WO2024045926A1 - 多媒体推荐方法、推荐装置、车机***和存储介质 - Google Patents
多媒体推荐方法、推荐装置、车机***和存储介质 Download PDFInfo
- Publication number
- WO2024045926A1 WO2024045926A1 PCT/CN2023/107958 CN2023107958W WO2024045926A1 WO 2024045926 A1 WO2024045926 A1 WO 2024045926A1 CN 2023107958 W CN2023107958 W CN 2023107958W WO 2024045926 A1 WO2024045926 A1 WO 2024045926A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- word
- query
- multimedia
- words
- user behavior
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 230000006399 behavior Effects 0.000 claims description 49
- 238000004364 calculation method Methods 0.000 claims description 33
- 238000012937 correction Methods 0.000 claims description 31
- 238000004590 computer program Methods 0.000 claims description 14
- 230000008569 process Effects 0.000 description 10
- 238000012545 processing Methods 0.000 description 10
- 238000012549 training Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 8
- 238000010586 diagram Methods 0.000 description 5
- 230000002452 interceptive effect Effects 0.000 description 5
- 230000003542 behavioural effect Effects 0.000 description 3
- 238000003491 array Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 239000000463 material Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000003796 beauty Effects 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 239000000835 fiber Substances 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/43—Querying
- G06F16/435—Filtering based on additional data, e.g. user or group profiles
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
- G06F16/48—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/216—Parsing using statistical methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Definitions
- the present application relates to the technical field of multimedia content search, and in particular to a multimedia recommendation method, recommendation device, vehicle system and storage medium.
- In-vehicle media services include content from multiple media sources, including aggregation of music, listening to books, and news. Users search for the content they want through voice or the car console.
- the vehicle-mounted system performs natural language understanding based on the query words input by the user, identifies the songs and singer names in the query words for music, and performs query output based on the clear singer and song names.
- the output accuracy is low.
- This application provides a multimedia recommendation method, recommendation device, vehicle-machine system and storage medium.
- the multimedia recommendation method in the implementation of this application includes:
- the trained predetermined model calculate a plurality of matching scores between the multimedia content and the correction word and a plurality of user behavior scores of the multimedia content, where the user behavior score is determined based on the historical playback status of the multimedia content;
- a plurality of the multimedia contents are sorted and a recommendation result is output.
- the multimedia recommendation method in the embodiment of the present application obtains multiple corrected multimedia contents by correcting the user input query words, and then outputs the recommended content after sorting according to the matching scores and user behavior scores calculated by the trained predetermined model, which can improve When users enter fuzzy query terms, they get more accurate output results and recommend content that is more suitable for users.
- the obtaining query terms includes:
- the input speech parsed text is used as the query word.
- modifying the query term and obtaining the corrected term includes:
- the similar words are used as the correction words.
- the calculation of the matching score includes:
- the matching score is calculated based on the data length, the length of the text and the weight of the target word.
- calculating the weight of the target word in the predetermined model output sentence includes;
- the product of the ratio and the frequency of occurrence is used as the weight of the target word.
- the matching score is calculated using the following formula:
- total_w word is the sum of the weights of the matched target words
- mappingword len is the length of the data in the matched multimedia content
- Text1 len is the length of the text of the query word
- Text2 len is the length of the text of the target word.
- the calculation of the user behavior score includes:
- the characteristic data including at least one of the number of plays, the number of players, the average play duration, the number of plays by the same user, and the play duration;
- the value of the product of the normalized value and the weight of the feature data is accumulated as the user behavior score.
- the normalized value is calculated using the following formula:
- X is the feature data
- Xmin is the minimum value of the feature data
- Xmax is the maximum value of the feature data
- the first acquisition module is used to acquire query words
- a correction module for correcting the query word based on spelling and similar words and obtaining the corrected word
- a second acquisition module used to acquire multiple multimedia contents related to the correction word
- a calculation module configured to calculate a plurality of matching scores between the multimedia content and the correction word and a plurality of user behavior scores of the multimedia content according to the trained predetermined model, and the user behavior score is based on the history of the media content. The playback situation is confirmed;
- a push module configured to sort a plurality of multimedia contents according to the matching score and the user behavior score and output recommendation results.
- the recommendation device in the embodiment of the present application obtains a plurality of corrected multimedia contents by correcting the query words input by the user, and then outputs the recommended content after sorting according to the matching scores and user behavior scores calculated by the trained predetermined model, which can improve the user experience. Get more accurate output results when entering fuzzy query terms and recommend content that is more relevant to users.
- the vehicle-machine system in the embodiment of the present application includes a memory and a processor.
- the memory stores a computer program.
- the processor executes the computer program, it implements the multimedia recommendation method as described in any of the above embodiments.
- the vehicle-machine system in the embodiment of the present application can perform recommendation display based on the user's behavioral data based on the text matching of the user's search content through the multimedia recommendation method.
- a non-volatile computer-readable storage medium containing a computer program according to an embodiment of the present application.
- the processor implements the multimedia recommendation described in any of the above embodiments. method.
- Figure 1 is a schematic flow chart of a multimedia recommendation method according to an embodiment of the present application.
- Figure 2 is a module schematic diagram of the recommendation device according to the embodiment of the present application.
- Figure 3 is a schematic diagram of the overall architecture of the multimedia recommendation method according to the embodiment of the present application.
- Figure 4 is a schematic diagram of the process from data collection to model processing in the multimedia recommendation method according to the embodiment of the present application.
- Figure 5 is a schematic diagram of the online search process in Figure 3.
- Figure 6 is a schematic flowchart of a multimedia recommendation method according to an embodiment of the present application.
- Figure 7 is a schematic flow chart of a multimedia recommendation method according to an embodiment of the present application.
- Figure 8 is a schematic flow chart of the multimedia recommendation method according to the embodiment of the present application.
- Figure 9 is a schematic flowchart of a multimedia recommendation method according to an embodiment of the present application.
- Figure 10 is a schematic structural diagram of a vehicle-machine system according to an embodiment of the present application.
- Recommendation device 1000 first acquisition module 110, correction module 120, second acquisition module 130, calculation module 140;
- Vehicle system 2000 memory 220, processor 220.
- the multimedia recommendation method in the embodiment of this application includes:
- S40 According to the trained predetermined model, calculate the matching scores of multiple multimedia contents and correction words and the user behavior scores of multiple multimedia contents.
- the user behavior scores are determined based on the historical playback status of the multimedia content;
- S50 Sort multiple multimedia contents based on matching scores and user behavior scores and output recommendation results.
- the recommendation device 1000 for recommending multimedia content in the embodiment of the present application includes a first acquisition module 110, a correction module 120, a second acquisition module 130 and a calculation module 130.
- the first acquisition module 110 is used to acquire query words.
- the correction module 120 is used to correct the query word and obtain the correction word based on spelling and similar words
- the second acquisition module 130 is used to obtain multiple multimedia contents related to the correction word
- the calculation module 130 is used to predetermine according to the training model, calculates the matching scores of multiple multimedia contents and corrective words and the user behavior scores of multiple multimedia contents.
- the user behavior scores are determined based on the historical playback status of the media content
- the push module is used to calculate multiple multimedia content based on the matching scores and user behavior scores. Sort multimedia content and output recommendation results.
- the multimedia recommendation method and recommendation device 1000 in the embodiment of the present application obtains multiple corrected multimedia contents by correcting the user input query words, and then outputs recommendations after sorting the matching scores and user behavior scores calculated based on the trained predetermined model. content, which can improve users' input of fuzzy query words to obtain more accurate output results and recommend content that is more suitable for users.
- step S10 may be first performed, and the query word may be obtained through the first acquisition module 110.
- the query word may be a word that the user inputs to the car terminal to be queried.
- step S20 can be taken to correct the query word based on spelling and similar words and obtain the corrected word.
- the spelling correction and similar word correction can correct the query word through pinyin correction and similar word replacement, and then the query word can be corrected. Get the corrected corrective word.
- step S30 may be taken to obtain multiple multimedia contents related to the correction word through the second acquisition device.
- the multimedia contents may be multiple media sources in the multimedia services provided to the user by the vehicle terminal, such as audio, video, songs and news. etc. content.
- the calculation module 130 can take step S40 to calculate the matching scores of multiple multimedia contents and correction words and the user behavior scores of multiple multimedia contents according to the trained predetermined model, and the user behavior scores are determined based on the historical playback status of the media content; wherein,
- the matching score can be understood as the degree of matching between multiple multimedia contents searched based on the modified words and the modified words.
- the user behavior can be understood as the user's historical playback status in multiple multimedia contents searched based on modified words. For example, the historical playback status may be the number of historical playbacks, historical playback duration, etc.
- the predetermined model can use model methods such as TF-IDF model and weight model to calculate matching scores and user behavior scores.
- the recommendation module can then take step S50 to sort the multiple multimedia contents and output the recommendation results according to the results of the matching scores and user behavior scores calculated in step S40.
- spelling correction, similar word replacement and text retrieval can use the lexicon pinyin table, similar lexicon, content library, etc. corresponding to the interactive data in the figure; the lexicon extraction and similar words in the big data platform corresponding to the offline training in the figure Training methods such as library extraction, user content statistics, data cleaning, data integration, and content statistics can be used for spelling correction, similar word replacement, and text retrieval.
- the matching score can be calculated and then the user behavior score can be calculated. Then the output recommendation results can be obtained based on the calculation results of matching scores and user behavior scores.
- the corresponding matching score calculation and user behavior score calculation can also be performed based on offline training of the interactive data structure.
- the matching score calculation can use the TF-IDF model corresponding to the interactive data in the figure and the TF-IDF model and weight model in the training platform corresponding to the offline training in the figure.
- the user behavior score calculation can use the content popularity, user content behavior, etc. corresponding to the interactive data in the figure, and the TF-IDF model and weight model in the training platform corresponding to the offline training in the figure.
- the vehicle-machine end can report the hidden data to the cloud and forward it to kafka (message channel).
- the big data platform consumes the kafka (message channel) data and enters the data warehouse.
- Business data in the cloud including basic information about users and content, is synchronized to the big data platform.
- the TF-IDF model can be trained on the text data of the user's voice search, which is used to calculate the weight of the words in the query words for later text matching.
- the TF model calculation method is the ratio of the number of occurrences of words in the text to the number of occurrences of all words in the text during prediction.
- tf i,j represents the i-th word in the j-th sentence of the query word
- n i,j represents the number of times the i-th word in the j-th sentence of the query word appears in the j-th sentence
- ⁇ k n k, j represents the sum of the times of all words in the j-th sentence of the i-th word in the j-th sentence of the query word.
- the query words input by the user are two sentences, the first sentence is "Play, fire truck", and the second sentence is "Oh, here is a song, I love the country more than the beauty.”
- the first sentence is segmented to get the words “play” and "fire truck”.
- Taking the word “play” as an example, the TF of "play” 1/2.
- the principle is as follows:
- the IDF calculation method can learn the number of times each word appears in the query words during model training, and predict according to the predicted Perform a dictionary lookup on the tested text words.
- ⁇ j:t i ⁇ d j ⁇ represents the number of sentences in which a single word appears in the query word.
- represents the total number of sentences in the query term.
- the model data is the value of the product of TF and IDF, which is used as the weight of each target word.
- the output of the model file can be saved in PMML format.
- the overall process of online search can be shown in Figure 5.
- the user on the car side inputs voice to request the voice service to request the search service for the text query.
- the search service performs text processing and data query within the search service and then calls the model from the model warehouse and returns the results. to the search service, and then return the final data to the vehicle terminal.
- obtaining the query terms includes:
- the input text is used as the query word or the input speech-parsed text is used as the query word, or the input text is used as the query word and the input speech-parsed text is used as the query word at the same time.
- correcting the query word and obtaining the corrected word includes:
- the correction module 120 is used for spelling based on pronunciation, querying the thesaurus to confirm the correctness of the query word; and for correcting the query word and obtaining the corrected word when the query word is incorrect; and for querying similar words based on semantics.
- the library is used to confirm similar words of the query term; and is used to use similar words as corrective words when the query term has similar words.
- the query word is corrected by confirming the spelling and semantics of the pronunciation of the query word from the lexicon.
- step S20 you can first correct the spelling of the query word and then replace similar words.
- Step S21 is taken to query the thesaurus based on the pronunciation of the spelling to confirm the correctness of the query word; and then step S22 is taken. If the query word is incorrect, correct the query word and obtain the corrected word; then take step S23 to query the similar vocabulary database based on semantics to confirm the similar words of the query word; then take step S24, if the query word has similar words , use similar words as corrective words.
- the calculation of the matching score includes:
- the calculation module 130 is used to calculate the data length in the multimedia content; and to determine the length of the text that needs to be matched; and to calculate the weight of the target word in the predetermined model output sentence; and to calculate the weight of the target word in the predetermined model output sentence; and to determine the length of the text according to the data length, the length of the text, and The weight of the target word is calculated to obtain the matching score.
- the matching score is calculated using data length, text length and target word weight, which can reflect the matching degree of the multimedia content corresponding to the target word in the query word.
- the matching score calculation in step S40 can first take step S41 to calculate the data length in the multimedia content; then step S42 can be taken to determine the length of the text that needs to be matched; and then step S43 can be taken to calculate the length of the predetermined model output sentence.
- the weight of the target word; then step S44 is taken to calculate the matching score based on the data length, the length of the text and the weight of the target word.
- the predetermined model may be a TF-IDF model and a weight model.
- calculating the weight of the target word in the predetermined model output sentence includes;
- S431 Calculate the ratio of the number of times the target word appears in the output sentence to the number of times all words appear in the output sentence;
- the calculation module 130 is used to calculate the ratio of the number of occurrences of the target word in the output sentence to the number of occurrences of all words in the output sentence; and to calculate the frequency of the target word in the document; and to combine the ratio and the frequency of occurrence.
- the product is used as the weight of the target word.
- the weight of the target word in the sentence that can reflect the target word is calculated by using the product of the frequency ratio and the frequency of occurrence.
- step S431 the calculation method of the TF model may be used to calculate the ratio of the number of occurrences of the target word in the output sentence to the number of occurrences of all words in the output sentence.
- step S432 the calculation method of the IDF model may be used to calculate the frequency of occurrence of the target word in the document.
- step S433 the product of the ratio calculated in step S431 and the frequency of occurrence in step S432 can be used as the weight of the target word.
- the match score is calculated using the following formula:
- total_w word is the sum of the weights of the matched target words
- mappingWord len is the length of the data in the matched multimedia content
- Text1 len is the length of the text of the query word
- Text2 len is the length of the text of the target word.
- the calculation formula used in the matching score can more accurately obtain the matching degree between the target word and the multimedia content.
- the total_w word in the calculation formula is the sum of the weights of the matched target words, where the weight can be the weight of the target words in the sentence obtained by predicting the output using the TF-IDF model in the prediction model; mappingWord len is the weight of the matched target words.
- the length of the data in the multimedia content Text1 len is the length of the text of the query word, that is, the length of the text of the query word entered by the user; Text2 len is the length of the text of the target word, that is, the length of the text of the target word in the query word length.
- the calculation of user behavior scores includes:
- S45 Calculate the normalized value of the feature data in the multimedia content.
- the feature data includes at least one of the number of plays, the number of players, the average play duration, the number of plays by the same user, and the play duration;
- S47 Accumulate the value of the product of the normalized value and the weight of the feature data as the user behavior score.
- the calculation module 130 is used to calculate the normalized value of the feature data in the multimedia content.
- the feature data includes at least one of the number of plays, the number of players, the average play duration, the number of plays by the same user, and the play duration; and is used to calculate features.
- the calculation of user behavior scores can display behavioral data such as user preferences for multimedia content.
- the user behavior score can be calculated after the user matching score is calculated, and step S45 can be used to calculate the normalized value of the feature data in the multimedia content, where the feature data includes the number of plays, the number of players, the average play duration, and the same At least one of the user's play times and play duration; and then step S46 can be used to calculate the weight of the feature data. Then step S47 can be taken to accumulate the value of the product of the normalized value and the weight of the feature data as the user behavior score.
- user clicks and music features are normalized, and then the information entropy of each feature is calculated.
- the weight of the feature data is determined based on the information entropy, which can then be used to calculate user behavior scores.
- the information entropy calculation formula is:
- p ij represents the proportion of normalized data values of each dimension data.
- x ij represents the data value in the data table or the i-th row and j-th column in the two-dimensional array. Represents the sum of all row data in each column of data.
- a certain column of data is: 1, 2, 4, 6.
- the information entropy calculation formula also includes:
- the weight calculation formula of information is:
- the normalized value is calculated using the following formula:
- X is the feature data
- Xmin is the minimum value of the feature data
- Xmax is the maximum value of the feature data
- normalization can facilitate sorting of feature data and subsequent calculation of user behavior scores.
- w j is the weight of each column of data
- X norm represents the normalized feature value of each column
- score i represents the calculated score of each row of data, which is the behavior score of each user.
- step S50 multiple multimedia contents are sorted according to the matching scores and user behavior scores and the recommendation results are output.
- the matching score is reconciled with the user's behavior score for the content, sorted from high to low, and then the content is output.
- tune The sum can be an accumulation in a certain proportion, for example, a match score of 0.6 plus a user behavior score of 0.4.
- the content output by different users is different.
- the query entered by user 10001 is "play Xiao Ming's songs”.
- the output effect of user 10001 is as follows:
- the query word entered by user 10002 is also "play Xiao Ming's songs”.
- the output effect of user 10002 is as follows:
- the vehicle-machine system 2000 in the embodiment of the present application includes a memory 210 and a processor 220.
- the memory 210 stores a computer program.
- the processor 220 executes the computer program, it implements any of the multimedia recommendation methods in the above embodiments.
- the vehicle-machine system 2000 can perform recommendation display based on the user's behavioral data based on the text matching of the user's search content through the multimedia recommendation method.
- the vehicle-machine system 2000 can be an intelligent vehicle-mounted system on a car, which can be used to implement functions such as vehicle-machine interaction and media playback.
- the memory 210 and the processor 220 may be provided on the vehicle system 2000.
- the non-volatile computer-readable storage medium containing a computer program when the computer program is executed by one or more processors, causes the processor to implement the multimedia recommendation method of any of the above embodiments.
- the processor of the computer program may be a central processing unit (Central Processing Unit, CPU).
- the processor can also be other general-purpose processors, Digital Signal Processor (DSP), Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA) or other Programmable logic devices, discrete gate or transistor logic devices, discrete hardware Components and other chips, or a combination of the above types of chips.
- DSP Digital Signal Processor
- ASIC Application Specific Integrated Circuit
- FPGA Field-Programmable Gate Array
- Programmable logic devices discrete gate or transistor logic devices, discrete hardware Components and other chips, or a combination of the above types of chips.
- the computer program can be stored in the memory.
- the memory as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer executable programs and modules, as described in the method in the above method embodiment.
- the processor executes various functional applications and data processing of the processor by running non-transient software programs, instructions and modules stored in the memory, that is, implementing the method in the above method embodiment.
- a "computer-readable medium” may be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
- Non-exhaustive list of computer readable media include the following: electrical connections with one or more wires (electronic device), portable computer disk cartridges (magnetic device), random access memory (RAM), Read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable compact disc read-only memory (CDROM).
- the computer-readable medium may even be paper or other suitable medium on which the program may be printed, as the paper or other medium may be optically scanned, for example, and subsequently edited, interpreted, or otherwise suitable as necessary. process to obtain the program electronically and then store it in computer memory.
- the processor can be a central processing unit (Central Processing Unit, CPU), or other general-purpose processor, digital signal processor (Digital Signal Processor, DSP), application specific integrated circuit (Application Specific Integrated Circuit, ASIC), or off-the-shelf programmable Gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
- CPU Central Processing Unit
- DSP Digital Signal Processor
- ASIC Application Specific Integrated Circuit
- FPGA off-the-shelf programmable Gate array
- a general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.
- each functional unit in various embodiments of the present application can be integrated into a processing module, each unit can exist physically alone, or two or more units can be integrated into one module.
- the above integrated modules can be implemented in the form of hardware or software function modules. If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.
- the storage media mentioned above can be read-only memory, magnetic disks or optical disks, etc.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Artificial Intelligence (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Library & Information Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
一种多媒体推荐方法、推荐装置(1000)、车机***(2000)和存储介质。多媒体推荐方法包括:(S10)获取查询词;(S20)基于拼写和相似词,修正查询词并得到修正词;(S30)获取与修正词相关的多个多媒体内容;(S40)根据训练的预定模型,计算多个多媒体内容与修正词的匹配分和多个多媒体内容的用户行为分,用户行为分根据媒体内容的历史播放情况确定;(S50)根据匹配分和用户行为分,对多个多媒体内容进行排序并输出推荐结果。
Description
本申请涉及多媒体内容搜索技术领域,尤其涉及一种多媒体推荐方法、推荐装置、车机***和存储介质。
在车载媒体服务中包含多种媒体来源内容,音乐、听书、新闻的内容聚合。用户通过语音或者车控台进行搜索自己想要找的内容。在相关技术中,车载***根据用户输入的查询词进行自然语言理解,并针对音乐识别查询词中的歌曲和歌手名称,基于明确的歌手和歌曲名称进行查询输出。但是,在没有明确的歌曲和歌手名称时,只进行模糊匹配搜索,输出准确率较低。
发明内容
本申请提供一种多媒体推荐方法、推荐装置、车机***和存储介质。
本申请实施方式的多媒体推荐方法包括:
获取查询词;
基于拼写和相似词,修正所述查询词并得到修正词;
获取与所述修正词相关的多个多媒体内容;
根据训练的预定模型,计算多个所述多媒体内容与所述修正词的匹配分和多个所述多媒体内容的用户行为分,所述用户行为分根据所述多媒体内容的历史播放情况确定;
根据所述匹配分和所述用户行为分,对多个所述多媒体内容进行排序并输出推荐结果。
本申请实施方式的多媒体推荐方法通过对用户输入端查询词进行修正得到修正后的多个多媒体内容,然后再根据训练的预定模型计算得到的匹配分和用户行为分排序后输出推荐内容,可以提升用户输入模糊查询词时得到更准确的输出结果并推荐更符合用户的内容。
在某些实施方式中,所述获取查询词包括:
将输入的文本作为所述查询词;和/或,
将输入的语音解析的文本作为所述查询词。
在某些实施方式中,修正所述查询词并得到修正词,包括:
基于读音的拼写,查询词库以确认所述查询词的正确性;
在所述查询词不正确的情况下,修正所述查询词并得到修正词;
基于语义,查询相似词库以确认所述查询词的相似词;
在查询词具有相似词的情况下,将所述相似词作为所述修正词。
在某些实施方式中,所述匹配分的计算包括:
计算所述多媒体内容中的数据长度;
确定需要匹配的文本的长度;
计算所述预定模型输出句子中的目标词的权重;
根据所述数据长度、所述文本的长度和所述目标词的权重计算得到所述匹配分。
在某些实施方式中,所述计算所述预定模型输出句子中的目标词的权重,包括;
计算所述目标词在所述输出句子中出现的次数与所述输出句子中所有词的出现次数的比值;
计算所述目标词在文档中出现的频率;
将所述比值和所述出现的频率的乘积作为所述目标词的权值。
在某些实施方式中,所述匹配分采用以下公式计算:
其中,total_wword为匹配的目标词的权重的加和;mapingwordlen为匹配到的多媒体内容中的数据长度;Text1len为查询词的文本的长度;Text2len为目标词的文本的长度。
在某些实施方式中,所述用户行为分的计算包括:
计算所述多媒体内容中的特征数据的归一化值,所述特征数据包括播放次数、播放人数、平均播放时长、同一用户的播放次数、播放时长中的至少一种;
计算所述特征数据的权重;
累加所述归一化值和所述特征数据的权重的乘积的值作为所述用户行为分。
在某些实施方式中,所述归一化值采用以下公式计算:
Xnorm=(X-Xmin)/(Xmax-Xmin)
其中,X为特征数据,Xmin为特征数据的最小值,Xmax为特征数据的最大值。
本申请实施方式的推荐装置包括:
第一获取模块,用于获取查询词;
修正模块,用于基于拼写和相似词,修正所述查询词并得到修正词;
第二获取模块,用于获取与所述修正词相关的多个多媒体内容;
计算模块,用于根据训练的预定模型,计算多个所述多媒体内容与所述修正词的匹配分和多个所述多媒体内容的用户行为分,所述用户行为分根据所述媒体内容的历史播放情况确定;
推送模块,用于根据所述匹配分和所述用户行为分,对多个所述多媒体内容进行排序并输出推荐结果。
本申请实施方式的推荐装置通过对用户输入端查询词进行修正得到修正后的多个多媒体内容,然后再根据训练的预定模型计算得到的匹配分和用户行为分排序后输出推荐内容,可以提升用户输入模糊查询词时得到更准确的输出结果并推荐更符合用户的内容。
本申请实施方式的车机***包括存储器和处理器,所述存储器存储有计算机程序,所述处理器执行所述计算机程序时实现如上述任一项实施方式所述多媒体推荐方法。
本申请实施方式的车机***通过多媒体推荐方法能够在用户搜索内容文本匹配的基础之上,根据用户的行为数据进行推荐展示。
本申请实施方式的包含计算机程序的非易失性计算机可读存储介质,当所述计算机程序被一个或多个处理器执行时,使得所述处理器实现上述任一项实施方式所述多媒体推荐方法。
本申请的附加方面和优点将在下面的描述中部分给出,部分将从下面的描述中变得明显,或通过本申请的实践了解到。
本申请的上述和/或附加的方面和优点从结合下面附图对实施方式的描述中将变得明显和容易理解,其中:
图1是本申请实施方式的多媒体推荐方法的流程示意图;
图2是本申请实施方式的推荐装置的模块示意图;
图3是本申请实施方式的多媒体推荐方法的整体架构示意图;
图4是本申请实施方式的多媒体推荐方法中数据采集到模型处理的过程示意图;
图5是图3中的在线搜索的过程示意图;
图6是本申请实施方式的多媒体推荐方法的流程示意图;
图7是本申请实施方式的多媒体推荐方法的流程示意图;
图8是本申请实施方式的多媒体推荐方法的流程示意图;
图9是本申请实施方式的多媒体推荐方法的流程示意图;
图10是本申请实施方式的车机***的结构示意图。
主要元件符号说明:
推荐装置1000、第一获取模块110、修正模块120、第二获取模块130、计算模块140;
车机***2000、存储器220、处理器220。
下面详细描述本申请的实施方式,所述实施方式的示例在附图中示出,其中自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。下面通过参考附图描述的实施方式是示例性的,仅用于解释本申请,而不能理解为对本申请的限制。
下文的公开提供了许多不同的实施方式或例子用来实现本申请的不同结构。为了简化本申请的公开,下文中对特定例子的部件和设置进行描述。当然,它们仅仅为示例,并且目的不在于限制本申请。此外,本申请可以在不同例子中重复参考数字和/或参考字母,这种重复是为了简化和清楚的目的,其本身不指示所讨论各种实施方式和/或设置之间的关系。此外,本申请提供了的各种特定的工艺和材料的例子,但是本领域普通技术人员可以意识到其他工艺的应用和/或其他材料的使用。
请参阅图1,本申请实施方式的多媒体推荐方法包括:
S10:获取查询词;
S20:基于拼写和相似词,修正查询词并得到修正词;
S30:获取与修正词相关的多个多媒体内容;
S40:根据训练的预定模型,计算多个多媒体内容与修正词的匹配分和多个多媒体内容的用户行为分,用户行为分根据多媒体内容的历史播放情况确定;
S50:根据匹配分和用户行为分,对多个多媒体内容进行排序并输出推荐结果。
请参阅图2,本申请实施方式的用于推荐多媒体内容的推荐装置1000包括第一获取模块110、修正模块120、第二获取模块130和计算模块130,第一获取模块110用于获取查询词;修正模块120用于基于拼写和相似词,修正查询词并得到修正词;第二获取模块130用于获取与修正词相关的多个多媒体内容;计算模块130用于根据训练的预定
模型,计算多个多媒体内容与修正词的匹配分和多个多媒体内容的用户行为分,用户行为分根据媒体内容的历史播放情况确定;推送模块用于根据匹配分和用户行为分,对多个多媒体内容进行排序并输出推荐结果。
本申请实施方式的多媒体推荐方法和推荐装置1000通过对用户输入端查询词进行修正得到修正后的多个多媒体内容,然后再根据训练的预定模型计算得到的匹配分和用户行为分排序后输出推荐内容,可以提升用户输入模糊查询词时得到更准确的输出结果并推荐更符合用户的内容。
具体地,推荐装置1000实现多媒体推荐方法可先采取步骤S10,可通过第一获取模块110来获取查询词,查询词可以是用户向车机端输入的想要查询的词语。获取用户输入的查询词后可采取步骤S20,基于拼写和相似词修正查询词并得到修正词,其中拼写修正和相似词修正可通过拼音改错以及相似词替换等方式对查询词进行修正,进而得到修正后的修正词。进一步可采取步骤S30通过第二获取装置获取与修正词相关的多个多媒体内容,多媒体内容可以是车机端向用户提供的多媒体服务中的多种媒体来源,例如,音频、视频、歌曲以及新闻等内容。
然后计算模块130可采取步骤S40,根据训练的预定模型,计算多个多媒体内容与修正词的匹配分和多个多媒体内容的用户行为分,用户行为分根据媒体内容的历史播放情况确定;其中,匹配分可理解为根据修改词进行搜索得到的多个多媒体内容与修正词的匹配程度。用户行为分可理解为根据修改词进行搜索得到的多个多媒体内容中用户历史播放的情况,例如,历史播放情况可以是历史播放次数、历史播放时长等。预定模型可采用TF-IDF模型、权重模型等模型方法对匹配分已经用户行为分进行计算。
再然后推荐模块可采取步骤S50根据步骤S40中计算得出的匹配分和用户行为分的结果,对多个多媒体内容进行排序并输出推荐结果。
多媒体推荐方法的整体架构可如图3所示:
用户通过在线搜索输入词语进行查询,获取查询词后可先基于拼写改错再经过相似词替换后进行对多媒体内容的文本检索。拼写改错、相似词替换以及文本检索可基于交互数据结合离线训练的方式进行。
例如,拼写改错、相似词替换以及文本检索可运用图中交互数据对应的词库拼音表、相似词库、内容库等;图中离线训练对应的大数据平台内的词库提取、相似词库提取、用户对内容统计、数据清洗、数据整合以及内容统计等训练方式可有用于拼写改错、相似词替换以及文本检索。
进一步地,文本检索得到结果后可进行匹配分计算然后再进行用户行为分计算,最
后可根据匹配分和用户行为分的计算结果得到输出推荐结果。相应的匹配分计算以及用户行为分的计算也可基于交互数据结构离线训练的方式进行。
例如,匹配分计算可运用图中交互数据对应的TF-IDF模型及图中离线训练对应的训练平台内的TF-IDF模型和权重模型。用户行为分计算可运用图中交互数据对应的内容热度、用户内容行为等及图中离线训练对应的训练平台内的TF-IDF模型和权重模型。
预定模型构建的过程可如图4所示:
首先可进行数据采集,车机端可将埋点数据上报至云端,转发给kafka(消息通道),大数据平台消费kafka(消息通道)数据进入数据仓库。云端的业务数据包括用户和内容的基础信息同步到大数据平台。
然后进行数据处理通过对埋点数据和业务数据进行数据建模,根据需求计算用户特征和音乐特征。对多媒体内容文本数据处理,提取词库、提取关联词库,根据提取的词库对内容文本进行自定义分词,剔除停用词,分词后的专辑、内容名、作者名称合并统一存储。
再对查询词进行词权重模型的训练,可通过对用户使用语音搜索的文本数据训练TF-IDF模型,用于计算查询词中的词在该查询词中的权重,用于后期的文本匹配。
TF模型计算方法,预测时文本内词出现的次数与该文本内所有词的出现次数比值。
其中,tfi,j表示查询词的第j个句子中的第i个词;ni,j表示查询词的第j个句子中的第i个词在第j个句子中出现的次数;∑knk,j表示查询词的第j个句子中的第i个词在第j个句子中所有词的次数和。
示例性地,用户输入的查询词为两句,第一句是“播放,消防车”,第二句是“哦,来一首,爱江山更爱美人”。以用户输入的第一句“播放,消防车”为例。第一句分词得到词语“播放”,“消防车”,以“播放”一词为例,“播放”的TF=1/2,原理如下:
tfi,j=f1,1表示查询词的第1个句子中的第1个词。ni,j=n1,1=1即“播放”一词在查询词的第1个句子中出现了1次。∑knk,j=n1,1+n2,1=1+1=2,即第一句分词得到“播放”一词,“消防车”一词,“播放”一词在查询词的第1个句子中出现了1次,“消防车”在查询词的第1个句子中出现了1次。因此第一句分词得到“播放”,“消防车”两个词,即第一句词的个数总和为2次。
IDF计算方法,可在模型训练时学习每个词在出现查询词中的次数,预测时根据预
测的文本词进行字典查询。
其中,{j:ti∈dj}表示查询词的中出现单个词的句子数。|D|表示查询词中总的句子数。
预测时模型数据为TF与IDF乘积的值,作为每个目标词的权值。模型文件的输出可保存为PMML格式。
在线搜索的整体过程可如图5所示,车机端的用户输入语音请求语音服务对文本query请求搜索服务,搜索服务进行搜索服务内文本处理和数据查询然后从模型仓库中调用模型并将结果返回至搜索服务,再将最终数据返回至车机端。
在某些实施方式中,获取查询词包括:
将输入的文本作为查询词;和/或,
将输入的语音解析的文本作为查询词。
如此,通过输入、语音解析的文本作为查询词可便于用户在不同场景下使用。
具体地,将输入的文本作为查询词或将输入的语音解析的文本作为查询词或者将输入的文本作为查询词和将输入的语音解析的文本同时作为查询词。
请参阅图6,在某些实施方式中,修正查询词并得到修正词(步骤S20),包括:
S21:基于读音的拼写,查询词库以确认查询词的正确性;
S22:在查询词不正确的情况下,修正查询词并得到修正词;
S23:基于语义,查询相似词库以确认查询词的相似词;
S24:在查询词具有相似词的情况下,将相似词作为修正词。
修正模块120用于基于读音的拼写,查询词库以确认查询词的正确性;及用于在查询词不正确的情况下,修正查询词并得到修正词;及用于基于语义,查询相似词库以确认查询词的相似词;以及用于在查询词具有相似词的情况下,将相似词作为修正词。
如此,通过对查询词的读音的拼写及语义从词库中进行确认从而修正查询词。
具体地,实现步骤S20,可先对查询词进行拼写的修正然后再进行相似词的替换,采取步骤S21,基于读音的拼写,查询词库以确认查询词的正确性;再采取步骤S22,在查询词不正确的情况下,修正查询词并得到修正词;然后采取步骤S23,基于语义,查询相似词库以确认查询词的相似词;再采取步骤S24,在查询词具有相似词的情况下,将相似词作为修正词。
例如,用户输入的查询词为“加州旅馆”具有拼写错误,通过步骤S21和步骤S22可将“加洲旅馆”修正为“加州旅馆”。进一步地,通过步骤S23和步骤S24可对修正后的“加州旅馆”进行相似词替换“加州旅馆”可替换为“Hotel California”。
请参阅图7,在某些实施方式中,匹配分的计算(步骤S40)包括:
S41:计算多媒体内容中的数据长度;
S42:确定需要匹配的文本的长度;
S43:计算预定模型输出句子中的目标词的权重;
S44:根据数据长度、文本的长度和目标词的权重计算得到匹配分。
计算模块130用于计算多媒体内容中的数据长度;及用于确定需要匹配的文本的长度;及用于计算预定模型输出句子中的目标词的权重;以及用于根据数据长度、文本的长度和目标词的权重计算得到匹配分。
如此,匹配分采用数据长度、文本的长度和目标词的权重计算可以体现查询词中目标词对应多媒体内容的匹配度。
具体地,步骤S40中的匹配分计算可先采取步骤S41计算多媒体内容中的数据长度;然后可采取步骤S42,确定需要匹配的文本的长度;再然后采取步骤S43,计算预定模型输出句子中的目标词的权重;再采取步骤S44,根据数据长度、文本的长度和目标词的权重计算得到匹配分。其中,预定模型可以是TF-IDF模型和权重模型。
请参阅图8,在某些实施方式中,计算预定模型输出句子中的目标词的权重(步骤S43),包括;
S431:计算目标词在输出句子中出现的次数与输出句子中所有词的出现次数的比值;
S432:计算目标词在文档中出现的频率;
S433:将比值和出现的频率的乘积作为目标词的权值。
计算模块130用于计算目标词在输出句子中出现的次数与输出句子中所有词的出现次数的比值;及用于计算目标词在文档中出现的频率;以及用于将比值和出现的频率的乘积作为目标词的权值。
如此,采用次数比值和出现频率乘积的方式计算能够体现目标词的在句子中的权重。
具体地,步骤S431中计算目标词在输出句子中出现的次数与输出句子中所有词的出现次数的比值可以采用TF模型的计算方法。步骤S432中计算目标词在文档中出现的频率可以采用IDF模型的计算方法。然后如步骤S433可将步骤S431中计算得到的比值与步骤S432中出现的频率的乘积作为目标词的权值。
在某些实施方式中,匹配分采用以下公式计算:
其中,total_wword为匹配到的目标词的权重的加和;mapingWordlen为匹配到的多媒体内容中的数据长度;Text1len为查询词的文本的长度;Text2len为目标词的文本的长度。
如此,匹配分采用的计算公式能够更加准确的得到目标词与多媒体内容的匹配程度。
具体地,计算公式中total_wword为匹配到的目标词的权重的加和,其中权重可以是采用预测模型中的TF-IDF模型预测输出得到的句子中的目标词的权重;mapingWordlen为匹配到的多媒体内容中的数据长度;Text1len为查询词的文本的长度,即用户输入的查询词具有的文本的长度;Text2len为目标词的文本的长度,即查询词中目标词具有的文本的长度。
请参阅图9,在某些实施方式中,用户行为分的计算包括:
S45:计算多媒体内容中的特征数据的归一化值,特征数据包括播放次数、播放人数、平均播放时长、同一用户的播放次数、播放时长中的至少一种;
S46:计算特征数据的权重;
S47:累加归一化值和特征数据的权重的乘积的值作为用户行为分。
计算模块130用于计算多媒体内容中的特征数据的归一化值,特征数据包括播放次数、播放人数、平均播放时长、同一用户的播放次数、播放时长中的至少一种;及用于计算特征数据的权重;以及用于累加归一化值和特征数据的权重的乘积的值作为用户行为分。
如此,用户行为分的计算能够显示出用户对多媒体内容的偏好等行为数据。
具体地,用户行为分的计算可在用户匹配分的计算之后,可采用步骤S45计算多媒体内容中的特征数据的归一化值,其中,特征数据包括播放次数、播放人数、平均播放时长、同一用户的播放次数、播放时长中的至少一种;然后可采用步骤S46,计算特征数据的权重。然后可采取步骤S47,累加归一化值和特征数据的权重的乘积的值作为用户行为分。
示例性地,采用用户点击和音乐特征归一化,然后计算每个特征信息熵的大小,根据信息熵确定特征数据的权重,然后可用于用户行为分的计算。
信息熵计算公式为:
其中,pij表示每个维度数据的归一化的数据值的比例。xij表示数据表,或者二维数组中第i行,第j列的数据值。表示每列数据中所有行数据加和。
举例来说,某一列数据为:1、2、4、6。总和为:1+2+4+6=13;计算后的值为1/13、2/13、4/13、6/13。
信息熵计算公式还包括:
其中,表示上述公式中计算得到的数据的加和。表示所有行数,Ln是log以e为底的对数,Ej表示第j列数据的信息熵。
信息的权重计算公式为:
其中,表示上述公式中Ej计算得到的数据的加和。
请参阅图,在某些实施方式中,归一化值采用以下公式计算:
Xnorm=(X-Xmin)/(Xmax-Xmin)
其中,X为特征数据,Xmin为特征数据的最小值,Xmax为特征数据的最大值。
如此,归一化可便于对特征数据进行排序以及后续用户行为分的计算。
具体地,Xnorm表示特征数据归一化的值,以用户对多媒体内容的播放次数的特征数据为例,比如数据为1、3、4、6,进行归一化值的计算得到的数据为:
数据“1”计算得到的结果是“(1-1)/(6-1)=0”,“3”计算得到的结果是“4”计算得到的结果是“6”计算得到的结果是“(6-1)/(6-1)=1”可得到0到1之间的归化值。
然后用户行为分的计算公式可以是:
其中,wj为每列数据的权重,Xnorm表示每列归一化后的特征值,scorei即表示计算得到的每行数据分数,即为每个用户的行为分。
综上,步骤S50中根据匹配分和用户行为分,对多个多媒体内容进行排序并输出推荐结果。匹配分和用户对内容的行为分进行调和,从高到低排序,然后再输出内容。调
和可以是以一定比例的累加,例如,0.6的匹配分加上0.4的用户行为分。不同用户所输出的内容不同。
示例性的,用户10001输入的查询词为“播放小明的歌曲”。
用户10001输出效果如下表:
用户10002输入的查询词同样为“播放小明的歌曲”。
用户10002输出效果如下表:
可以理解,根据用户的不同将推荐匹配用户并符合用户行为的内容。
请参阅图10,本申请实施方式的车机***2000包括存储器210和处理器220,存储器210存储有计算机程序,处理器220执行计算机程序时实现上述实施方式中任一项多媒体推荐方法。
本申请实施方式的车机***2000通过多媒体推荐方法能够在用户搜索内容文本匹配的基础之上,根据用户的行为数据进行推荐展示。
具体地,车机***2000可以是汽车上的智能车载***,可用于实现车机交互、媒体播放等功能。存储器210和处理器220可设置在车机***2000上。
本申请实施方式的包含计算机程序的非易失性计算机可读存储介质,当计算机程序被一个或多个处理器执行时,使得处理器实现上述任一项实施方式多媒体推荐方法。
计算机程序的处理器可以为中央处理器(Central Processing Unit,CPU)。处理器还可以为其他通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现场可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件
组件等芯片,或者上述各类芯片的组合。
计算机程序可以被存储在存储器中,存储器作为一种非暂态计算机可读存储介质,可用于存储非暂态软件程序、非暂态计算机可执行程序以及模块,如上述方法实施例中的方法所对应的程序指令/模块。处理器通过运行存储在存储器中的非暂态软件程序、指令以及模块,从而执行处理器的各种功能应用以及数据处理,即实现上述方法实施例中的方法。
流程图中或在此以其他方式描述的任何过程或方法描述可以被理解为,表示包括一个或更多个用于实现特定逻辑功能或过程的步骤的可执行指令的代码的模块、片段或部分,并且本申请的优选实施方式的范围包括另外的实现,其中可以不按所示出或讨论的顺序,包括根据所涉及的功能按基本同时的方式或按相反的顺序,来执行功能,这应被本申请的实施例所属技术领域的技术人员所理解。
在流程图中表示或在此以其他方式描述的逻辑和/或步骤,例如,可以被认为是用于实现逻辑功能的可执行指令的定序列表,可以具体实现在任何计算机可读介质中,以供指令执行***、装置或设备(如基于计算机的***、包括处理模块的***或其他可以从指令执行***、装置或设备取指令并执行指令的***)使用,或结合这些指令执行***、装置或设备而使用。就本说明书而言,"计算机可读介质"可以是任何可以包含、存储、通信、传播或传输程序以供指令执行***、装置或设备或结合这些指令执行***、装置或设备而使用的装置。计算机可读介质的更具体的示例(非穷尽性列表)包括以下:具有一个或多个布线的电连接部(电子装置),便携式计算机盘盒(磁装置),随机存取存储器(RAM),只读存储器(ROM),可擦除可编辑只读存储器(EPROM或闪速存储器),光纤装置,以及便携式光盘只读存储器(CDROM)。另外,计算机可读介质甚至可以是可在其上打印所述程序的纸或其他合适的介质,因为可以例如通过对纸或其他介质进行光学扫描,接着进行编辑、解译或必要时以其他合适方式进行处理来以电子方式获得所述程序,然后将其存储在计算机存储器中。
处理器可以是中央处理单元(Central Processing Unit,CPU),还可以是其他通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现成可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
应当理解,本申请的实施方式的各部分可以用硬件、软件、固件或它们的组合来实现。在上述实施方式中,多个步骤或方法可以用存储在存储器中且由合适的指令执行系
统执行的软件或固件来实现。例如,如果用硬件来实现,和在另一实施方式中一样,可用本领域公知的下列技术中的任一项或他们的组合来实现:具有用于对数据信号实现逻辑功能的逻辑门电路的离散逻辑电路,具有合适的组合逻辑门电路的专用集成电路,可编程门阵列(PGA),现场可编程门阵列(FPGA)等。
本技术领域的普通技术人员可以理解实现上述实施例方法携带的全部或部分步骤是可以通过程序来指令相关的硬件完成,所述的程序可以存储于一种计算机可读存储介质中,该程序在执行时,包括方法实施例的步骤之一或其组合。
此外,在本申请的各个实施例中的各功能单元可以集成在一个处理模块中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个模块中。上述集成的模块既可以采用硬件的形式实现,也可以采用软件功能模块的形式实现。所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时,也可以存储在一个计算机可读取存储介质中。
上述提到的存储介质可以是只读存储器,磁盘或光盘等。
尽管上面已经示出和描述了本申请的实施例,可以理解的是,上述实施例是示例性的,不能理解为对本申请的限制,本领域的普通技术人员在本申请的范围内可以对上述实施方式进行变化、修改、替换和变型。
Claims (11)
- 一种多媒体推荐方法,其特征在于,包括:获取查询词;基于拼写和相似词,修正所述查询词并得到修正词;获取与所述修正词相关的多个多媒体内容;根据训练的预定模型,计算多个所述多媒体内容与所述修正词的匹配分和多个所述多媒体内容的用户行为分,所述用户行为分根据所述多媒体内容的历史播放情况确定;根据所述匹配分和所述用户行为分,对多个所述多媒体内容进行排序并输出推荐结果。
- 根据权利要求1所述的方法,其特征在于,所述获取查询词包括:将输入的文本作为所述查询词;和/或,将输入的语音解析的文本作为所述查询词。
- 根据权利要求1所述的方法,其特征在于,所述基于拼写和相似词,修正所述查询词并得到修正词,包括:基于读音的拼写,查询词库以确认所述查询词的正确性;在所述查询词不正确的情况下,修正所述查询词并得到修正词;基于语义,查询相似词库以确认所述查询词的相似词;在查询词具有相似词的情况下,将所述相似词作为所述修正词。
- 根据权利要求1所述的方法,其特征在于,所述匹配分的计算包括:计算所述多媒体内容中的数据长度;确定需要匹配的查询词的文本的长度;计算所述预定模型输出句子中的目标词的权重;根据所述数据长度、所述文本的长度和所述目标词的权重计算得到所述匹配分。
- 根据权利要求4所述的方法,其特征在于,所述计算所述预定模型输出句子中的目标词的权重,包括;计算所述目标词在所述输出句子中出现的次数与所述输出句子中所有词的出现次 数的比值;计算所述目标词在文档中出现的频率;将所述比值和所述出现的频率的乘积作为所述目标词的权值。
- 根据权利要求1所述的方法,其特征在于,所述匹配分采用以下公式计算:
其中,total_wword为匹配到的目标词的权重的加和;mapingWordlen为多媒体内容中的数据长度;Text1len为查询词的文本的长度;Text2len为目标词的文本的长度。 - 根据权利要求1所述的方法,其特征在于,所述用户行为分的计算包括:计算所述多媒体内容中的特征数据的归一化值,所述特征数据包括播放次数、播放人数、平均播放时长、同一用户的播放次数、播放时长中的至少一种;计算所述特征数据的权重;累加所述归一化值和所述特征数据的权重的乘积的值作为所述用户行为分。
- 根据权利要求7所述的方法,其特征在于,所述归一化值采用以下公式计算:
Xnorm=(X-Xmin)/(Xmax-Xmin)其中,X为特征数据,Xmin为特征数据的最小值,Xmax为特征数据的最大值。 - 一种推荐装置,其特征在于,所述推荐装置包括:第一获取模块,用于获取查询词;修正模块,用于基于拼写和相似词,修正所述查询词并得到修正词;第二获取模块,用于获取与所述修正词相关的多个多媒体内容;计算模块,用于根据训练的预定模型,计算多个所述多媒体内容与所述修正词的匹配分和多个所述多媒体内容的用户行为分,所述用户行为分根据所述媒体内容的历史播放情况确定;推送模块,用于根据所述匹配分和所述用户行为分,对多个所述多媒体内容进行排序并输出推荐结果。
- 一种车机***,其特征在于,所述车机***包括存储器和处理器,所述存储器存储有计算机程序,所述处理器执行所述计算机程序时实现如权利要求1-8中任一项所述多媒体推荐方法。
- 一种包含计算机程序的非易失性计算机可读存储介质,其特征在于,当所述计算机程序被一个或多个处理器执行时,使得所述处理器实现权利要求1-8中任一项所述多媒体推荐方法。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211042204.2A CN115422382A (zh) | 2022-08-29 | 2022-08-29 | 多媒体推荐方法、推荐装置、车机***和存储介质 |
CN202211042204.2 | 2022-08-29 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2024045926A1 true WO2024045926A1 (zh) | 2024-03-07 |
Family
ID=84200684
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2023/107958 WO2024045926A1 (zh) | 2022-08-29 | 2023-07-18 | 多媒体推荐方法、推荐装置、车机***和存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN115422382A (zh) |
WO (1) | WO2024045926A1 (zh) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115422382A (zh) * | 2022-08-29 | 2022-12-02 | 浙江极氪智能科技有限公司 | 多媒体推荐方法、推荐装置、车机***和存储介质 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101131706A (zh) * | 2007-09-28 | 2008-02-27 | 北京金山软件有限公司 | 一种查询修正方法及*** |
CN104462573A (zh) * | 2014-12-29 | 2015-03-25 | 北京奇艺世纪科技有限公司 | 一种视频搜索结果展示方法及装置 |
US20150178265A1 (en) * | 2013-12-20 | 2015-06-25 | Google Inc. | Content Recommendation System using a Neural Network Language Model |
CN111324700A (zh) * | 2020-02-21 | 2020-06-23 | 北京声智科技有限公司 | 资源召回方法、装置、电子设备及计算机可读存储介质 |
CN115422382A (zh) * | 2022-08-29 | 2022-12-02 | 浙江极氪智能科技有限公司 | 多媒体推荐方法、推荐装置、车机***和存储介质 |
-
2022
- 2022-08-29 CN CN202211042204.2A patent/CN115422382A/zh active Pending
-
2023
- 2023-07-18 WO PCT/CN2023/107958 patent/WO2024045926A1/zh unknown
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101131706A (zh) * | 2007-09-28 | 2008-02-27 | 北京金山软件有限公司 | 一种查询修正方法及*** |
US20150178265A1 (en) * | 2013-12-20 | 2015-06-25 | Google Inc. | Content Recommendation System using a Neural Network Language Model |
CN104462573A (zh) * | 2014-12-29 | 2015-03-25 | 北京奇艺世纪科技有限公司 | 一种视频搜索结果展示方法及装置 |
CN111324700A (zh) * | 2020-02-21 | 2020-06-23 | 北京声智科技有限公司 | 资源召回方法、装置、电子设备及计算机可读存储介质 |
CN115422382A (zh) * | 2022-08-29 | 2022-12-02 | 浙江极氪智能科技有限公司 | 多媒体推荐方法、推荐装置、车机***和存储介质 |
Also Published As
Publication number | Publication date |
---|---|
CN115422382A (zh) | 2022-12-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107818781B (zh) | 智能交互方法、设备及存储介质 | |
US20190043506A1 (en) | Methods and systems for transcription | |
US9582757B1 (en) | Scalable curation system | |
US10755177B1 (en) | Voice user interface knowledge acquisition system | |
US7983915B2 (en) | Audio content search engine | |
CN111324728A (zh) | 文本事件摘要的生成方法、装置、电子设备及存储介质 | |
US20080270344A1 (en) | Rich media content search engine | |
US20080270110A1 (en) | Automatic speech recognition with textual content input | |
CN105956053B (zh) | 一种基于网络信息的搜索方法及装置 | |
CN109545185B (zh) | 交互***评价方法、评价***、服务器及计算机可读介质 | |
US8731930B2 (en) | Contextual voice query dilation to improve spoken web searching | |
CN109828981B (zh) | 一种数据处理方法及计算设备 | |
WO2024045926A1 (zh) | 多媒体推荐方法、推荐装置、车机***和存储介质 | |
US11935315B2 (en) | Document lineage management system | |
TW201822190A (zh) | 語音辨識系統及其方法、詞彙建立方法與電腦程式產品 | |
CN113157727B (zh) | 提供召回结果的方法、设备和存储介质 | |
CN116662518A (zh) | 问答方法、装置、电子设备及可读存储介质 | |
CN112686051A (zh) | 语义识别模型训练方法、识别方法、电子设备、存储介质 | |
JP5426292B2 (ja) | 意見分類装置およびプログラム | |
El Hannani et al. | Evaluation of the effectiveness and efficiency of state-of-the-art features and models for automatic speech recognition error detection | |
CN108140034B (zh) | 使用主题模型基于接收的词项选择内容项目 | |
CN113076740A (zh) | 政务服务领域的同义词挖掘方法及装置 | |
CN111639160A (zh) | 领域识别的方法、交互的方法、电子设备及存储介质 | |
CN116484829A (zh) | 用于信息处理的方法和设备 | |
CN116028626A (zh) | 文本匹配方法、装置、存储介质以及电子设备 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 23858962 Country of ref document: EP Kind code of ref document: A1 |