizing the basic attribute of the target user. It should be appreciated that the manner in which the joint feature vector corresponding to any one of the search result content texts is generated may include, without limitation, expanding, stitching or summing the corresponding search character text feature vector, the user basic attribute feature vector, and the web page attribute feature vector corresponding to any one of the search result content texts, where one of the search result content texts corresponds to one of the joint feature vectors.
Further, after the text ranking model is obtained, the joint feature vector corresponding to each search result content text in the plurality of search result content texts can be input into the text ranking model, so that the ranking result of each search result content text can be output according to the text ranking model. The text ordering model is trained according to a plurality of sample joint feature vectors corresponding to the plurality of sample texts and ordering results of the plurality of sample texts. It should be appreciated that in the embodiment of the present application, the top n search result content texts among the outputted ranking results may be determined as search object content texts. That is, the first n search result content texts may be combined, and the combined text may be determined as the search object content text.
Alternatively, in some possible embodiments, after determining, based on the search string, a plurality of search result content texts displayed by the search engine to the user, the first n search result content texts displayed in the display interface of the search engine may also be directly determined as the search object content texts. Alternatively, in some possible embodiments, the search record further includes a browsing record and browsing time of the target user, so by acquiring a uniform resource locator (Uniform Resource Locator, URL) included in the browsing record within a preset period of time, the URL may be accessed to acquire the corresponding content text as the search target content text. It should be understood that the browse record herein is a click record of the target user, for example, assuming that the list of related documents (in the search scenario, web pages in general) presented to the user by the search engine is y= { Y1, Y2, …, ym }, m is an integer greater than 0, where Y3 and Y5 may be determined as search object content text if the user clicks on Y3 and Y5.
In some possible embodiments, after determining the content text of the search object, if the content text feature vector corresponding to the content text of the search object is preset, the corresponding content text feature vector may be queried directly based on the preset identifier of the content text of the search object. For example, assuming that the preset identifier of the search object content text is a hash value of the search object content text, after the search object content text is acquired, hash calculation may be performed on the acquired search object content text to obtain the hash value of the search object content text, and then, through the hash value, a content text feature vector corresponding to the hash value may be queried from a preset content text feature vector lookup table.
Alternatively, in some possible embodiments, the word segmentation may be performed on the search object content text to obtain a plurality of words that form the search object content text. It should be understood that the embodiment of the present application may perform word segmentation on the text of the search object content based on word segmentation tools, so as to obtain a plurality of words that constitute the text of the search object content, where the word segmentation tools include, but are not limited to jieba, standardAnalyzer, chineseAnalyzer, CJKAnalyzer, IKAnalyzer, paoding, imdict, and the like, and are not limited thereto. For example, assuming that the search object content text is "i like Zhou Jielun", the search object content text "i like Zhou Jielun" is segmented to obtain "i", "like", "Zhou Jielun". Then, a plurality of word vectors corresponding to a plurality of words are obtained from the word vector lookup table by obtaining a preset word vector lookup table, wherein one word corresponds to one word vector. Therefore, the word vectors corresponding to each word in the words constituting the search object content text are searched from the word vector lookup table, so that the word vectors can be obtained, and further the content text word vector corresponding to the search object content text can be generated according to the word vectors.
It should be appreciated that the manner in which the content text feature vectors corresponding to the search object content text are generated based on the plurality of word vectors may include expansion, concatenation, summation, or the like, and is not limited herein. For example, assuming that the word vector corresponding to "me" is [1,2], "like" is [3,4], and "Zhou Jielun" is [5,6], the content text feature vector corresponding to the search target content text "me like Zhou Jielun" is [ [1,2], [3,4], and [5,6] ] can be obtained by expanding the plurality of word vectors. For another example, assuming that the word vector corresponding to "me" is [1,2], "like" is [3,4], "Zhou Jielun" is [5,6], the content text feature vector corresponding to the search object content text "me like Zhou Jielun" may also be expressed as a sum of a plurality of vectors, that is, the content text feature vector is [9,12]. For another example, assuming that the word vector corresponding to "me" is [1,2], "like" is [3,4], "Zhou Jielun" is [5,6], the content text feature vector corresponding to the search object content text "me like Zhou Jielun" may be a vector obtained by splicing a plurality of vectors, that is, the content text feature vector is [1,2,3,4,5,6]. For the purpose of description, the following embodiments of the present application will take a vector obtained by expanding a content text word vector as a plurality of word vectors as an example.
Alternatively, in some possible embodiments, the search object content text is segmented. The preset stop word list can also be obtained to obtain a stop word list comprising a plurality of stop words. It should be appreciated that in Search Engine Optimization (SEO), to save memory space and improve search efficiency, the search engine automatically ignores certain Words or Words, known as Stop Words, when indexing pages or processing search requests. Stop Words are somewhat equivalent to Filter Words (Filter Words), however, the Filter Words are more extensive, keywords containing sensitive information are treated as Filter Words, and the stop Words themselves are not limited. In general terms, the term "disuse" can be broadly divided into two categories: the first category is words that are used quite widely, even too frequently. Words such as "I", "just" appear on almost every document, and query such word search engines cannot guarantee that truly relevant search results can be given, so that the search scope is difficult to be reduced, the accuracy of the search results is improved, and the search efficiency is reduced; the second category is words that occur very frequently in text, but do not have great practical significance. For example, including mood aid words, adverbs, prepositions, conjunctions, etc., which are generally not themselves explicitly meaningful, words that have a certain effect only by putting them into a complete sentence. Common words such as "at", "and", "followed" and the like. Therefore, after the word segmentation processing is performed to obtain the words forming the text of the search object content, the words not belonging to the stop word list can be determined from the words forming the text of the search result, so as to be used as the words to be processed subsequently.
For example, assuming that the search object content text is "i liked Zhou Jielun very much", the word segmentation of the search object content text "i liked Zhou Jielun very much" may result in "i", "very", "like", "Zhou Jielun". Wherein the stop word list is obtained, it is assumed that the stop words in the stop word list include "I, very, most, too, slave, since, on, when" and the like. Thus, "like", "Zhou Jielun" can be obtained by deactivating the words by the plurality of words obtained after the word segmentation process described above. Then, a plurality of word vectors corresponding to a plurality of words are obtained from the word vector lookup table by obtaining a preset word vector lookup table, wherein one word corresponds to one word vector. The word vectors corresponding to each word in the plurality of words composing the search object content text are inquired from the word vector inquiry table, so that the plurality of word vectors can be obtained, and further the content text feature vector corresponding to the search object content text can be generated according to the plurality of word vectors.
102. And obtaining a song text feature vector corresponding to each song included in the song library.
In some possible embodiments, a song text feature vector corresponding to each song included in a song library of the music class APP is obtained, wherein the song text feature vector includes a plurality of word vectors corresponding to a plurality of words constituting song information. That is, for songs in the song library, text information (i.e., song information) of each song may be extracted, and generally, the text information of the song mainly includes 3-aspect information, i.e., lyric-related information, song style-related information, and name-related information. Wherein the lyric related information mainly comprises song names, lyric contents and the like; the related information of the song style mainly comprises the style name of the song, the hierarchical information (such as rock-heavy metal) of the category to which the song belongs and the like; the person name related information mainly includes singer name, word author name, song author name, and the like. After the text information of the song is extracted from the 3 dimensions, word segmentation processing can be performed on the text information of the song, and after vectorization is performed on each word obtained by the word segmentation processing, a text feature vector of the song corresponding to each song can be obtained. The method comprises the steps of obtaining a preset word vector lookup table, and then inquiring the word vector lookup table to determine a plurality of word vectors corresponding to a plurality of words forming song information from the word vector lookup table, so that a song text feature vector corresponding to the song information is generated according to the plurality of word vectors.
Alternatively, in some possible embodiments, if a text feature vector of a song corresponding to each song in the song library has been preset, the text feature vector of the corresponding song may be queried based on the song identifier of each song. The song identifier may be a song name, a singer name, or the like, or may also be a hash value corresponding to the song name and the singer name, or may also be a character string which is composed of numbers, letters, symbols, or the like and can uniquely mark a song, or the like, which is not limited herein.
103. And determining the text similarity between the content text of the search object and the song information of each song based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, and determining at least one song from the song library to recommend to the target user according to the text similarity.
In some possible embodiments, based on the text feature vector of the content and the text feature vector of the song corresponding to each song included in the song library, a text similarity between the content text of the search object and the song information of each song may be determined, and then, based on the text similarity between the content text of the search object and the song information of each song, at least one song may be determined from the song library to be recommended to the target user.
Referring to fig. 2, fig. 2 is a schematic view of a scenario of a song recommendation method provided in the present embodiment. As shown in fig. 2, according to the search string, a plurality of search result content texts corresponding to the search string can be matched from the text library, wherein the plurality of search result content texts corresponding to the search string are ranked in combination with the text ranking model, so that ranking results of each search result content text can be obtained, and the first n search result content texts in the ranking results are determined as search object content texts, so that a content text feature vector corresponding to the search object content text can be obtained. Further, by acquiring song information corresponding to each song included in the song library, a corresponding song text feature vector may be generated based on the song information, and then by inputting the content text feature vector and the song text feature vector corresponding to each song included in the song library into the text similarity classification model, it may be determined whether the search object content text is similar to the song information based on the output result of the text similarity classification model, where if it is determined that the search object content text is similar to the song information, the corresponding song may be recommended as a recommended song to the target user.
Specifically, by acquiring a text similarity classification model, a content text feature vector and a song text feature vector corresponding to any song can be input into the text similarity classification model, wherein the text similarity classification model is obtained by training according to the content text feature vector corresponding to the sample search object content text, the song text feature vector corresponding to the sample song and the text similarity classification result identification. The text similarity between the song text information corresponding to any song and the text of the search object content can be determined according to the text similarity classification result identification by acquiring the text similarity classification result identification output by the text similarity classification model. Text similarity classification models herein include, without limitation, BERT pre-trained language models (BidirectionalEncoder Representations from Transformer), GPT-2 models, XLnet models, word2vec models, and the like. For convenience of description, the embodiment of the present application will be described by taking the BERT pre-training language model as an example.
Referring to fig. 3, fig. 3 is a schematic diagram of an application scenario for determining text similarity according to an embodiment of the present application. As shown in fig. 3, assume that the text of the search object content is text a, and the song information corresponding to song 1 in the song library is text B. Let the text a= { a1, a2, …, ai }, where i is an integer greater than 0, ai is a word obtained by word segmentation of the text a, and the text b= { B1, B2, …, bj }, where j is an integer greater than 0, bj is a word obtained by word segmentation of the text B. It should be appreciated that assuming that the word vector corresponding to a certain word ai is Eai, the content text feature vector (i.e., word vector sequence) corresponding to text a may be represented as { Ea1,Ea2,…,Eai }, and the song text feature vector (i.e., word vector sequence) corresponding to text B may be represented as { Eb1,Eb2,…,Ebj }. The text similarity between the text A and the text B can be determined based on the text similarity classification result identification output by the BERT pre-training language model by inputting the content text feature vector { Ea1,Ea2,…,Eai } and the song text feature vector { Eb1,Eb2,…,Ebj } into the BERT pre-training language model. For example, if the text similarity classification result identifier includes an identifier 1 and an identifier 0, where 1 indicates similarity and 0 indicates dissimilarity. Assuming that the text similarity classification result output by the BERT pre-training language model in fig. 3 is identified as the identifier 1, it may be determined that the text similarity between the text a and the text B is similar, so song 1 may be recommended to the target user. As shown in fig. 3, in the input of the BERT pre-training language model, a CLS symbol is inserted in front of the text a, and the output vector corresponding to the symbol is used as a semantic representation of the whole text for text similarity classification. Two text inputs are also split by a SEP symbol.
Alternatively, in some possible embodiments, the input of the BERT pre-training language model may include text vectors and location vectors in addition to the content text feature vectors corresponding to the search object content text and the song text feature vectors corresponding to the song. The text vector value is automatically learned in the model training process, is used for describing global semantic information of the text, and is fused with semantic information of the single word/word. The location vector is used to represent the location of each word that forms the text of the search object content in the text of the search object content, or the location of each word that forms the song information in the song information, it should be understood that, because of the difference in semantic information carried by the words/words that appear in different locations of the text, for example, "i love you" and "i love me", the BERT pre-training language model may append a different vector to each word/word in a different location to distinguish.
For example, referring to fig. 4, fig. 4 is a schematic diagram of input vectors of a BERT pre-training language model provided by an embodiment of the present application. As shown in fig. 4, assume that the text of the search object content is text C, and the song information corresponding to song 2 in the song library is text D. Let the text c= { C1, C2, …, ci }, where i is an integer greater than 0, ci is a word obtained by word segmentation of the text C, the text d= { D1, D2, …, dj }, where j is an integer greater than 0, and dj is a word obtained by word segmentation of the text D. It should be understood that, assuming that the word vector corresponding to a certain word ci is Eci, the word vector sequence corresponding to the text C may be represented as { Ec1,Ec2,…,Eci }, and the word vector sequence corresponding to the text D may be represented as { Ed1,Ed2,…,Edj }. Assuming that the sequence of text vectors is { Esc1,Esc2,…,Esci,Esd1,Esd2,…,Esdi }, the sequence of position vectors is { Epc1,Epc2,…,Epci,Epd1,Epd2,…,Epdi }, the sum of word vectors, text vectors, and position vectors can be used as the input vector for the BERT pre-trained language model. As shown in fig. 4, in the input of the BERT pre-training language model, a CLS symbol is inserted in front of the text C, and the output vector corresponding to the symbol is used as a semantic representation of the whole text for text similarity classification. Two text inputs are also split by a SEP symbol.
In the embodiment of the application, the search record of the target user in the preset time period is obtained, so that the corresponding search object content text can be obtained according to the search record. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. Based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, the text similarity between the content text of the search object and the song information of each song can be determined, and then at least one song can be determined from the song library according to the text similarity and recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved.
Referring to fig. 5, fig. 5 is another flow chart of a song recommendation method according to an embodiment of the present application. The method provided by the embodiment of the present application may be illustrated by the implementation manner provided in the following steps 201 to 204:
201. And acquiring a content text of a search object corresponding to the search record of the target user in a preset time period, and acquiring a content text feature vector corresponding to the content text of the search object.
202. And obtaining a song text feature vector corresponding to each song included in the song library.
The specific implementation manner of the steps 201 to 202 may refer to descriptions of the steps 101 to 102 in the corresponding embodiment of fig. 1, and will not be described herein.
203. And calculating Euclidean distance between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song included in the song library, and converting the Euclidean distance into a similarity value to be used as the text similarity between the text of the content to be searched and the song information of each song.
In some possible embodiments, the text similarity calculation method is mainly divided into two categories, i.e., supervised and unsupervised. The supervised method is to use a supervised model such as a naive Bayes classifier to judge the text similarity or calculate the similarity. The non-supervision method is to directly calculate the distance or similarity between the texts by using methods such as euclidean distance, etc., wherein the common similarity calculation method includes euclidean distance, manhattan distance, minkowski distance, cosine similarity, etc., and the method is not limited herein. For convenience of description, the embodiment of the present application is mainly illustrated by taking the euclidean distance as an example. Specifically, by calculating euclidean distances between the content text feature vectors and the song text feature vectors corresponding to each song included in the song library, respectively, the euclidean distances can be converted into similarity values as text similarity between the search object content text and the song information of each song. For example, assuming that the content text feature vector a= { a1, a2,..am }, the song text feature vector b= { B1, B2,..and bm }, which corresponds to the search object content text, the euclidean distance D between the content text feature vector and the song text feature vector, which corresponds to the song 1, can be calculated based on equation 1:
After the euclidean distance between the text feature vector of the content and the text feature vector of the song corresponding to the song 1 is calculated, the euclidean distance may be converted into a similarity value, for example, the euclidean distance may be converted into the similarity value based on the formula 2:
In addition to the conversion formula of the above formula 2, the formula for converting the euclidean distance into the similarity value may also be defined in different ways according to different requirements, which is not limited herein.
204. And obtaining a preset similarity threshold value, and obtaining songs with text similarity not smaller than the preset similarity threshold value from a song library to be recommended to a target user.
In some possible embodiments, after euclidean distances between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song included in the song library are calculated, and the euclidean distances are converted into text similarity between the content text of the search object and the song information of each song, a preset similarity threshold value can be obtained, and then songs with the text similarity not smaller than the preset similarity threshold value are obtained from the song library so as to be recommended to the target user. Or the calculated text similarity is arranged in a descending order, so that k songs corresponding to the first k text similarities after the descending order are obtained to be recommended to a target user, wherein k is an integer greater than 0. Or the calculated text similarity is arranged in an ascending order, so that k songs corresponding to the k text similarity after the ascending order are obtained to be recommended to the target user, wherein k is an integer greater than 0.
In the embodiment of the application, the search record of the target user in the preset time period is obtained, so that the corresponding search object content text can be obtained according to the search record. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. The euclidean distance between the content text feature vector and the corresponding song text feature vector of each song included in the song library is calculated, so that the euclidean distance can be converted into a similarity value to serve as the text similarity between the content text of the search object and the song information of each song. And finally, acquiring at least one song with text similarity not smaller than the preset similarity threshold value from the song library by acquiring the preset similarity threshold value so as to be recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved, the operability is strong, and the applicability is high.
Referring to fig. 6, fig. 6 is a schematic structural diagram of a song recommendation apparatus according to an embodiment of the present application. The song recommending device provided by the embodiment of the application comprises:
A first feature vector obtaining module 31, configured to obtain a content text of a search object corresponding to a search record of a target user within a preset period of time, and obtain a feature vector of a content text corresponding to the content text of the search object, where the feature vector of the content text includes a plurality of word vectors corresponding to a plurality of words that compose the content text of the search object;
a second feature vector obtaining module 32, configured to obtain a song text feature vector corresponding to each song included in the song library, where the song text feature vector includes a plurality of word vectors corresponding to a plurality of words that compose song information;
And the song recommendation module 33 is configured to determine, based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, a text similarity between the search object content text and song information of each song, and determine, according to the text similarity, that at least one song is recommended to the target user from the song library.
Referring to fig. 7, fig. 7 is a schematic diagram of another structure of a song recommendation apparatus according to an embodiment of the present application, in which:
In some possible embodiments, the first feature vector obtaining module 31 includes a first search object content text obtaining unit 311 and a content text feature vector obtaining unit 312, where the first search object content text obtaining unit 311 includes:
A log file obtaining subunit 3111, configured to obtain a log file corresponding to a search engine used by the target user, where the log file includes a search record of the target user, and the search record includes a search string used by the target user for searching and a search time;
A search result content text obtaining subunit 3112, configured to obtain a search string with the largest occurrence number in a search record of a target user within a preset period, determine a plurality of search result content texts based on the search string, and obtain a web attribute feature vector corresponding to each search result content text;
A joint feature vector obtaining subunit 3113, configured to obtain a search character text feature vector corresponding to the search string, and obtain a user basic attribute feature vector of the target user, and generate a joint feature vector corresponding to each search result content text based on the search character text feature vector, the user basic attribute feature vector, and a web page attribute feature vector corresponding to each search result content text;
The text sorting subunit 3114 is configured to obtain a text sorting model, input each joint feature vector into the text sorting model to obtain a sorting result of each text of the search result content output by the text sorting model, where the text sorting model is obtained by training according to a plurality of sample joint feature vectors corresponding to a plurality of sample texts and the sorting result of the plurality of sample texts;
The ranking result processing subunit 3115 is configured to determine the top n search result content texts in the ranking results as search target content texts.
In some possible implementations, the search record includes a browsing record of the target user, and the first feature vector obtaining module 31 further includes a second search object content text obtaining unit 313, where the second search object content text obtaining unit 313 is specifically configured to:
and acquiring a Uniform Resource Locator (URL) included in the browsing record in a preset time period, and accessing the URL to acquire a corresponding content text as a search object content text.
With reference to the second aspect, in a possible implementation manner, the apparatus further includes a word vector acquisition module 34, where the word vector acquisition module 34 includes:
A word segmentation processing unit 341, configured to perform word segmentation processing on the search object content text to obtain a plurality of words that compose the search object content text;
a word vector lookup table obtaining unit 342, configured to obtain a preset word vector lookup table, where the word vector lookup table includes a plurality of word vectors corresponding to a plurality of words, and one word corresponds to one word vector;
the word vector determining unit 343 is configured to determine, from the word vector lookup table, a word vector corresponding to each word in the plurality of words that compose the text of the search target content.
In some possible embodiments, the song recommendation module 33 includes a first text similarity determining unit 331, and the first text similarity determining unit 331 includes:
a classification model obtaining subunit 3311, configured to obtain a text similarity classification model, and input the content text feature vector and a song text feature vector corresponding to any song into the text similarity classification model, where the text similarity classification model is obtained by training according to a content text feature vector corresponding to a sample search object content text, a song text feature vector corresponding to a sample song, and a text similarity classification result identifier;
And a result identifier processing subunit 3312, configured to obtain a text similarity classification result identifier output by the text similarity classification model, and determine, according to the text similarity classification result identifier, a text similarity between the song text information corresponding to the arbitrary song and the search object content text.
In some possible embodiments, the song recommendation module 33 includes a second text similarity determining unit 332, and the second text similarity determining unit 332 includes:
a distance determining unit 3321, configured to calculate euclidean distances between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song included in the song library, respectively;
and a distance conversion unit 3322 for converting the euclidean distance into a similarity value as a text similarity between the text of the search object content and the song information of each song.
In some possible embodiments, the song recommendation module 33 includes a recommended song determining unit 333, and the recommended song determining unit 333 is specifically configured to:
Acquiring a preset similarity threshold value, and acquiring song recommendation with text similarity not smaller than the preset similarity threshold value from the song library to the target user; or alternatively
And descending the text similarity, and obtaining k songs corresponding to the first k text similarity after descending the sequence to be recommended to the target user, wherein k is an integer greater than 0.
In a specific implementation, the song recommendation apparatus may execute the implementation provided by each step in fig. 1 and fig. 5 through each function module built in the song recommendation apparatus. For example, the first feature vector obtaining module 31 may be configured to perform the above-mentioned steps to obtain the content text of the search object, obtain the feature vector of the content text, and the like, and the detailed description of the implementation provided by the above-mentioned steps will be omitted herein. The second feature vector obtaining module 32 may be configured to perform the implementation manner described in the related steps of obtaining the text feature vector of the song corresponding to each song in the above steps, and specifically, the implementation manner provided in the above steps may be referred to, which is not described herein again. The song recommendation module 33 may be configured to determine the text similarity based on the content text feature vector and the song text feature vector in the above steps, and determine the recommended song according to the text similarity, which is specifically referred to the implementation provided in the above steps and will not be described herein. The word vector obtaining template 34 may be used to perform the implementation manners of word segmentation on the text of the search object content in the above steps, obtaining a word vector lookup table, obtaining word vectors corresponding to the words after word segmentation, and the like, and specifically, the implementation manners provided in the above steps may be referred to, which are not described herein again.
In the embodiment of the application, the song recommendation device can acquire the corresponding search object content text according to the search record by acquiring the search record of the target user in the preset time period. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. Based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, the text similarity between the content text of the search object and the song information of each song can be determined, and then at least one song can be determined from the song library according to the text similarity and recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved, the flexibility is high, and the application range is wide.
Referring to fig. 8, fig. 8 is a schematic structural diagram of a terminal device according to an embodiment of the present application. As shown in fig. 8, the terminal device in the present embodiment may include: one or more processors 401 and a memory 402. The processor 401 and the memory 402 are connected via a bus 403. The memory 402 is used for storing a computer program comprising program instructions, and the processor 401 is used for executing the program instructions stored in the memory 402 for performing the following operations:
Obtaining a search object content text corresponding to a search record of a target user within a preset time period, and obtaining a content text feature vector corresponding to the search object content text, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the search object content text;
Obtaining a song text feature vector corresponding to each song included in a song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information;
And determining the text similarity between the content text of the search object and the song information of each song based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, and determining at least one song from the song library to recommend to the target user according to the text similarity.
In some possible embodiments, the processor 401 is configured to:
Acquiring a log file corresponding to a search engine used by the target user, wherein the log file comprises a search record of the target user, and the search record comprises a search string used for searching by the target user and search time;
acquiring a search character string with the largest occurrence number in a search record of a target user in a preset time period, determining a plurality of search result content texts based on the search character string, and acquiring a webpage attribute feature vector corresponding to each search result content text;
Acquiring a search character text feature vector corresponding to the search character string, acquiring a user basic attribute feature vector of the target user, and generating a joint feature vector corresponding to each search result content text based on the search character text feature vector, the user basic attribute feature vector and a web page attribute feature vector corresponding to each search result content text;
obtaining a text ordering model, and inputting each joint feature vector into the text ordering model to obtain an ordering result of each search result content text output by the text ordering model, wherein the text ordering model is obtained by training according to a plurality of sample joint feature vectors corresponding to a plurality of sample texts and the ordering result of the plurality of sample texts;
and determining the first n search result content texts in the sorting results as search object content texts.
In some possible embodiments, the search record includes a browsing record and a browsing time of the target user; the processor 401 is configured to:
and acquiring a Uniform Resource Locator (URL) included in the browsing record in a preset time period, and accessing the URL to acquire a corresponding content text as a search object content text.
In some possible embodiments, the processor 401 is configured to:
Word segmentation processing is carried out on the search object content text to obtain a plurality of words forming the search object content text;
Acquiring a preset word vector lookup table, wherein the word vector lookup table comprises a plurality of word vectors corresponding to a plurality of words, and one word corresponds to one word vector;
and determining word vectors corresponding to each word in the plurality of words composing the text of the search object content from the word vector lookup table.
In some possible embodiments, the processor 401 is configured to:
a text similarity classification model is obtained, the content text feature vector and the song text feature vector corresponding to any song are input into the text similarity classification model, wherein the text similarity classification model is obtained through training according to the content text feature vector corresponding to the sample search object content text, the song text feature vector corresponding to the sample song and the text similarity classification result identification;
And acquiring a text similarity classification result identifier output by the text similarity classification model, and determining the text similarity between the song text information corresponding to any song and the text of the search object content according to the text similarity classification result identifier.
In some possible embodiments, the processor 401 is configured to:
Calculating Euclidean distance between the text feature vectors of the content and the text feature vectors of the songs corresponding to each song in the song library respectively;
And converting the Euclidean distance into a similarity value to serve as the text similarity between the text of the content of the search object and the song information of each song.
In some possible embodiments, the processor 401 is configured to:
Acquiring a preset similarity threshold value, and acquiring song recommendation with text similarity not smaller than the preset similarity threshold value from the song library to the target user; or alternatively
And descending the text similarity, and obtaining k songs corresponding to the first k text similarity after descending the sequence to be recommended to the target user, wherein k is an integer greater than 0.
It should be appreciated that in some possible embodiments, the processor 401 may be a central processing unit (central processing unit, CPU), which may also be other general purpose processors, digital signal processors (DIGITAL SIGNAL processors, DSPs), application Specific Integrated Circuits (ASICs), off-the-shelf programmable gate arrays (field programmable GATE ARRAY, FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory 402 may include read only memory and random access memory and provides instructions and data to the processor 401. A portion of memory 402 may also include non-volatile random access memory. For example, the memory 402 may also store information of device type.
In a specific implementation, the terminal device may execute, through each built-in functional module, an implementation manner provided by each step in fig. 1 and fig. 5, and specifically may refer to an implementation manner provided by each step, which is not described herein again.
In the embodiment of the application, the terminal equipment can acquire the corresponding search object content text according to the search record by acquiring the search record of the target user in the preset time period. And obtaining a content text feature vector corresponding to the content text of the search object, wherein the content text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing the content text of the search object. And obtaining a song text feature vector corresponding to each song included in the song library, wherein the song text feature vector comprises a plurality of word vectors corresponding to a plurality of words composing song information. Based on the content text feature vector and the song text feature vector corresponding to each song included in the song library, the text similarity between the content text of the search object and the song information of each song can be determined, and then at least one song can be determined from the song library according to the text similarity and recommended to the target user. By adopting the embodiment of the application, the accuracy of recommending songs for new users can be improved, the flexibility is high, and the application range is wide.
The embodiment of the present application further provides a computer readable storage medium, where the computer readable storage medium stores a computer program, where the computer program includes program instructions, and when the program instructions are executed by a processor, implement a song recommendation method provided by each step in fig. 1 and fig. 5, and specifically, reference may be made to an implementation manner provided by each step, which is not described herein again.
The computer readable storage medium may be the song recommendation apparatus provided in any one of the foregoing embodiments or an internal storage unit of the terminal device, for example, a hard disk or a memory of the electronic device. The computer readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, a smart memory card (SMART MEDIA CARD, SMC), a Secure Digital (SD) card, a flash memory card (FLASH CARD), or the like, which are provided on the electronic device. Further, the computer-readable storage medium may also include both an internal storage unit and an external storage device of the electronic device. The computer-readable storage medium is used to store the computer program and other programs and data required by the electronic device. The computer-readable storage medium may also be used to temporarily store data that has been output or is to be output.
The terms "first," "second," "third," "fourth" and the like in the claims and in the description and drawings of the present application are used for distinguishing between different objects and not for describing a particular sequential order. Furthermore, the terms "comprise" and "have," as well as any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those listed steps or elements but may include other steps or elements not listed or inherent to such process, method, article, or apparatus.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Those of skill in the art will explicitly and implicitly appreciate that the embodiments described herein may be combined with other embodiments. The term "and/or" as used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes such combinations. Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, in computer software, or in a combination of the two, and that the elements and steps of the examples have been generally described in terms of function in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.
The method and related apparatus provided in the embodiments of the present application are described with reference to the flowchart and/or schematic structural diagrams of the method provided in the embodiments of the present application, and each flow and/or block of the flowchart and/or schematic structural diagrams of the method may be implemented by computer program instructions, and combinations of flows and/or blocks in the flowchart and/or block diagrams. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or structural diagram block or blocks. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or structures.