Voice dialogue version intelligent question answering system based on large model offline knowledge base

Abstract

The invention discloses a voice dialogue version intelligent question answering system based on a large model offline knowledge base, and relates to the technical field of informatization management, the voice dialogue version intelligent question answering system comprises an information collection module, an offline database, a voice recognition module, a natural language processing module, a data retrieval module and a voice playing module, a data text is generated after processing and stored in an offline database, a system converts voice input of a user into a user voice text, a natural language processing module analyzes the user voice text, an adaptive data text is screened from the offline database, the playing frequency of the adaptive data text is retrieved, and a priority answer score is analyzed. The answer data text is converted into voice and then played to the user, when the user inputs the answer data text through voice and does not solve the problem of the user, the data text is sequentially played through voice according to the priority answer scores from high to low, the off-line knowledge base is constructed and optimized, and the system can normally operate and provide accurate question and answer service in the environment without a network.

Description

Technical Field

[0001] The present invention relates to the technical field of voice intelligent question answering, and specifically to a voice dialogue version intelligent question answering system based on a large model offline knowledge base. Background Art

[0002] With the hot use of big models, various companies have also begun to use big model services. In actual use, due to the privacy and importance of corporate information and internal knowledge, leakage will cause immeasurable losses to the company. Therefore, most companies will choose to use privately deployed models, which solves the problem of information leakage. This is the privately deployed enterprise knowledge base. However, these big models usually need to be trained and inferred in the cloud, which limits their application in some specific scenarios, such as scenarios that require high security, low latency or specific hardware environments.

[0003] In the existing market, users need to manually input query content through keyboard or touch screen for question and answer, which is extremely inconvenient and inefficient in some scenarios (such as driving or walking), and brings poor user experience. On the one hand, the existing intelligent question and answer system often relies on Internet connection to obtain real-time data or update the knowledge base, which cannot be used in an environment without network. It is necessary to use the network to search for online document knowledge and integrate answers. In the case of poor network or environment that does not allow external network connection, the function is greatly limited. On the one hand, even if some systems support offline knowledge base, they often lack internal learning and self-optimization capabilities, and the knowledge base is updated slowly and costly. Problems and solutions generated at the operation and maintenance site cannot be synchronized in time, and the knowledge base is relatively lagging. On the other hand, for existing systems, the query results are often displayed directly, and the entire document has a large amount of content. Users need to search from it, and it is difficult to find key document content.

[0004] Therefore, people need a voice dialogue version of intelligent question and answer system based on a large model offline knowledge base to solve the above problems. Summary of the invention

[0005] The purpose of the present invention is to provide a voice dialogue version intelligent question and answer system based on a large model offline knowledge base to solve the problems raised in the above background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: a voice dialogue version intelligent question and answer system based on a large model offline knowledge base, the system comprising: an information collection module, an offline database, a voice recognition module, a natural language processing module, a data retrieval module and a voice playback module;

[0007] The information collection module collects company information, generates information text after processing and stores it in an offline database;

[0008] Converting the user's voice input into user voice text through the voice recognition module;

[0009] Analyze the user's voice text through the natural language processing module and select the adapted material text from the offline database;

[0010] Retrieving the number of times the adaptation material text is played through the data retrieval module, analyzing the priority answer score in the adaptation material text in consideration of the number of times the text is played, and using the adaptation material text with the highest priority answer score as the answer material text;

[0011] The voice playback module converts the answer material text into voice and then plays it to the user. When the user's voice input of the answer material text does not solve the user's problem, the voice playback module plays the material text in order of priority answer scores from high to low.

[0012] Furthermore, the information collection module manually collects company information from documents, manuals, and FAQ lists, and pre-processes the information to generate information text. The company information is collected after authorization is granted. The pre-processing process includes cleaning data, removing noise, and standardizing the format. The information collection module stores the processed information text in an offline database. Through data pre-processing, a series of processing is performed on the original data in the processes of data mining, data analysis, and machine learning to improve data quality, reduce noise, and facilitate the extraction of useful information, and convert the original data into useful data for subsequent data analysis. Data pre-processing includes: data deduplication, removing duplicate records in the data set, and achieving this by comparing unique identifiers or key fields in the records; missing value processing, filling missing values ​​in the data set, and processing using interpolation, mean, median, mode, and other methods; outlier processing, detecting and processing outliers in the data set, and the outliers are deleted or replaced with acceptable values; data standardization, standardizing the data format to a consistent format for easy processing and analysis; data conversion, data conversion is essentially converting the format of the data, and the main purpose is to facilitate data processing and analysis.

[0013] Furthermore, the offline database is used as an offline knowledge base. It does not need to be connected to an external network during use. An offline database is built to ensure information security and to ensure that key information can be permanently stored in a accessible location. The intelligent refinement capability of the large model is used to accurately refine and simplify the query results, ensure the length of time for voice answers, reduce invalid information for the questioner, introduce a self-learning and optimization mechanism, automatically update and optimize the knowledge base and question-answering model based on user feedback and usage, and promptly update the offline database when new problems or new solutions arise on the operation and maintenance line.

[0014] Furthermore, the speech recognition module converts the user's speech input into user speech text through an algorithm based on deep learning. Front-end processing: This step includes noise reduction, echo cancellation and automatic gain control, the purpose of which is to improve the quality of the speech signal; Feature extraction: Extract useful features from the audio signal, such as Mel-frequency cepstral coefficients (MFCC) or linear predictive coding (LPC); Acoustic model: Use statistical models (such as hidden Markov models HMMs or deep neural networks DNNs) to recognize speech fragments; Language model: Used to estimate the possibility of a given sentence, using an n-gram model or a neural network model; Decoder: Combine the acoustic model and language model to find the most likely text sequence.

[0015] Furthermore, the natural language processing module analyzes the user's voice text to extract nouns A, verbs B and adjectives C contained in the text, and the system uses natural language generation (NLG) to convert the internal representation into natural language text.

[0016] Furthermore, the natural language processing module in any document text F i Search for noun A, verb B, and adjective C in the result, and find that noun A appears a number of times. i , verb B appears B times i , adjective C appears C times i , when the number of characters between noun A and adjective C is less than G, it is recorded as an accurate noun phrase. The total number of characters is D i group of precise noun phrases, where i represents the i-th document text in the I-th document text in the offline database, and calculates F i Adaptation score E for user voice text i :

[0017] E i =K A *A i +K B *B i +K C *C i +K D *D i ;

[0018] Where G is the number of characters used to determine whether a noun and an adjective form a precise noun phrase, and K is A is the weight of the given adaptation score noun, K B is the existing adaptation scoring verb weight, K C is the existing adaptation score adjective weight, K D Calculate the weight of the precise noun phrase K for the precise noun phrase to be calculated. D :

[0019]

[0020] Substitute {i=1, 2, ..., I} one by one, calculate the adaptation score {F1, F2, ..., F I}, select the H data texts with the highest adaptation scores as the adaptation data texts adapted to the user's voice text, denoted as {J1, J2, ..., J H}, mark the parts of speech, analyze the degree of adaptation by part of speech, and accurately determine whether the data text can solve the user's problem.

[0021] Furthermore, the data retrieval module retrieves the historical playback times of the adaptation material texts {J1, J2, ..., JH}, where the hth adaptation material text J h The number of times played is L h , considering the number of playbacks to calculate the data text J h Priority answer rating W h :

[0022] W h =k E *E h +k L *L h ;

[0023] where k E For data text J h Score W for the priority answer h The adaptation score weight, k L For data text J h Score W for the priority answer h The weight of the number of times a message is played is taken into account. Considering that for any enterprise, historical commonly used information is often the most valuable and easiest to search for information, the number of times a message is played is taken into account here to calculate the priority answer score of the information text.

[0024] Furthermore, {h=1, 2, ..., H} are substituted one by one to calculate the priority answer scores {W1, W2, ..., W H}, sort the H adaptation material texts into {M1, M2, ..., M H}.

[0025] Furthermore, the voice playback module converts the adaptation document text M1 into voice through NLG and then plays it to the user. When the user voice inputs the answer document text but does not solve the user's problem, the voice playback module voice plays the document text in descending order according to the priority answer score until the user issues a voice command that the answer has been obtained. Considering that the number of document texts for the system to answer the question is H, the documents are sorted here and voice answers are provided to the user in descending order according to the priority answer score.

[0026] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: in order to solve the trouble of manual operation by users, the system is connected to voice recording and voice recognition technology, so that voice questions can be asked to input query content, thereby improving the convenience of user operation and enhancing the user experience; building and optimizing an offline knowledge base, so that the system can operate normally and provide accurate question and answer services even in a networkless environment; introducing a self-learning and optimization mechanism, automatically updating and optimizing the knowledge base and question and answer model according to user feedback and usage, summarizing online operation and maintenance problems and solutions, and updating them to the knowledge base in a timely and synchronous manner for easy query and use again; using the intelligent refining capability of the large model to accurately refine and simplify the query results, ensuring that users can quickly obtain useful information, and screening and refining the output answers to summarize useful information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0028] Figure 1 It is a structural diagram of the voice dialogue version intelligent question and answer system based on a large model offline knowledge base of the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] See also Figure 1 ,The present invention provides a technical solution: a voice dialogue version intelligent question and answer system based on a large model offline knowledge base, the system includes: an information collection module, an offline database, a voice recognition module, a natural language processing module, a data retrieval module and a voice playback module;

[0031] The information collection module collects company information, generates information text after processing and stores it in an offline database;

[0032] Converting the user's voice input into user voice text through the voice recognition module;

[0033] Analyze the user's voice text through the natural language processing module and select the adapted material text from the offline database;

[0034] Retrieving the number of times the adaptation material text is played through the data retrieval module, analyzing the priority answer score in the adaptation material text in consideration of the number of times the text is played, and using the adaptation material text with the highest priority answer score as the answer material text;

[0035] The voice playback module converts the answer material text into voice and then plays it to the user. When the user's voice input of the answer material text does not solve the user's problem, the voice playback module plays the material text in order of priority answer scores from high to low.

[0036] The information collection module manually collects company information from documents, manuals, and FAQ lists, and pre-processes the information to generate information text. The pre-processing process includes cleaning data, removing noise, and standardizing the format. The information collection module stores the processed information text in an offline database. Through data pre-processing, a series of processing is performed on the original data in the processes of data mining, data analysis, and machine learning to improve data quality, reduce noise, and facilitate the extraction of useful information, and convert the original data into useful data for subsequent data analysis. Data pre-processing includes: data deduplication, removing duplicate records in the data set, and achieving this by comparing unique identifiers or key fields in the records; missing value processing, filling in missing values ​​in the data set, and processing using interpolation, mean, median, mode, and other methods; outlier processing, detecting and processing outliers in the data set, and the outliers are deleted or replaced with acceptable values; data standardization, standardizing the data format to a consistent format for easy processing and analysis; data conversion, data conversion is essentially converting the format of the data, and the main purpose is to facilitate data processing and analysis.

[0037] The offline database is used as an offline knowledge base. It does not need to be connected to an external network during use. An offline database is built to ensure information security and to ensure that key information can be permanently stored in a location where it can be consulted. The intelligent refinement capability of the large model is used to accurately refine and simplify the query results, ensure the length of time for voice answers, reduce invalid information for the questioner, introduce a self-learning and optimization mechanism, automatically update and optimize the knowledge base and question-answering model based on user feedback and usage, and update the offline database in a timely manner when new problems or new solutions arise on the operation and maintenance line.

[0038] The speech recognition module converts the user's voice input into user voice text through an algorithm based on deep learning. Front-end processing: This step includes noise reduction, echo cancellation and automatic gain control, with the aim of improving the quality of the speech signal; Feature extraction: Extract useful features from the audio signal, such as Mel-frequency cepstral coefficients (MFCC) or linear predictive coding (LPC); Acoustic model: Use statistical models (such as hidden Markov models HMMs or deep neural networks DNNs) to recognize speech fragments; Language model: Used to estimate the possibility of a given sentence, using an n-gram model or a neural network model; Decoder: Combine the acoustic model and language model to find the most likely text sequence.

[0039] The natural language processing module analyzes the user's voice text and extracts nouns A, verbs B and adjectives C contained in the text. The system uses natural language generation (NLG) to convert the internal representation into natural language text.

[0040] The natural language processing module is used in any document text F i Search for noun A, verb B, and adjective C in the result, and find that noun A appears a number of times. i , verb B appears B times i , adjective C appears C times i , when the number of characters between noun A and adjective C is less than G, it is recorded as an accurate noun phrase. The total number of characters is D i group of precise noun phrases, where i represents the i-th document text in the I-th document text in the offline database, and calculates F i Adaptation score E for user voice text i :

[0041] E i =K A *A i +K B *B i +K C *C i +K D *D i ;

[0042] Where G is the number of characters used to determine whether a noun and an adjective form a precise noun phrase, and K is A is the weight of the given adaptation score noun, K B is the existing adaptation scoring verb weight, K C is the existing adaptation score adjective weight, K D Calculate the weight of the precise noun phrase K for the precise noun phrase to be calculated. D :

[0043]

[0044] Substitute {i=1, 2, ..., I} one by one, calculate the adaptation score {F1, F2, ..., F I}, select the H data texts with the highest adaptation scores as the adaptation data texts adapted to the user's voice text, denoted as {J1, J2, ..., J H}, mark the parts of speech, analyze the degree of adaptation by part of speech, and accurately determine whether the data text can solve the user's problem.

[0045] The data retrieval module retrieves the adaptation material text {J1, J2, ..., J H} historical playback times, where the hth adapted data text J h The number of times played is L h , considering the number of playbacks to calculate the data text J h Priority answer rating W h :

[0046] W h =k E *E h +k L *L h ;

[0047] where k E For data text J h Score W for the priority answer h The adaptation score weight, k L For data text J h Score W for the priority answer h The weight of the number of times a message is played is taken into account. Considering that for any enterprise, historical commonly used information is often the most valuable and easiest to search for information, the number of times a message is played is taken into account here to calculate the priority answer score of the information text.

[0048] Substitute {h=1, 2, ..., H} one by one, and calculate the priority answer scores {W1, W2, ..., W H}, sort the H adaptation material texts into {M1, M2, ..., M H}.

[0049] The voice playback module converts the adaptation document text M1 into voice through NLG and plays it to the user. When the user voice inputs the answer document text and fails to solve the user's problem, the voice playback module voice plays the document text in descending order according to the priority answer score until the user issues a voice command that the answer has been obtained. Considering that the number of document texts for the system to answer the question is H, the documents are sorted here and the voice answer is provided to the user in descending order according to the priority answer score.

[0050] Embodiment 1: The information collection module collects company information from documents, manuals, and FAQ lists, and pre-processes the information to generate information text. The pre-processing process includes cleaning data, removing noise, and standardizing the format. The information collection module stores the processed information text in an offline database. Through data pre-processing, a series of processing is performed on the original data in the processes of data mining, data analysis, and machine learning to improve data quality, reduce noise, and facilitate the extraction of useful information, and convert the original data into useful data for subsequent data analysis. Data pre-processing includes: data deduplication, removing duplicate records in the data set, and achieving this by comparing unique identifiers or key fields in the records; missing value processing, filling missing values ​​in the data set, and processing using interpolation, mean, median, mode, and other methods; outlier processing, detecting and processing outliers in the data set, and the outliers are deleted or replaced with acceptable values; data standardization, standardizing the data format to a consistent format for easy processing and analysis; data conversion, data conversion is essentially converting the format of the data, and the main purpose is to facilitate data processing and analysis. The offline database is used as an offline knowledge base. It does not need to be connected to an external network during use. The construction of an offline database ensures information security and also ensures that key information can be permanently stored in a location that can be consulted. The speech recognition module converts the user's voice input into user speech text through an algorithm based on deep learning. Front-end processing: This step includes noise reduction, echo cancellation, and automatic gain control, the purpose of which is to improve the quality of the speech signal; feature extraction: extract useful features from the audio signal, such as Mel-frequency cepstral coefficients (MFCC) or linear predictive coding (LPC); acoustic model: use statistical models (such as hidden Markov models HMMs or deep neural networks DNNs) to recognize speech fragments; language model: used to estimate the possibility of a given sentence, using an n-gram model or a neural network model; decoder: combine the acoustic model and the language model to find the most likely text sequence. The natural language processing module analyzes the user's speech text and extracts nouns A, verbs B, and adjectives C contained in the text. The system uses natural language generation (NLG) to convert the internal representation into natural language text. The natural language processing module generates a natural language sequence in any data text F i Search for noun A, verb B and adjective C in the database, and find that noun A appears 10 times, verb B appears 12 times, and adjective C appears 16 times. When the number of characters between noun A and adjective C is less than 3, it is recorded as an accurate noun phrase. Statistics show that there are 6 groups of accurate noun phrases, where i represents the i-th data text in the I data texts in the offline database, where i = 10. Calculate F 10 Adaptation score E for user voice text 10 :

[0051] E 10=0.3*10+0.2*12+0.2*16+K D *6

[0052] Where G is the number of characters used to determine whether a noun and an adjective form a precise noun phrase, and K is A is the existing adaptation score noun weight, set to 0.3, K B is the existing adaptation score verb weight, set to 0.2, K C is the existing adaptation score adjective weight, set to 0.2, K D Calculate the weight of the precise noun phrase K for the precise noun phrase to be calculated. D :

[0053]

[0054] Calculate F 10 Adaptation score E for user voice text 10 =9.5.

[0055] Substitute {i=1, 2, ..., I} one by one and calculate the adaptation scores {F1, F2, ..., F 1000}, select the 10 data texts with the highest adaptation scores as the adaptation data texts adapted to the user's voice text, denoted as {J1, J2, ..., J 10}, mark the parts of speech, analyze the degree of adaptation by part of speech, and accurately determine whether the data text can solve the user's problem.

[0056] The data retrieval module retrieves the matching data text {J1, J2, ..., J 10}, where the number of times the first adapted document J1 is played is 56, and the adaptation score is 10. Considering the number of times it is played, the priority answer score W of document J1 is calculated. h :

[0057] W1 = 0.9 * 10 + 0.1 * 56 = 6.5;

[0058] where k E For data text J h Score W for the priority answer h The adaptation score weight is set to 0.9, k L For data text J h Score W for the priority answer hThe playback weight is set to 0.1. Considering that historical commonly used data is often the most valuable and easily searched data for any enterprise, the playback count is considered here to calculate the priority answer score of the data text. Substitute {h = 1, 2, ..., 10} one by one to calculate the priority answer scores of 10 suitable data texts {W1, W2, ..., W 10}, sort the H adaptation material texts into {M1, M2, ..., M 10}, the voice playback module converts the adaptation document text M1 into voice through NLG and plays it to the user. When the user voice inputs the answer document text but does not solve the user's question, the voice playback module voice plays the document text in descending order according to the priority answer score until the user issues a voice command that the answer has been obtained. Considering that the number of document texts for the system to answer the question is H, the documents are sorted here and the voice answers are provided to the user in descending order according to the priority answer score.

[0059] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A voice dialogue version intelligent question answering system based on a large model offline knowledge base, characterized by: The system includes: an information collection module, an offline database, a speech recognition module, a natural language processing module, a data retrieval module and a speech playback module; The information collection module collects company information, generates information text after processing and stores it in an offline database; Converting the user's voice input into user voice text through the voice recognition module; Analyze the user's voice text through the natural language processing module and select the adapted material text from the offline database; Retrieving the number of times the adaptation material text is played through the data retrieval module, analyzing the priority answer score in the adaptation material text in consideration of the number of times the text is played, and using the adaptation material text with the highest priority answer score as the answer material text; The voice playback module converts the answer material text into voice and then plays it to the user. When the user's voice input of the answer material text does not solve the user's problem, the voice playback module plays the material text in order of priority answer scores from high to low.

2. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 1 is characterized by: The information collection module manually collects company information from documents, manuals, and FAQ lists, and preprocesses the information to generate information text. The preprocessing process includes cleaning data, removing noise, and standardizing the format. The information collection module stores the processed information text in an offline database.

3. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 1 is characterized by: The offline database is used as an offline knowledge base and does not need to be connected to an external network during use.

4. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 1 is characterized by: The speech recognition module converts the user's speech input into user speech text through an algorithm based on deep learning.

5. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 4 is characterized by: The natural language processing module analyzes the user's voice text and extracts noun A, verb B and adjective C contained in the text.

6. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 5 is characterized by: The natural language processing module is used in any document text F i Search for noun A, verb B, and adjective C in the result, and find that noun A appears a number of times. i , verb B appears B times i , adjective C appears C times i , when the number of characters between noun A and adjective C is less than G, it is recorded as an accurate noun phrase. The total number of characters is D i group of precise noun phrases, where i represents the i-th document text in the I-th document text in the offline database, and calculates F i Adaptation score E for user voice text i : E i =K A *A i +K B *B i +K C *C i +K D *D i ; Where G is the number of characters used to determine whether a noun and an adjective form a precise noun phrase, and K is A is the weight of the given adaptation score noun, K B is the existing adaptation scoring verb weight, K C is the existing adaptation score adjective weight, K D Calculate the weight of the precise noun phrase K for the precise noun phrase to be calculated. D : Substitute {i=1, 2, ..., I} one by one, calculate the adaptation score {F1, F2, ..., F I }, select the H data texts with the highest adaptation scores as the adaptation data texts adapted to the user's voice text, denoted as {J1, J2, ..., J H }.

7. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 6 is characterized by: The data retrieval module retrieves the adaptation material text {J1, J2, ..., J H } historical playback times, where the hth adapted data text J h The number of times played is L h , considering the number of playbacks to calculate the data text J h Priority answer rating W h : W h =k E *E h +k L *L h ; where k E For data text J h Score W for the priority answer h The adaptation score weight, k L For data text J h Score W for the priority answer h The weight of the number of plays.

8. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 7 is characterized by: Substitute {h=1, 2, ..., H} one by one, and calculate the priority answer scores {W1, W2, ..., W H }, sort the H adaptation material texts into {M1, M2, ..., M H }.

9. The voice dialogue version intelligent question answering system based on a large model offline knowledge base according to claim 8 is characterized by: The voice playback module converts the adaptation material text M1 into voice through NLG and plays it to the user. When the user's voice input of the answer material text does not solve the user's problem, the voice playback module plays the material text in order from high to low priority answer scores until the user issues a voice command that the answer has been obtained.

Images