Personalized input method and device, storage medium and equipment

Abstract

The invention discloses a personalized input method and device, a storage medium and equipment. The method comprises the following steps: firstly, acquiring historical input data of a target user in each APP; according to the historical input data, constructing a user portrait of the target user in the category to which each APP belongs; then, according to the user portraits of the target user in the categories to which the APPs belong, constructing an overall user portrait of the target user; and inputting the overall user portrait of the target user into a preset large language model in combination with a first prompt instruction prompt, and determining personalized input content of the target user in the target input task according to output content of the large language model. Therefore, the user portrait of the target user in the specific category to which each APP belongs is combined with the large language model to be applied to the target input tasks in various scenes, the personalized input effect of the target user in the target input tasks is improved, and then the input experience of the target user is improved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to a personalized input method, device, storage medium and equipment. Background Art

[0002] With the rapid development of information technologies such as artificial intelligence and the Internet of Things, the application scenarios of human-computer interaction are becoming more and more extensive. As an important bridge for human-computer interaction, input methods have greatly promoted the rapid input and processing of text information. Therefore, it is particularly important to realize personalized input for different users to improve the input experience of each user.

[0003] At present, there are usually two ways to achieve personalized input for different users: one is to label the user and then match the existing domain vocabulary, but this method is difficult to accurately match the user and the domain vocabulary, and too many domain vocabulary will affect the decoding effect. The second is to directly build a personalized vocabulary for each person, but this method requires continuous and regular analysis of each user's input history to update each person's personalized vocabulary, which has a high maintenance cost. It can be seen that the two existing methods of achieving personalized input for different users cannot achieve the ideal effect and reduce the user's input experience. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to provide a personalized input method, apparatus, storage medium and device, which can improve the personalized user input effect and thus enhance the user's input experience.

[0005] The embodiment of the present application provides a personalized input method, including:

[0006] Obtain the historical input data of the target user in each application APP; and construct a user profile of the target user in the category to which each APP belongs based on the historical input data;

[0007] Build an overall user profile of the target user based on the user profile of the target user in the category to which each APP belongs;

[0008] The overall user portrait of the target user is combined with the first prompt instruction prompt and input into a preset large language model, and the personalized input content of the target user in the target input task is determined according to the output content of the large language model.

[0009] In a possible implementation, constructing a user profile of the target user in each category to which each APP belongs according to the historical input data includes:

[0010] Cluster each APP to determine the category to which each APP belongs;

[0011] Count the target user's input history texts in different apps and usage time of different apps in the history input data;

[0012] According to the input history text of the target user in different APPs and the usage time of different APPs, a user portrait of the target user in the category to which each APP belongs is constructed.

[0013] In a possible implementation, clustering each APP to determine the category to which each APP belongs includes:

[0014] According to the classification results of the APP market, each APP is clustered to determine the category to which each APP belongs;

[0015] Alternatively, clustering is performed on each APP based on the classification results of the APP market and the sub-classification results of each classification to determine the category to which each APP belongs;

[0016] Alternatively, based on the functional attribute description of each APP, a natural language processing (NLP) model is used to generate a functional attribute description vector for each APP, and by calculating the distance between the functional attribute description vectors of each APP, each APP is clustered to determine the category to which each APP belongs.

[0017] In a possible implementation, constructing a user profile of the target user in the category to which each APP belongs according to the input history text of the target user in different APPs and the usage time of different APPs includes:

[0018] Extracting keywords from the input history texts of the target user in different APPs, and performing deduplication processing on the keywords to obtain deduplicated keywords;

[0019] The deduplicated keywords and the usage time of different APPs by the target user are input into a preset large language model in combination with a second prompt instruction prompt, so as to obtain a user portrait of the target user in the category to which each APP belongs output by the large language model.

[0020] In a possible implementation, the target input task is an input task involving a large language model; the overall user profile of the target user, combined with the first prompt instruction prompt, is input into a preset large language model, and according to the output content of the large language model, the personalized input content of the target user in the target input task is determined, including:

[0021] Determine the target scenario to which the target input task belongs and the target APP used by the target user, so as to select a user profile corresponding to the category to which the target APP belongs from the overall user profile of the target user;

[0022] The user portrait corresponding to the category to which the target APP belongs is combined with the first prompt instruction prompt and input into a preset large language model to obtain personalized input content output by the large language model for the target user when completing the target input task in the target APP.

[0023] In a possible implementation, the target input task is not an input task in which a large language model participates; the step of inputting the overall user profile of the target user, combined with the first prompt instruction prompt, into a preset large language model, and determining the personalized input content of the target user in the target input task according to the output content of the large language model includes:

[0024] The overall user portrait of the target user is combined with the first prompt instruction prompt, and input into a preset large language model, and according to the output content of the large language model, a personalized vocabulary of the target user in the category to which each APP belongs is generated;

[0025] Based on the personalized word library of the target user in each category to which the APP belongs and the category to which the APP corresponding to the target input task belongs, the personalized input content of the target user in the target input task is determined.

[0026] In a possible implementation, the target input task includes at least one of a word or sentence-based association task, a context-based automatic dialogue task, and an input completion task.

[0027] The embodiment of the present application also provides a personalized input device, including:

[0028] An acquisition unit, used to acquire historical input data of a target user in each application APP; and to construct a user profile of the target user in the category to which each APP belongs based on the historical input data;

[0029] A construction unit, configured to construct an overall user profile of the target user according to the user profiles of the target user in the categories to which each APP belongs;

[0030] A determination unit is used to input the overall user portrait of the target user, combined with a first prompt instruction prompt, into a preset large language model, and determine the personalized input content of the target user in the target input task according to the output content of the large language model.

[0031] In a possible implementation, the acquiring unit includes:

[0032] The clustering subunit is used to cluster each APP and determine the category to which each APP belongs;

[0033] A statistics subunit, used to count the input history texts of the target user in different APPs and the usage time of different APPs in the historical input data;

[0034] The construction subunit is used to construct a user portrait of the target user in the category to which each APP belongs based on the input history text of the target user in different APPs and the usage time of different APPs.

[0035] In a possible implementation, the clustering subunit is specifically used for:

[0036] According to the classification results of the APP market, each APP is clustered to determine the category to which each APP belongs;

[0037] Alternatively, clustering is performed on each APP based on the classification results of the APP market and the sub-classification results of each classification to determine the category to which each APP belongs;

[0038] Alternatively, based on the functional attribute description of each APP, a natural language processing (NLP) model is used to generate a functional attribute description vector for each APP, and by calculating the distance between the functional attribute description vectors of each APP, each APP is clustered to determine the category to which each APP belongs.

[0039] In a possible implementation, the construction subunit includes:

[0040] An extraction subunit is used to extract keywords from the input history texts of the target user in different apps, and perform deduplication processing on the keywords to obtain deduplicated keywords;

[0041] The first input sub-unit is used to input the deduplicated keywords, the usage time of the target user for different APPs, and a second prompt instruction prompt into a preset large language model to obtain a user portrait of the target user in the category to which each APP belongs output by the large language model.

[0042] In a possible implementation, the target input task is an input task involving a large language model; and the determining unit includes:

[0043] A first determination subunit is used to determine a target scenario to which the target input task belongs and a target APP used by the target user, so as to select a user portrait corresponding to the category to which the target APP belongs from the overall user portrait of the target user;

[0044] The second input sub-unit is used to input the user portrait corresponding to the category to which the target APP belongs, combined with the first prompt instruction prompt, into a preset large language model to obtain personalized input content output by the large language model for the target user when completing the target input task in the target APP.

[0045] In a possible implementation, the target input task is not an input task involving a large language model; and the determining unit includes:

[0046] A generating subunit, configured to input the overall user portrait of the target user, combined with the first prompt instruction prompt, into a preset large language model, and generate a personalized vocabulary of the target user in the category to which each APP belongs according to the output content of the large language model;

[0047] The second determining subunit is used to determine the personalized input content of the target user in the target input task based on the personalized vocabulary of the target user in each APP category and the APP category corresponding to the target input task.

[0048] In a possible implementation, the target input task includes at least one of a word or sentence-based association task, a context-based automatic dialogue task, and an input completion task.

[0049] The embodiment of the present application also provides a personalized input device, including: a processor, a memory, and a system bus;

[0050] The processor and the memory are connected via the system bus;

[0051] The memory is used to store one or more programs, and the one or more programs include instructions. When the instructions are executed by the processor, the processor executes any one of the implementations of the above-mentioned personalized input method.

[0052] An embodiment of the present application further provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any one of the implementations of the above-mentioned personalized input method.

[0053] The embodiment of the present application further provides a computer program product. When the computer program product is run on a terminal device, the terminal device executes any one of the implementation modes of the above-mentioned personalized input method.

[0054] A personalized input method, apparatus, storage medium and device provided in an embodiment of the present application first obtain historical input data of a target user in each application (Application, APP); and based on the historical input data, construct a user portrait of the target user in the category to which each APP belongs; then, based on the user portrait of the target user in the category to which each APP belongs, construct an overall user portrait of the target user; then, the overall user portrait of the target user is input into a preset large language model (large language model, LLM for short) in combination with a first prompt instruction prompt, and the personalized input content of the target user in the target input task is determined based on the output content of the large language model.

[0055] It can be seen that since the personality traits exhibited by the same person in different scenarios (or APPs) are not completely consistent, this application first constructs a user portrait of the target user in the specific category of each APP based on the historical input data of the target user in the APPs in different scenarios, and then combines the user portrait in the specific category of each APP with the large language model and applies it to the target input tasks in various scenarios, thereby improving the personalized input effect of the target user in the target input task, thereby improving the target user's input experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0057] Figure 1 A flowchart of a personalized input method provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the overall process of personalized input provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the composition of a personalized input device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] With the continuous development and popularization of the Internet and artificial intelligence, the mode of human-computer interaction is also constantly evolving. However, as the most traditional input method, typing is still the most widely accepted and popular interaction mode.

[0061] However, with the continuous development of input methods, the basic pinyin-to-text input method has become mature, and the effects of various commonly used input method manufacturers are highly homogenized, and it is difficult to form a generational leading advantage. Correspondingly, it is difficult for users to achieve significant improvement and progress in input experience. At present, in order to improve the user's input experience, the following two methods are usually used to achieve personalized input for different users:

[0062] (1) First label the user, and then match the existing domain vocabulary to achieve personalized input.

[0063] Specifically, this method requires building vocabulary in different fields in advance. Generally speaking, it is divided according to the major categories of the field first, and then the sub-lexicon is subdivided for each different field. For example, the major categories of the field may include: natural sciences, humanities, life, sports, video games, medicine, etc. Each major category will also be subdivided into subcategories, such as medicine will be divided into different subcategories such as traditional Chinese medicine, acupuncture, Western medicine, surgery, etc. Because each subcategory has some special words that are different from other subcategories, the distinction between them is still relatively strong. When a user who accurately matches a specific vocabulary enters a specific word under this subcategory, the input method will give a higher score to the entries with similar pronunciation in the vocabulary when decoding the pinyin, so that the entry candidate will be positioned higher and easier to choose. There are two difficulties in this method of building a vocabulary in advance. One is that it is necessary to build and maintain a huge classified vocabulary. This requires a dedicated team to regularly update the vocabulary according to the new hot words and new words in various fields, which requires huge human and material resources. Another difficulty is how to tag users and match fixed user lexicons with users for loading. The general method is to select when new users register their accounts, or to manually add domain lexicons during use. However, both methods cannot achieve the ideal matching effect.

[0064] It can be seen that in the user personalized input solution of building a domain vocabulary, in addition to building and maintaining a large-scale domain vocabulary, the biggest problem is how to match users with domain vocabulary. This problem has two aspects. The first is how to bind users and domains and find a suitable domain vocabulary; the second is the problem that too many domain vocabulary will affect the decoding effect.

[0065] For the existing domain word library binding scheme, it is generally mainly the user's own choice. In this link, the user may be impatient or unprofessional, and the vocabulary selected is not highly matched, or no choice is made at all. This requires an algorithm to match the domain word library with a specific user. Some algorithms use some models combined with the user's historical input to label the user, thereby building a user portrait. However, because the dimension of ordinary users' contact with the world is not very low, many fields will also be involved. This simple labeling method often labels the user too much, thereby loading many domain words. If too many domain words are bound to a user, it will affect the decoding effect. Because each word is a specially weighted word, the decoding order of normal words is later, or the weights between domain words are equal, resulting in the problem that each domain word is easy to fail to come out.

[0066] (2) Directly build a personalized vocabulary library for each person to achieve personalized input.

[0067] Specifically, this method needs to analyze the user's historical input, extract the words with a relatively high probability of appearing in the historical input, and use a variety of methods including term frequency-inverse document frequency (TF-IDF) and clustering to obtain a user's common words in a period of time. Then, the word library is screened and processed to build a word library dedicated to each user.

[0068] It can be seen that the problem with building a user-specific vocabulary is that it is expensive. It is necessary to constantly and regularly analyze the input history of each user to update the user vocabulary, and the maintenance cost is high. Therefore, some solutions will also label users to build user portraits. By building user portraits, the vocabulary of the user's frequently used fields is bound to the user, and the domain vocabulary is used to expand the user vocabulary to achieve the integration of the two methods. However, although these existing methods of realizing personalized input for different users have certain effects, they also have certain limitations and cannot achieve the ideal effect, which reduces the user's input experience.

[0069] To solve the above-mentioned defects, the present application provides a personalized input method, which first obtains the historical input data of the target user in each application APP; and based on the historical input data, constructs a user portrait of the target user in the category to which each APP belongs; then, based on the user portrait of the target user in the category to which each APP belongs, constructs an overall user portrait of the target user; then, the overall user portrait of the target user is combined with a first prompt instruction prompt, and input into a preset large language model, and based on the output content of the large language model, determines the personalized input content of the target user in the target input task.

[0070] It can be seen that since the personality traits exhibited by the same person in different scenarios (or APPs) are not completely consistent, this application first constructs a user portrait of the target user in the specific category of each APP based on the historical input data of the target user in the APPs in different scenarios, and then combines the user portrait in the specific category of each APP with the large language model and applies it to the target input tasks in various scenarios, thereby improving the personalized input effect of the target user in the target input task, thereby improving the target user's input experience.

[0071] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0072] First embodiment

[0073] See also Figure 1 , is a flow chart of a personalized input method provided in this embodiment, the method comprising the following steps:

[0074] S101: Obtain historical input data of the target user in each application APP; and construct a user profile of the target user in the category to which each APP belongs based on the historical input data.

[0075] In this embodiment, any user who realizes personalized input through the personalized input method provided by the present application is defined as a target user. In addition, it should be noted that the applicant believes that each person's user portrait and personality characteristics are not single. It is unreasonable to simply construct a user portrait and personality characteristics description of a person based on the user's historical input as in the existing method. For example, some people are honest and diligent at work, with a mild temper, but they may become irritable in the field of games, and become a caring boyfriend or a nanny, a nanny, etc. in the shopping scene. In other words, the personality characteristics of the same person in different scenes may sometimes be mutually exclusive. Therefore, it is unreasonable to only construct a set of user portraits to realize personalized input as in the existing method.

[0076] In this regard, this application proposes to construct personalized user portraits of users (such as target users) in different fields and scenarios (such as different types of APPs) based on the input characteristics and personality characteristics of users (such as target users) in different fields. And combined with the large language model LLM in different field scenarios (such as different types of APPs), give users (such as target users) different input experiences to improve personalized input effects. In this way, not only user-level personalization is achieved, but also personalization of a user (such as a target user) in different field scenarios (such as different types of APPs) is achieved, thereby improving the input experience of users (such as target users).

[0077] Among them, the large language model LLM can be a language model based on deep learning, which can generate new language expressions such as text, sentences, paragraphs and even articles based on the input text content. The large language model LLM uses a large-scale language data set to train language rules and patterns through autoregressive generation, and can simulate human instructions to generate language expressions (such as text data). Specifically, when the large language model LLM generates new text data, it predicts the possibility of the next language unit based on the previously generated content until the complete text data is generated.

[0078] Specifically, in order to improve the personalization effect of the target user and thus enhance his input experience, it is first necessary to obtain the historical input data of the target user in each APP with the permission of the target user (the specific content is not limited, and may include but is not limited to the historical input content and the target user's usage time for each APP, etc.), and then construct a user portrait of the target user in the category to which each APP belongs based on the obtained historical input data to execute the subsequent step S102.

[0079] Among them, it should be noted that the user portraits of the target users in the categories to which each APP belongs constructed by this application are different from the “user portraits” mentioned in the existing methods. The user portraits generated in the existing methods refer to labeling users, such as users are “sports enthusiasts”, “academic tyrants”, “like history”, etc. However, the basic idea of ​​this application to construct user portraits of target users in the categories to which each APP belongs is that the personality characteristics of the same person (such as the target user) in each different scenario field may not be the same, so their personality characteristics are described for different fields. Therefore, the construction method adopted by this application is to summarize and characterize the personality characteristics of the target users in different field scenarios, forming a series of non-label summary descriptions as user portraits of the target users in the categories to which each APP belongs, which are used to perform subsequent target input tasks.

[0080] In an optional implementation manner, the implementation process of "building a user profile of the target user in each category to which the APP belongs according to the historical input data" in step S101 may include the following steps AC:

[0081] Step A: Cluster each APP to determine the category to which each APP belongs.

[0082] It should be noted that each person (such as the target user) has different preferences for APPs due to factors such as personality, gender, education level and hobbies. For example, a girl who loves shopping may prefer shopping APPs with shopping functions, a sophisticated girl may like APPs with both sharing and shopping functions, and a boy who likes to watch football games may like APPs that can provide sports news, live broadcasts, forums and other functions. Each specific person may also show different personalities in different social fields. For example, a person may behave maturely and steadily at work, but may be a fanatical fan in the basketball circle, and his replies and comments may be sharp and excited. In shopping APPs, he may behave more at home and buy some baby products and household products. Therefore, fixed labels cannot fully represent a person's personality traits. It is necessary to evaluate a person's personality traits in each type of scenario to build a more reasonable user portrait.

[0083] Therefore, in this implementation, in order to accurately construct the user portrait of the target user in the category to which each APP belongs, the first step is to cluster each APP with each APP as the granularity to determine the category to which each APP belongs, that is, to merge the categories of APPs with similar attributes to reduce the categories of field classification. For example, there are many APPs with the same functional attributes, but different names, etc., that is, to determine that these APPs belong to the same category.

[0084] Moreover, the present application does not limit the method for clustering each APP. An optional implementation method is that each APP can be clustered according to the classification results of the APP market to determine the category to which each APP belongs. Alternatively, each APP can be clustered according to the classification results of the APP market and the sub-classification results of each classification to determine the category to which each APP belongs. This is because the APP market often divides a class of APPs together and marks a label. Therefore, the classification method under these label categories can be used directly, or the APPs within a label can be further classified. Alternatively, the functional attribute description vectors of each APP can be generated by using the natural language processing (Natural Language Processing, NLP) model (such as the Bert model, etc.) according to the functional attribute description of each APP, and the distance between the functional attribute description vectors of each APP can be calculated to cluster each APP to determine the category to which each APP belongs, such as APP clusters with similar vector distances can be regarded as one category. In this way, APPs of similar field categories can be clustered into one category through these methods.

[0085] For example: Assume that the target user has eight apps installed on his mobile phone, namely "Alibaba, JD.com, MTMall, WeChat, Book, Hu, PUBG, and Che". According to the first classification method mentioned in the previous paragraph, they can be divided into three categories according to the classification of the APP market, namely: shopping software (Alibaba, JD.com, MTMall), social software (Book, Hu, PUBG) and other software (Che). Alternatively, according to the first classification method mentioned in the previous paragraph, the APPs within a label can be further classified and refined into four categories: shopping software (Alibaba, JD.com, MTMall), instant messaging software (Book, Hu, PUBG), social interest platforms (Book, Hu, PUBG) and other software (Che). Finally, if necessary, according to the third classification method mentioned in the previous paragraph, the APP descriptions can be clustered and divided. For example, if the descriptions of MTMall and Alibaba both include that they belong to the same parent company, they may be clustered into one category and distinguished from the Alibaba APP. In this case, it is possible to divide it into five categories: shopping software category 1 (Jindong), shopping software category 2 (Taobao, Tmall), instant messaging software (Xinxin), social interest platform (Shushu, Huhu, Pu) and other software (Chemou).

[0086] Step B: Count the target user's input history texts in different apps and the usage time of different apps in the historical input data.

[0087] In this implementation, after determining the category to which each APP belongs through step A, the input behavior of the target user in each category of APP (specifically, the historical input data generated) can be further analyzed, such as counting the target user's input history text in different APPs and the usage time of different APPs in the historical input data, etc., to execute the subsequent step C.

[0088] It should be noted that the historical input data of the target user in each APP obtained by this application is obtained only after the target user agrees, and the purpose is to improve the input experience of the target user, not to steal the target user's personal information. In addition, in order to protect the target user's information, the obtained information will not be uploaded to the cloud, but will only be calculated locally.

[0089] It should also be noted that the information considered in this application includes the input history input data of the target user in each APP category group. Of course, it is not limited to this information. All information that can reflect the input characteristics of users in different APP groups without involving user sensitive information can be considered for use. In addition, since the personalized input method proposed in the application is for each type of APP cluster, the input characteristics of the target user on it are counted, so information similar to the input history text of the target user in different APPs and the use time of different APPs in the historical input data will be collected. In this way, by executing the subsequent step C, it is possible to roughly determine which user group the target user belongs to based on the target user's use time of different APPs. Based on the input history text of the target user in different APPs, the use characteristics of the target user in the category cluster to which each APP belongs can be further extracted, thereby constructing a clearer user portrait of the target user in the category to which each APP belongs.

[0090] Step C: Build a user profile of the target user in the category to which each APP belongs based on the target user's input history text in different APPs and the usage time of different APPs.

[0091] In this implementation, after the input history text of the target user in different APPs and the use time of different APPs in the historical input data are counted by step B, it is further possible to start with simple features, for example, the approximate gender, hobbies and occupations can be judged according to the APPs commonly used by the user (i.e., the use time of the APP is longer). For example, assuming that in the social interest platform (a certain book, a certain book, a certain book) determined by step A, generally speaking, girls will prefer a certain book APP, and boys will prefer a certain book and a certain book APP. There may also be boys who will install a certain book APP, but the number of times it is opened and the use time are less, and the time spent on a certain book and a certain book APP is very much. Then more information can be obtained through the use time. For example, if the target user uses a certain book and a certain book APP for a long time to discuss sports topics, it is more likely to be a boy. Then, according to the input history text of the target user in different APPs, a more accurate user portrait of the target user in the category to which different APPs belong can be constructed.

[0092] Specifically, an optional implementation method is that after counting the target user's input history texts in different APPs and the usage time of different APPs, the target user can first extract keywords from the input history texts in different APPs, and deduplicate these keywords to obtain deduplicated keywords, and then integrate the deduplicated keywords and the target user's usage time for different APPs into the prompt instruction prompt (defined as the second prompt instruction here), and input it into the preset large language model to use the large language model LLM output for the target user in the category to which each APP belongs. Among them, the second prompt instruction prompt can be written in many ways, and the specific content is not limited in this application.

[0093] For example, if the target user is in a scenario where he is using a shopping app, the second prompt can be: "This user is using a 'shopping' app. The words he often enters in this scenario include 'baby carriage, pacifier, milk powder, soy sauce, flour', etc. Please use a few sentences to speculate on the possible personality characteristics of this user. Try to be concise and avoid repetition." Through prompt content like this, a description of the target user's user profile in the shopping app category can be generated based on the words that the target user has entered in the past and the output of the large language model LLM.

[0094] S102: Construct an overall user profile of the target user based on the user profile of the target user in the category to which each APP belongs.

[0095] In this embodiment, after constructing the user portrait of the target user in each category to which the APPs belong through step S101, the overall user portrait of the target user can be further constructed based on the user portrait of the target user in each category to which the APPs belong, so as to execute the subsequent step S103.

[0096] Specifically, after constructing the user portrait of the target user in each APP category, the user portraits in each APP category can be further integrated to form a complete user description. Of course, the dimensions of each person's overall user portrait are also different, because the categories of the APPs involved may be different. For example, a male's commonly used APPs include "Taobao, JD.com, Huhu, PUBG, and Xinxin". His overall user portrait may be "This is a male who likes shopping, sports, consumption, and social networks. In the shopping scenario, he is a good family man. He may have a baby and will consider buying baby products and household items, and occasionally buy some electronic products. In the sports community, he is an avid fan of the xx team, likes football, likes xx players, and is willing to spend time watching football games. In the knowledge community, he is an emotional expert who browses a large number of emotional articles and spends time answering emotional questions."

[0097] S103: The overall user portrait of the target user is input into a preset large language model in combination with the first prompt instruction prompt, and the personalized input content of the target user in the target input task is determined according to the output content of the large language model.

[0098] In this embodiment, after constructing the overall user portrait of the target user through step S102, the overall user portrait of the target user (including the user portrait of the target user in each category to which the APP belongs) can be further used in more different input tasks (defined as target input here) to enhance the input experience of the target user. Specifically, the overall user portrait of the target user can be integrated into the first prompt instruction prompt and input into the preset large language model LLM, so as to utilize the output content of the large language model LLM to more accurately predict the personalized input content of the target user in the target input task.

[0099] Among them, this application does not limit the specific content and type of the target input task, which can be set according to actual conditions and experience. For example, the target input task may include but is not limited to word or sentence-based association tasks, context-based automatic dialogue tasks, input completion tasks, etc.

[0100] Specifically, an optional implementation method is that when the target input task is an input task involving a large language model LLM, the implementation process of this step S103 may include: first, determining the target scenario to which the target input task belongs and the target APP used by the target user, so as to select the user portrait corresponding to the category to which the target APP belongs from the overall user portrait of the target user.

[0101] Among them, it should be noted that since the overall user portrait of the target user pre-generated in this application includes the user portrait of the target user in each APP category (i.e., different application scenarios) (describing the personality characteristics of the target user in each different scenario), these user portraits can be used in different scenarios according to the different scenarios to which the target input task belongs. For example, for the user portrait of the target user in the shopping APP category, the effect of using it in the shopping scenario will be better. Therefore, it is necessary to first target the target scenario to which the target input task belongs and the target APP used by the target user. Then select the user portrait corresponding to the category to which the current APP (i.e., the target APP) belongs from the overall user portrait of the target user. After that, it can be integrated into various target input tasks (including but not limited to the task of generating the following dialogue based on the previous text, the task of converting the chat style, the generative decoding task, etc.) for use.

[0102] Then, in this implementation, since the target input task is an input task involving the large language model LLM, the user portrait corresponding to the category to which the current APP (i.e., the target APP) belongs can be selected from the overall user portrait of the target user and embedded into a specific position in the first prompt instruction prompt, and input into the preset large language model to obtain the personalized input content output by the large language model for the target user to complete the target input task in the target APP. In this way, the reasoning and understanding ability of the large language model can be used to bring new input experiences, such as conversion of chat styles and automatic chat, thereby achieving personalized input effects.

[0103] For example: suppose the target input task is to use the user portrait of the target user in a shopping scenario to perform a context-based association task. At this time, the first prompt instruction prompt can be written in many ways. An example may be: "The user is currently in a 'shopping' scenario, and the previous sentence history entered is 'XXX'. His user portrait is 'XXXXXX', etc. Please predict what may be entered in the second half of this sentence." Through prompt content similar to this, before the user (i.e., the target user) has entered the second half of the sentence, possible options can be given based on his input habits and the reasoning and understanding capabilities of the large language model, and provided to the target user, so as to further improve the user input experience.

[0104] Another optional implementation method is that when the target input task is not an input task involving a large language model, the implementation process of this step S103 may include: first, the overall user portrait of the target user is input into the preset large language model in combination with the first prompt instruction prompt, and the personalized word library of the target user in each APP category is generated according to the output content of the large language model. Then, based on the personalized word library of the target user in each APP category and the APP category corresponding to the target input task, the personalized input content of the target user in the target input task is determined.

[0105] In this implementation, since the target input task is not an input task involving a large language model, a personalized vocabulary library for the target user in each APP category can be generated directly based on the user portrait of the target user in each APP category contained in the target user's overall user portrait, combined with the large language model LLM, and loaded into the target user's exclusive personalized vocabulary library, thereby simplifying the target user's exclusive vocabulary library construction process and improving the target user's experience of using traditional input methods.

[0106] Specifically, different from the existing method of building a personalized vocabulary by collecting the words commonly used by users in history, this application uses the user portraits of the target users in the categories to which each APP belongs, and generates a personalized vocabulary for the target users in the categories to which different APPs belong through a non-real-time large language model. This method is low-cost, can reuse the generated user portraits, and can also achieve separate coverage on different APPs. This can be distinguished from the traditional user personalized vocabulary generation solution, which uses similar vocabulary for different APPs. And after this application generates the personalized vocabulary for the target user in the categories to which each APP belongs, it can be directly loaded during the traditional pinyin decoding process without the real-time participation of the large language model, thereby improving the input efficiency of the target user.

[0107] To facilitate understanding of the personalized input method proposed in this application, this application also provides a schematic diagram of the overall implementation process of personalized input. The overall implementation process is based on the fact that the personality characteristics of the same person in different scenarios (or APPs) are not completely consistent. The user portrait description of the target in each specific field is inferred and constructed based on the historical input data of the target user in the APP in different scenarios. And the user portrait is combined with the large language model and applied to the target input tasks in various scenarios to achieve a personalized input experience for the target user. Among them, the historical input data of the target user in the APP in different scenarios is stored locally in this application and will not be uploaded to the cloud server. In addition, this personalized function will only be put into use after the target user agrees to the user privacy terms, and can be actively enabled by the target user, not by default.

[0108] Specifically, Figure 2 As shown, first, target users in different APP categories (such as Figure 2 The historical input data (including but not limited to input history text and usage time of different APPs) in the M category to which the target user belongs is analyzed, and the user portrait of the target user in different APP categories is determined using the large language model LLM. Then, the generated user portraits of the target user in different APP categories are applied to different input tasks based on the large language model, such as association, dialogue, rewriting and completion tasks, to rewrite the prompt of the input large language model to obtain personalized input results. The generated user portraits of the target user in different APP categories can also be used to build a personalized vocabulary library exclusive to the target user. Thereby reducing the complexity of input, and by learning some commonly used sentence patterns and input habits of the target users, the target users can truly have some different experiences in the input process, making the target users feel that input is simpler and more convenient.

[0109] In summary, the personalized input method provided in this embodiment first obtains the historical input data of the target user in each application (Application, APP); and based on the historical input data, constructs a user portrait of the target user in the category to which each APP belongs; then, based on the user portrait of the target user in the category to which each APP belongs, constructs an overall user portrait of the target user; then, the overall user portrait of the target user is combined with the first prompt instruction prompt, and input into a preset large language model (large language model, LLM for short), and determines the personalized input content of the target user in the target input task based on the output content of the large language model.

[0110] It can be seen that since the personality traits exhibited by the same person in different scenarios (or APPs) are not completely consistent, this application first constructs a user portrait of the target user in the specific category of each APP based on the historical input data of the target user in the APPs in different scenarios, and then combines the user portrait in the specific category of each APP with the large language model and applies it to the target input tasks in various scenarios, thereby improving the personalized input effect of the target user in the target input task, thereby improving the target user's input experience.

[0111] Second embodiment

[0112] This embodiment will introduce a personalized input device. For related content, please refer to the above method embodiment.

[0113] See also Figure 3, is a schematic diagram of the composition of a personalized input device provided in this embodiment, the device 300 includes:

[0114] The acquisition unit 301 is used to acquire the historical input data of the target user in each application APP; and to construct a user profile of the target user in the category to which each APP belongs based on the historical input data;

[0115] A construction unit 302 is used to construct an overall user profile of the target user according to the user profiles of the target user in the categories to which each APP belongs;

[0116] The determination unit 303 is used to input the overall user portrait of the target user, combined with the first prompt instruction prompt, into a preset large language model, and determine the personalized input content of the target user in the target input task according to the output content of the large language model.

[0117] In an implementation of this embodiment, the acquiring unit 301 includes:

[0118] The clustering subunit is used to cluster each APP and determine the category to which each APP belongs;

[0119] A statistics subunit, used to count the input history texts of the target user in different APPs and the usage time of different APPs in the historical input data;

[0120] The construction subunit is used to construct a user portrait of the target user in the category to which each APP belongs based on the input history text of the target user in different APPs and the usage time of different APPs.

[0121] In an implementation of this embodiment, the clustering subunit is specifically used for:

[0122] According to the classification results of the APP market, each APP is clustered to determine the category to which each APP belongs;

[0123] Alternatively, clustering is performed on each APP based on the classification results of the APP market and the sub-classification results of each classification to determine the category to which each APP belongs;

[0124] Alternatively, based on the functional attribute description of each APP, a natural language processing (NLP) model is used to generate a functional attribute description vector for each APP, and by calculating the distance between the functional attribute description vectors of each APP, each APP is clustered to determine the category to which each APP belongs.

[0125] In one implementation of this embodiment, the construction subunit includes:

[0126] An extraction subunit is used to extract keywords from the input history texts of the target user in different apps, and perform deduplication processing on the keywords to obtain deduplicated keywords;

[0127] The first input sub-unit is used to input the deduplicated keywords, the usage time of the target user for different APPs, and a second prompt instruction prompt into a preset large language model to obtain a user portrait of the target user in the category to which each APP belongs output by the large language model.

[0128] In an implementation of this embodiment, the target input task is an input task involving a large language model; the determining unit 303 includes:

[0129] A first determination subunit is used to determine a target scenario to which the target input task belongs and a target APP used by the target user, so as to select a user portrait corresponding to the category to which the target APP belongs from the overall user portrait of the target user;

[0130] The second input sub-unit is used to input the user portrait corresponding to the category to which the target APP belongs, combined with the first prompt instruction prompt, into a preset large language model to obtain personalized input content output by the large language model for the target user when completing the target input task in the target APP.

[0131] In an implementation of this embodiment, the target input task is not an input task involving a large language model; the determining unit 303 includes:

[0132] A generating subunit, configured to input the overall user portrait of the target user, combined with the first prompt instruction prompt, into a preset large language model, and generate a personalized vocabulary of the target user in the category to which each APP belongs according to the output content of the large language model;

[0133] The second determining subunit is used to determine the personalized input content of the target user in the target input task based on the personalized vocabulary of the target user in each APP category and the APP category corresponding to the target input task.

[0134] In an implementation of this embodiment, the target input task includes at least one of a word or sentence-based association task, a context-based automatic dialogue task, and an input completion task.

[0135] Furthermore, an embodiment of the present application also provides a personalized input device, including: a processor, a memory, and a system bus;

[0136] The processor and the memory are connected via the system bus;

[0137] The memory is used to store one or more programs, and the one or more programs include instructions. When the instructions are executed by the processor, the processor executes any one of the implementation methods of the above-mentioned personalized input method.

[0138] Furthermore, an embodiment of the present application also provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes any implementation method of the above-mentioned personalized input method.

[0139] Furthermore, an embodiment of the present application also provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute any one of the implementation methods of the above-mentioned personalized input method.

[0140] It can be known from the description of the above implementation mode that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., including several instructions for enabling a computer device (which can be a personal computer, a server, or a network communication device such as a media gateway, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0141] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0142] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0143] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A personalized input method, characterized in that: include: Obtain the historical input data of the target user in each application APP; And based on the historical input data, construct a user profile of the target user in the category to which each APP belongs; Build an overall user profile of the target user based on the user profile of the target user in the category to which each APP belongs; The overall user portrait of the target user is combined with the first prompt instruction prompt and input into a preset large language model, and the personalized input content of the target user in the target input task is determined according to the output content of the large language model.

2. The method according to claim 1, characterized in that The step of constructing a user profile of the target user in each category to which each APP belongs according to the historical input data includes: Cluster each APP to determine the category to which each APP belongs; Count the target user's input history texts in different apps and usage time of different apps in the history input data; According to the input history text of the target user in different APPs and the usage time of different APPs, a user portrait of the target user in the category to which each APP belongs is constructed.

3. The method according to claim 2, characterized in that The clustering process of each APP to determine the category to which each APP belongs includes: According to the classification results of the APP market, each APP is clustered to determine the category to which each APP belongs; Alternatively, clustering is performed on each APP based on the classification results of the APP market and the sub-classification results of each classification to determine the category to which each APP belongs; Alternatively, based on the functional attribute description of each APP, a natural language processing (NLP) model is used to generate a functional attribute description vector for each APP, and by calculating the distance between the functional attribute description vectors of each APP, each APP is clustered to determine the category to which each APP belongs.

4. The method according to claim 2, characterized in that: The step of constructing a user profile of the target user in the category to which each APP belongs according to the input history text of the target user in different APPs and the usage time of different APPs includes: Extracting keywords from the input history texts of the target user in different APPs, and performing deduplication processing on the keywords to obtain deduplicated keywords; The deduplicated keywords and the usage time of different APPs by the target user are input into a preset large language model in combination with a second prompt instruction prompt, so as to obtain a user portrait of the target user in the category to which each APP belongs output by the large language model.

5. The method according to claim 1, characterized in that: The target input task is an input task involving a large language model; the overall user profile of the target user, combined with the first prompt instruction prompt, is input into a preset large language model, and according to the output content of the large language model, the personalized input content of the target user in the target input task is determined, including: Determine the target scenario to which the target input task belongs and the target APP used by the target user, so as to select a user profile corresponding to the category to which the target APP belongs from the overall user profile of the target user; The user portrait corresponding to the category to which the target APP belongs is combined with the first prompt instruction prompt and input into a preset large language model to obtain personalized input content output by the large language model for the target user when completing the target input task in the target APP.

6. The method according to claim 1, characterized in that The target input task is not an input task in which a large language model participates; the overall user profile of the target user, combined with the first prompt instruction prompt, is input into a preset large language model, and according to the output content of the large language model, the personalized input content of the target user in the target input task is determined, including: The overall user portrait of the target user is combined with the first prompt instruction prompt, and input into a preset large language model, and according to the output content of the large language model, a personalized vocabulary of the target user in the category to which each APP belongs is generated; Based on the personalized word library of the target user in each category to which the APP belongs and the category to which the APP corresponding to the target input task belongs, the personalized input content of the target user in the target input task is determined.

7. The method according to any one of claims 1 to 6, characterized in that: The target input task includes at least one of a word or sentence-based association task, a context-based automatic dialogue task, and an input completion task.

8. A personalized input device, characterized in that: include: An acquisition unit, used to acquire the historical input data of the target user in each application APP; And based on the historical input data, construct a user profile of the target user in the category to which each APP belongs; A construction unit, configured to construct an overall user profile of the target user according to the user profiles of the target user in the categories to which each APP belongs; A determination unit is used to input the overall user portrait of the target user, combined with a first prompt instruction prompt, into a preset large language model, and determine the personalized input content of the target user in the target input task according to the output content of the large language model.

9. A personalized input device, characterized in that: include: Processor, memory, system bus; The processor and the memory are connected via the system bus; The memory is used to store one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by the processor, the processor executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the method according to any one of claims 1 to 7.

Images