CN114519094A - Method and device for conversational recommendation based on random state and electronic equipment - Google Patents

Abstract

The application discloses a method and a device for recommending dialect based on a random state and electronic equipment, wherein the method comprises the following steps: performing text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information; querying historical dialogue data according to the first statement to obtain a second statement, wherein the occurrence time of the second statement is earlier than that of the first statement, and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is minimum; extracting the intention of the second sentence to obtain a first intention characteristic; generating a state tracking label of the first statement according to the first will characteristic; obtaining at least one first score from the first sentence, the state tracking label and the at least one conversational input scoring model; multiplying each first fraction of the at least one first fraction by a random function to obtain at least one second fraction; and recommending the dialect corresponding to the largest second score in the at least one second score to the answering device.

Description

Method, device and electronic device for speech recommendation based on random state

technical field

The present invention relates to the technical field of artificial intelligence, in particular to a method, device and electronic device for speech recommendation based on a random state.

Background technique

At present, the traditional vocabulary recommendation model is basically based on collaborative filtering or the TOP list obtained by the sorting model. Generally speaking, for most scenarios, the effect of the above vocabulary recommendation model can meet the needs of the scene. However, in the dialogue with people, the collaborative filtering or sorting model adopted by the traditional word recommendation method is to extract the high-frequency words corresponding to the high-frequency intentions and reply, which will cause the customer to ask questions every time. Even in the same conversation, as long as the intent of the conversation is similar, the response will always be some repeated high-frequency words, that is, several high-frequency words will appear repeatedly. As a result, a vicious circle is created, so that high-frequency words are always recommended, so that the frequency is further increased, and the subsequent access weight will also increase, making it easier to continue to access. It makes the whole dialogue process appear dull, lacking in fun and innovation, and the user experience is poor.

SUMMARY OF THE INVENTION

In order to solve the above problems in the prior art, the embodiments of the present application provide a random state-based speech recommendation method, device, and electronic device, which can improve the fun and creativity in a dialogue scene with a person, thereby improving user experience.

In a first aspect, the embodiments of the present application provide a random state-based speech recommendation method, including:

Perform text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information;

Query historical dialogue data according to the first sentence, and obtain a second sentence, wherein the historical dialogue data is used to record the dialogue data generated before the current moment in the dialogue event to which the first sentence belongs, and the occurrence time of the second sentence is earlier than that of the first sentence The occurrence time of , and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is the smallest;

Perform willingness extraction on the second sentence to obtain the first willingness feature;

generating a state tracking tag of the first sentence according to the first intention feature, wherein the state tracking tag is used to identify the user's own intention direction and demand intensity when speaking the voice information;

The first sentence, the state tracking label and the at least one discourse are input into the scoring model to obtain at least one first score, wherein the scoring model is used to score each discourse in the at least one discourse, and the at least one first score is the same as that of the at least one discourse. One-to-one correspondence with at least one vocabulary;

Multiplying each of the at least one first score by a random function to obtain at least one second score, wherein the at least one second score corresponds to the at least one first score one-to-one;

A phrase corresponding to the largest second score among the at least one second score is recommended to the replying device, so that the replying device generates a reply sentence according to the recommended phrase to reply to the user's current voice information.

In a second aspect, embodiments of the present application provide a random state-based speech recommendation device, including:

an analysis module, configured to perform text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information;

The query module is used for querying historical dialogue data according to the first sentence to obtain the second sentence, wherein the historical dialogue data is used to record the dialogue data generated before the current moment in the dialogue event to which the first sentence belongs, and the occurrence time of the second sentence. It is earlier than the occurrence time of the first statement, and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is the smallest;

The processing module is configured to perform willingness extraction on the second sentence to obtain a first willingness feature, and generate a state tracking label of the first sentence according to the first willingness feature, wherein the state tracking label is used to identify the user's own behavior when speaking the voice information. Willingness direction and demand intensity;

The scoring module is used to input the first sentence, the state tracking label and the at least one discourse into the scoring model to obtain at least one first score, wherein the scoring model is used to score each discourse in the at least one discourse, at least One first score is in one-to-one correspondence with at least one vocabulary, and each first score in the at least one first score is multiplied by a random function to obtain at least one second score, wherein the at least one second score is associated with the at least one The first scores are in one-to-one correspondence;

The recommendation module is configured to recommend the phrase corresponding to the largest second score in the at least one second score to the replying device, so that the replying device generates a reply sentence according to the recommended phrase to reply to the current voice information of the user.

In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor, the processor is connected to a memory, the memory is used for storing a computer program, and the processor is used for executing the computer program stored in the memory, so that the electronic device executes the program as described in the first method on the one hand.

In a fourth aspect, embodiments of the present application provide a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and the computer program causes a computer to execute the method according to the first aspect.

In a fifth aspect, embodiments of the present application provide a computer program product, where the computer program product includes a non-transitory computer-readable storage medium storing a computer program, and the computer is operable to cause the computer to execute the method of the first aspect.

Implementing the embodiments of the present application has the following beneficial effects:

In the embodiment of the present application, by analyzing and extracting the will of the sentence (the second sentence) preceding the sentence (the first sentence) spoken by the user at the current moment in the current dialogue, the second sentence is generated according to the will of the second sentence. The state tracking label for a statement. Then, the first sentence and the state tracking tag are used to score at least one adapted phrase, and the scoring result is perturbed by a random function, and then the phrase corresponding to the highest score in the perturbation result is selected for recommendation. Therefore, through the state tracking tag, each time a speech recommendation is made, in addition to the current dialogue, the reference parameters also carry the dialogue willingness of the previous moment, so that the recovery process has more initiative and can be By marking different state tracking tags to affect the direction of speech recommendation, and then introducing logical judgments into the model, the results of the model output have interpretable logic. At the same time, the random function is introduced to perturb the output results of the model with random probability, which can increase the fun and creativity in the question-and-answer process and improve the customer experience.

Description of drawings

In order to illustrate the technical solutions in the embodiments of the present application more clearly, the following briefly introduces the drawings that are used in the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

1 is a schematic diagram of the hardware structure of a random state-based speech recommendation device provided by an embodiment of the present application;

2 is a schematic flowchart of a random state-based speech recommendation method provided by an embodiment of the present application;

3 is a schematic flowchart of a method for performing text conversion processing on the current voice information of a user to obtain a first sentence corresponding to the voice information provided by an embodiment of the present application;

4 is a schematic flowchart of a method for determining a dialect category of speech information according to an acoustic feature provided by an embodiment of the present application;

FIG. 5 is a block diagram of functional modules of a random state-based speech recommendation device according to an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

The terms "first", "second", "third" and "fourth" in the description and claims of the present application and the drawings are used to distinguish different objects, rather than to describe a specific order . Furthermore, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes unlisted steps or units, or optionally also includes For other steps or units inherent to these processes, methods, products or devices.

Reference herein to "an embodiment" means that a particular feature, result, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor a separate or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

First, referring to FIG. 1 , FIG. 1 is a schematic diagram of a hardware structure of a random state-based term recommendation apparatus according to an embodiment of the present application. The random state-based term recommendation device 100 includes at least one processor 101 , a communication line 102 , a memory 103 and at least one communication interface 104 .

In this embodiment, the processor 101 may be a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more of them for controlling An integrated circuit implemented by the program program of this application.

Communication line 102, which may include a path, communicates information between the above-mentioned components.

The communication interface 104 may be any device such as a transceiver (such as an antenna, etc.) for communicating with other devices or communication networks, such as Ethernet, RAN, wireless local area networks (WLAN), and the like.

The memory 103 may be read-only memory (ROM) or other types of static storage devices that can store static information and instructions, random access memory (RAM) or other types of static storage devices that can store information and instructions It can also be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disk storage, optical disk storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or capable of carrying or storing desired program code in the form of instructions or data structures and capable of being executed by a computer Access any other medium without limitation.

In this embodiment, the memory 103 can exist independently, and is connected to the processor 101 through the communication line 102 . The memory 103 may also be integrated with the processor 101 . The memory 103 provided by the embodiments of the present application may generally be non-volatile. The memory 103 is used for storing computer-executed instructions for executing the solutions of the present application, and the execution is controlled by the processor 101 . The processor 101 is configured to execute the computer-executed instructions stored in the memory 103, thereby implementing the methods provided in the following embodiments of the present application.

In an optional implementation manner, the computer-executed instructions may also be referred to as application code, which is not specifically limited in this application.

In alternative embodiments, the processor 101 may include one or more CPUs, such as CPU0 and CPU1 in FIG. 1 .

In an optional embodiment, the random state-based vocabulary recommendation apparatus 100 may include multiple processors, such as the processor 101 and the processor 107 in FIG. 1 . Each of these processors can be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. A processor herein may refer to one or more devices, circuits, and/or processing cores for processing data (eg, computer program instructions).

In an optional implementation manner, if the vocabulary recommendation device 100 based on the random state is a server, for example, it may be an independent server, or may provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, Cloud servers for basic cloud computing services such as cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms. Then, the phrase recommendation apparatus 100 based on the random state may further include an output device 105 and an input device 106 . The output device 105 is in communication with the processor 101 and can display information in a variety of ways. For example, the output device 105 may be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, or a projector (projector) Wait. The input device 106 is in communication with the processor 101 and can receive user input in a variety of ways. For example, the input device 106 may be a mouse, a keyboard, a touch screen device or a sensing device, or the like.

The above-mentioned random state-based speech recommendation apparatus 100 may be a general-purpose device or a special-purpose device. The embodiments of the present application do not limit the type of the speech recommendation apparatus 100 based on the random state.

Secondly, it should be noted that the embodiments disclosed in this application can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

The basic technologies of artificial intelligence generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

Finally, the random state-based speech recommendation method in this application can be applied to scenarios such as e-commerce sales, offline physical sales, business promotion, AI phone calls, and social platform promotion. In this application, the AI telephone outbound call scenario is used as an example to illustrate the random state-based speech recommendation method. The random state-based speech recommendation method in other scenarios is similar to the implementation in the AI outbound telephone call scenario, and is not repeated here. narrative.

The following will describe the random state-based speech recommendation method disclosed in this application:

Referring to FIG. 2 , FIG. 2 is a schematic flowchart of a random state-based speech recommendation method provided by an embodiment of the present application. The random state-based discourse recommendation method includes the following steps:

201: Perform text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information.

Because of my country's vast territory and numerous terrains, there are many complex terrains that are insurmountable. The complex terrain makes it difficult for people to communicate through these terrains in the era of underdeveloped transportation, which makes our country form a variety of social, historical and cultural environments, and therefore, a variety of dialects have been formed.

In this embodiment, the voice information may be the voice information input by the user through the audio collection device. For example, for the AI phone call scenario, the voice information may be the voice information sent by the user to the AI through the audio collection device of the communication device. The voice to reply or ask a question. Therefore, considering the diversity and wide use of dialects in our country, it is necessary to take into account the situation of users communicating through dialects in practical applications.

Based on this, this embodiment provides a method for performing text conversion processing on the current voice information of the user to obtain the first sentence corresponding to the voice information. As shown in FIG. 3 , the method includes:

301 : Acquire acoustic features of speech information.

In this embodiment, an acoustic model can be pre-trained, for example, a multi-layer long short-term memory network, a multi-layer convolutional neural network, and the like. Thus, by inputting the to-be-recognized speech into the acoustic model, the acoustic features of the to-be-recognized speech are extracted. Exemplarily, the acoustic features may include a feature sequence of the to-be-recognized speech, a posterior probability distribution of phonemes in the to-be-recognized speech, and an acoustic vector of the to-be-recognized speech.

Specifically, the output of the low-level network in the acoustic model can be used as the feature sequence of the speech to be recognized, and the output of the high-level network can be used as the acoustic vector of the speech to be recognized. The posterior probability distribution of the phonemes in the speech to be recognized refers to the probability that each phoneme in the speech to be recognized is recognized as different phonemes.

302: Determine the dialect category of the speech information according to the acoustic feature.

In this embodiment, a method for determining the dialect category of speech information according to acoustic features is provided. As shown in FIG. 4 , the method includes:

401: Determine the energy distribution, prosody distribution, fundamental frequency and average speech power of the speech information according to the acoustic features.

Specifically, the posterior probability distribution of the phonemes in the acoustic features and the acoustic vector can be analyzed to obtain the energy distribution and prosody distribution of the speech information; the feature sequence in the acoustic features can be analyzed to obtain the average of the speech information. Speech power; and fundamental frequency can be obtained by analyzing the pronunciation characteristics of speech information.

402: Determine the dialect area corresponding to the speech information according to the fundamental frequency and the average speech power.

Specifically, although dialects are only used in a certain area, they also have a complete system. All dialects have a phonetic structure system, a lexical structure system and a grammatical structure system, which can meet the needs of social communication in the region. All in all, the dialects in the same dialect area often show the language characteristics of "similarities and differences, and similarities". The "same" among them is often reflected in the underlying characteristics of the sound, that is, the frequency of pronunciation, the speed of speech, etc. In other words, for each dialect, the fundamental frequency of its pronunciation and the average speech power of its pronunciation have certain regional commonality.

Therefore, in this embodiment, the fundamental frequency feature and the average voice power feature of the dialect of each dialect area can be extracted and stored in advance, and then the fundamental frequency and average voice power of the voice information can be compared with the pre-collected respective The method of feature matching between the fundamental frequency feature and the average speech power feature of the dialect area, for example, the method of calculating the similarity or the Euclidean distance, determines the dialect area corresponding to the speech information.

403: Encode the energy distribution and the rhythm distribution respectively to obtain an energy distribution vector and a rhythm distribution vector.

In this embodiment, the energy distribution can indicate the change of loudness in the speech information, and the prosody distribution can indicate the change of the pitch in the speech information. Specifically, the energy distribution vector can be obtained by acquiring the energy spectrum of the speech information and performing a graph embedding process on the energy spectrum. Similarly, the prosodic distribution vector can be obtained by acquiring the frequency distribution map of the speech information, and then performing a graph embedding process on the frequency distribution map.

404: Vertically splicing the energy distribution vector and the rhythm distribution vector to obtain a timbre distribution vector.

In an optional embodiment, the energy distribution vector and the rhythm distribution vector may be summed and then averaged, and the obtained average vector may be used as the timbre distribution vector.

405: Perform matching in the timbre library corresponding to the dialect area according to the timbre distribution vector, and determine the dialect category of the voice information according to the matching result.

Specifically, as described in step 402, the dialects in the same dialect area will show the language characteristics of "similarity with differences, and differences with similarities". For example, the flow of tones, tails, etc.

Based on this, in this embodiment, the timbre distribution features of each dialect in each dialect area can be extracted and stored in advance, and then the timbre distribution vector of the voice information is combined with the pre-collected timbre distribution of each dialect in the corresponding dialect area. The method of feature matching for timbre distribution features, for example, the method of calculating similarity or Euclidean distance, determines the dialect category corresponding to the voice information.

303 : Obtain an audio transposition formula corresponding to the dialect category, and convert the voice information into standard voice through the audio transposition formula.

In this embodiment, the audio transposition formula is used to identify the conversion feature between the corresponding dialect pronunciation and the Mandarin pronunciation. Specifically, through the dialect transposition formula, the dialect speech can be converted into the corresponding Mandarin speech, that is, the standard speech mentioned in this application.

In this embodiment, by collecting a large number of texts in different dialects but with the same content, the differences and laws between different dialects relative to Mandarin can be determined by training, for example: differences and laws in pronunciation, differences in tone and The rules, the corresponding relationship of proprietary words, etc., then form the audio transposition formula for converting different dialects into Mandarin.

304: Acquire the pinyin text of the voice information according to the standard voice.

In this implementation manner, the corresponding audio features can be obtained by performing feature extraction on standard speech, such as spectrum conversion, nonlinear spectrum conversion, and feature coefficient conversion. The audio feature may be a degree mapping under the auditory critical band scale corresponding to the standard speech, for example: the mapping of the standard speech in the Bark domain, and the standard speech in the Equivalent Rectangular Bandwidth (ERB) domain The mapping of standard speech can be used to quantify the audio features of standard speech through audio features.

Then, the audio feature can be matched in a preset neural network to obtain pinyin text matching the audio feature. Specifically, the pinyin text may be composed of at least one first pinyin vowel text, and the first pinyin vowel text refers to any one initial or final.

305: Perform matching in a preset vocabulary database according to the pinyin text to obtain a first sentence.

Specifically, after the pinyin text is obtained, matching can be performed in a preset vocabulary according to each first pinyin text in the at least one first pinyin text in the pinyin text, and a match with the at least one first pinyin text can be obtained. The text corresponds to at least one first character on a one-to-one basis. Then, according to the correspondence between the at least one first character and the at least one first pinyin text, the at least one first character is arranged according to the arrangement order of the at least one first pinyin text in the pinyin text, that is, The first sentence is available.

202: Query historical dialogue data according to the first sentence to obtain a second sentence.

In this embodiment, the historical dialogue data is used to record the dialogue data generated before the current time by the dialogue event to which the first sentence belongs, the occurrence time of the second sentence is earlier than the occurrence time of the first sentence, and the occurrence time of the second sentence is The absolute value of the difference between the time and the occurrence time of the first sentence is the smallest. In short, the second sentence is the last sentence that the user said before the current sentence.

Specifically, taking the scenario of an outgoing AI phone call as an example, two interrelated statement queues can be stored in the historical dialogue data, one of which is used to store user statements issued by the user, and the other pair is used to store the statements issued by AI. AI statement. At the same time, each user sentence in the user sentence queue and each AI sentence in the AI sentence queue contain a dialogue identifier and a dialogue occurrence time, and a user sentence and an AI sentence with the same dialogue identifier are a question-and-answer pair, that is, a dialogue Identifies the same user sentence as a reply to an AI sentence, or a dialog identifies the same AI sentence as a reply to a user sentence. In this way, the logic of the question and answer in the historical dialogue data can be ensured, and the sentences of the user and the AI can be saved separately for easy search.

Therefore, in this embodiment, by querying the user sentence queue, it is possible to determine the sentence with the dialogue occurrence time earlier than the occurrence time of the first sentence and the smallest absolute value of the difference between the occurrence time and the occurrence time of the first sentence as the first sentence. Two sentences.

In an optional embodiment, the user's historical dialogues may be stored according to the order of the dialogue occurrence time. In this case, only the sentence preceding the first sentence needs to be determined in the user's sentence queue to obtain the second sentence.

203: Perform intention extraction on the second sentence to obtain the first intention feature.

In this embodiment, semantic extraction may be performed on the second sentence, and then matching is performed in a preset willing database according to the obtained semantic vector, so as to determine the first willing feature of the second sentence. Specifically, the willingness features pre-stored in the willingness database are set according to the applicable field. For example, for the AI phone outbound call scenario, the willingness features can be preset including: interest, amount, repayment time, dividend ratio, etc. Related Willing Traits. Therefore, when performing matching, the desired feature most similar to the semantic vector of the second sentence can be matched as the first desired feature by calculating the similarity.

204: Generate a state tracking label of the first sentence according to the first intention feature.

In this embodiment, the status tracking tag is used to identify the user's own willingness direction and demand strength when speaking the current voice information, and can be used to determine the user's interest in the current outgoing call or the recommended service. direction, to guide the follow-up reply selection, which can improve the order rate of this outbound call.

Specifically, the user's demand intensity in the previous round of dialogue can be determined by obtaining the first demand score corresponding to the second sentence, and then the willingness of the first sentence can be extracted, and the first demand score can be scored according to the obtained second willingness feature. The update is performed to obtain a second demand score corresponding to the first sentence, so as to determine the current willingness intensity of the user. Finally, the second demand score and the first desire feature are combined to obtain a state tracking label that can reflect the user's desire direction and demand intensity at the same time, for example: "previous item_first desire feature_second demand score".

Exemplarily, when the first sentence is the first sentence, since it is the first sentence, the previous sentence (the second sentence) does not exist. Therefore, the second statement can be regarded as empty, and the corresponding willingness is also empty. At this time, the first demand score corresponding to the second statement can be analyzed according to the user's historical business data, according to the user's order rate in the historical business data. to determine, now assume 6. Therefore, the state tracking label of the first sentence is "previous term_0_6". When the first sentence is not the first sentence, the previous sentence must exist. At this time, the first demand score corresponding to the previous sentence (the second sentence) can be determined by analyzing the semantics and emotions of the previous sentence. If the willingness feature is "how much is the amount", and the corresponding second demand score is 8, the state tracking label of the first sentence is "previous item_how much is the amount_8".

At the same time, this embodiment provides a method for updating the first demand score according to the second intention feature to obtain a second demand score corresponding to the first sentence. Specifically, the second desired feature can be encoded to obtain the first character string, and at the same time, the call parameters of the call to which the first sentence belongs are obtained, and the missing values in the call parameters are complemented to obtain the target parameter. Thus, the target parameter, the first demand score and the first character string are spliced together to obtain the second character string, and then the second character string is input into the logistic regression model to obtain the second demand score.

In this embodiment, for the completion of the missing value, the missing value can be replaced with a preset substitute value, for example: a specific value such as 999, to obtain the target parameter; or, according to the data type of the missing value, the corresponding data type can be obtained The historical data set of , replace the missing value with the median of at least one data value included in the historical data set to obtain the target parameter; or replace the missing value with the mean of at least one data value included in the historical data set to obtain the target parameter.

In an optional embodiment, the second desire feature may also be directly used as the state tracking label of the first sentence, that is, the desire feature of the previous sentence may be used as the state tracking label of the current sentence. Exemplarily, when the first sentence is the first sentence, since it is the first sentence and the previous sentence does not exist, the state tracking label of the first sentence is "previous item_0", which means that the previous sentence is empty. When the first sentence is not the first sentence, the preceding sentence must exist. In this case, the intention feature of the preceding sentence can be determined as the state tracking label of the first sentence. Exemplarily, if the willingness feature of the previous sentence is "how much is the amount", the state tracking label of the first sentence is "the previous item_how much is the amount".

205: Input the first sentence, the state tracking label, and the at least one phrase into the scoring model, to obtain at least one first score corresponding to the at least one phrase one-to-one.

In this embodiment, the scoring model is used to score each speech in the at least one speech, and a rankingbert model may be selected as the scoring model. Specifically, the first sentence is analyzed and processed, and the obtained analysis results and status tracking labels are arranged in the following manner to obtain input data:

[cls, customer intent, sep, words, sep]+[whether there is status tracking]+[status tracking label]

Wherein, "customer intention" is the willingness feature of the first sentence, and the extraction method is similar to the extraction method of the willingness feature of the second sentence in step 203, and details are not repeated here. The discourse is the discourse type to which the first sentence belongs.

Therefore, compared with the traditional input sequence of the rankingbert model "cls, customer intent, sep, huashu, sep", the state tracking tag is added to the input sequence in this embodiment, so that logical judgment is introduced into the model, so that the model The output results have interpretable logic.

206 : Multiply each of the at least one first score by a random function to obtain at least one second score corresponding to the at least one first score one-to-one.

In this embodiment, a random function is provided, as shown in formula ①:

Among them, x represents the discourse type to which the first sentence belongs, C _x represents the number of conversations in the historical dialogue data that were successfully completed by the discourse type, S _x represents the number of conversations using the discourse type in the historical dialogue data, and a represents the weight Coefficient, random([1, n _x ]) is a random number related to n _x , and n _x is the number of available words matched by the willingness feature of the first sentence.

Therefore, formula ② is satisfied between the first score, the second score and the random function:

y _i =z _i ×p(x).........②

Wherein, _yi represents the i-th second fraction in the at least one second fraction, _zi represents the i-th first fraction in the at least one first fraction, and i is an integer greater than or equal to 1.

207 : Recommend the phrase corresponding to the largest second score in the at least one second score to the replying device as a recommended phrase, so that the replying device generates a reply sentence according to the recommended phrase to reply to the user's current voice information.

In this embodiment, the largest top 5 digits in the first score group may be selected, and then the largest five first scores are multiplied by a random function respectively, and the words corresponding to the largest score in the result are selected for recommendation. In this way, the number of computations can be reduced, and the recommendation efficiency can be improved.

To sum up, in the random state-based discourse recommendation method provided by the present invention, the intention of the sentence (the second sentence) preceding the sentence (the first sentence) spoken by the user at the current moment in the current dialogue is analyzed. Extraction, thereby generating the state tracking label of the first sentence according to the will of the second sentence. Then, the first sentence and the state tracking tag are used to score at least one adapted phrase, and the scoring result is perturbed by a random function, and then the phrase corresponding to the highest score in the perturbation result is selected for recommendation. Therefore, through the state tracking tag, each time a speech recommendation is made, in addition to the current dialogue, the reference parameters also carry the dialogue willingness of the previous moment, so that the recovery process has more initiative and can be By marking different state tracking tags to affect the direction of speech recommendation, and then introducing logical judgments into the model, the results of the model output have interpretable logic. At the same time, the random function is introduced to perturb the output results of the model with random probability, which can increase the fun and creativity in the question-and-answer process and improve the customer experience.

Referring to FIG. 5 , FIG. 5 is a block diagram of functional modules of a random state-based vocabulary recommendation apparatus provided by an embodiment of the present application. As shown in FIG. 5 , the random state-based vocabulary recommendation device 500 includes:

The analysis module 501 is used to perform text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information;

The query module 502 is configured to query historical dialogue data according to the first sentence to obtain a second sentence, wherein the historical dialogue data is used to record the dialogue data generated before the current moment in the dialogue event to which the first sentence belongs, and the occurrence of the second sentence. The time is earlier than the occurrence time of the first statement, and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is the smallest;

The processing module 503 is configured to perform willingness extraction on the second sentence, obtain a first willingness feature, and generate a state tracking label of the first sentence according to the first willingness feature, wherein the state tracking label is used to identify the user himself when he speaks the voice information direction of willingness and intensity of demand;

The scoring module 504 is configured to input the first sentence, the state tracking label and the at least one discourse into the scoring model to obtain at least one first score, wherein the scoring model is used to score each discourse in the at least one discourse, At least one first score is in one-to-one correspondence with at least one vocabulary, and each first score in the at least one first score is multiplied by a random function to obtain at least one second score, wherein the at least one second score is associated with at least one second score. A first score corresponds one-to-one;

The recommending module 505 is configured to recommend the phrase corresponding to the largest second score in the at least one second score to the replying device, so that the replying device generates a reply sentence according to the recommended phrase to reply to the current voice information of the user.

In the embodiment of the present invention, in terms of generating the state tracking label of the first sentence according to the first intention feature, the processing module 503 is specifically used for:

Obtain the first demand score corresponding to the second sentence;

Perform willingness extraction on the first sentence to obtain the second willingness feature;

updating the first demand score according to the second desire feature, to obtain a second demand score corresponding to the first sentence;

Combining the second demand score and the first willingness feature to obtain a state tracking label.

In the embodiment of the present invention, in terms of updating the first demand score according to the second desire feature, and obtaining the second demand score corresponding to the first sentence, the processing module 503 is specifically used for:

Encoding the second will feature to obtain the first character string;

Obtain the call parameters of the call to which the first sentence belongs, and complete the missing values in the call parameters to obtain the target parameters;

Splicing the target parameter, the first demand score and the first string to obtain the second string;

Input the second string into the logistic regression model to get the second demand score.

In the embodiment of the present invention, in the aspect of completing the missing values in the call parameters to obtain the target parameters, the processing module 503 is specifically used for:

Replace missing values with preset substitute values to get target parameters;

Or, according to the data type of the missing value, obtain the historical data set corresponding to the data type, replace the missing value with the median value of at least one data value included in the historical data set, and obtain the target parameter;

Or, replace missing values with the mean of at least one data value included in the historical dataset to obtain the target parameter.

In an embodiment of the present invention, the random function is shown in formula ③:

Among them, x represents the discourse type to which the first sentence belongs, C _x represents the number of conversations in the historical dialogue data that were successfully completed by the discourse type, S _x represents the number of conversations using the discourse type in the historical dialogue data, and a represents the weight Coefficient, random([1, n _x ]) is a random number related to n _x , and n _x is the number of available words matched by the willingness feature of the first sentence.

In the embodiment of the present invention, in terms of performing text conversion processing on the current voice information of the user to obtain the first sentence corresponding to the voice information, the analysis module 501 is specifically used for:

Obtain acoustic features of speech information;

Determine the dialect category of the speech information according to the acoustic features;

Obtain the audio transposition formula corresponding to the dialect category, and convert the voice information into standard speech through the audio transposition formula, wherein the audio transposition formula is used to identify the conversion feature between the corresponding dialect pronunciation and the Mandarin pronunciation;

Obtain the pinyin text of the voice information according to the standard voice;

The first sentence is obtained by matching in the preset vocabulary database according to the pinyin text.

In the embodiment of the present invention, in terms of determining the dialect category of the speech information according to the acoustic features, the analysis module 501 is specifically used for:

Determine the energy distribution, prosody distribution, fundamental frequency and average speech power of speech information according to acoustic features;

Determine the dialect area corresponding to the voice information according to the fundamental frequency and the average voice power;

Encoding the energy distribution and the rhythm distribution respectively to obtain the energy distribution vector and the rhythm distribution vector;

The energy distribution vector and the rhythm distribution vector are longitudinally spliced to obtain the timbre distribution vector;

According to the timbre distribution vector, matching is performed in the timbre library corresponding to the dialect area, and the dialect category of the voice information is determined according to the matching result.

Referring to FIG. 6 , FIG. 6 is a schematic structural diagram of an electronic device according to an embodiment of the present application. As shown in FIG. 6 , the electronic device 600 includes a transceiver 601 , a processor 602 and a memory 603 . They are connected by bus 604 . The memory 603 is used to store computer programs and data, and can transmit the data stored in the memory 603 to the processor 602 .

The processor 602 is used to read the computer program in the memory 603 to perform the following operations:

Perform text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information;

Query historical dialogue data according to the first sentence, and obtain a second sentence, wherein the historical dialogue data is used to record the dialogue data generated before the current moment in the dialogue event to which the first sentence belongs, and the occurrence time of the second sentence is earlier than that of the first sentence The occurrence time of , and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is the smallest;

Perform willingness extraction on the second sentence to obtain the first willingness feature;

generating a state tracking tag of the first sentence according to the first intention feature, wherein the state tracking tag is used to identify the user's own intention direction and demand intensity when speaking the voice information;

The first sentence, the state tracking label and the at least one discourse are input into the scoring model to obtain at least one first score, wherein the scoring model is used to score each discourse in the at least one discourse, and the at least one first score is the same as that of the at least one discourse. One-to-one correspondence with at least one vocabulary;

Multiplying each of the at least one first score by a random function to obtain at least one second score, wherein the at least one second score corresponds to the at least one first score one-to-one;

The speech corresponding to the largest second score among the at least one second score is recommended to the replying device as a recommended speech, so that the replying device generates a reply sentence according to the recommended speech to reply to the user's current voice information.

In the embodiment of the present invention, in terms of generating the state tracking label of the first sentence according to the first intention feature, the processor 602 is specifically configured to perform the following operations:

Obtain the first demand score corresponding to the second sentence;

Perform willingness extraction on the first sentence to obtain the second willingness feature;

updating the first demand score according to the second desire feature, to obtain a second demand score corresponding to the first sentence;

Combining the second demand score and the first willingness feature to obtain a state tracking label.

In the embodiment of the present invention, in terms of updating the first demand score according to the second desire feature to obtain the second demand score corresponding to the first sentence, the processor 602 is specifically configured to perform the following operations:

Encoding the second will feature to obtain the first character string;

Obtain the call parameters of the call to which the first sentence belongs, and complete the missing values in the call parameters to obtain the target parameters;

Splicing the target parameter, the first demand score and the first string to obtain the second string;

Input the second string into the logistic regression model to get the second demand score.

In the embodiment of the present invention, in terms of completing the missing values in the call parameters to obtain the target parameters, the processor 602 is specifically configured to perform the following operations:

Replace missing values with preset substitute values to get target parameters;

Or, according to the data type of the missing value, obtain the historical data set corresponding to the data type, replace the missing value with the median value of at least one data value included in the historical data set, and obtain the target parameter;

Or, replace missing values with the mean of at least one data value included in the historical dataset to obtain the target parameter.

In an embodiment of the present invention, the random function is shown in formula ④:

Among them, x represents the discourse type to which the first sentence belongs, C _x represents the number of conversations in the historical dialogue data that were successfully completed by the discourse type, S _x represents the number of conversations using the discourse type in the historical dialogue data, and a represents the weight Coefficient, random([1, n _x ]) is a random number related to n _x , and n _x is the number of available words matched by the willingness feature of the first sentence.

In an embodiment of the present invention, in terms of performing text conversion processing on the current voice information of the user to obtain the first sentence corresponding to the voice information, the processor 602 is specifically configured to perform the following operations:

Obtain acoustic features of speech information;

Determine the dialect category of the speech information according to the acoustic features;

Obtain the audio transposition formula corresponding to the dialect category, and convert the voice information into standard speech through the audio transposition formula, wherein the audio transposition formula is used to identify the conversion feature between the corresponding dialect pronunciation and the Mandarin pronunciation;

Obtain the pinyin text of the voice information according to the standard voice;

The first sentence is obtained by matching in the preset vocabulary database according to the pinyin text.

In the embodiment of the present invention, in terms of determining the dialect category of the speech information according to the acoustic features, the processor 602 is specifically configured to perform the following operations:

Determine the energy distribution, prosody distribution, fundamental frequency and average speech power of speech information according to acoustic features;

Determine the dialect area corresponding to the voice information according to the fundamental frequency and the average voice power;

Encoding the energy distribution and the rhythm distribution respectively to obtain the energy distribution vector and the rhythm distribution vector;

The energy distribution vector and the rhythm distribution vector are longitudinally spliced to obtain the timbre distribution vector;

According to the timbre distribution vector, matching is performed in the timbre library corresponding to the dialect area, and the dialect category of the voice information is determined according to the matching result.

It should be understood that the random state-based speech recommendation device in this application may include smart phones (such as Android mobile phones, iOS mobile phones, Windows Phone mobile phones, etc.), tablet computers, handheld computers, notebook computers, and MID (Mobile Internet Devices, Mobile Internet Devices, etc.) Abbreviation: MID), robots or wearable devices, etc. The above random state-based speech recommendation device is only an example, not exhaustive, including but not limited to the above random state-based speech recommendation device. In practical applications, the above-mentioned random state-based speech recommendation apparatus may further include: an intelligent vehicle-mounted terminal, a computer device, and the like.

From the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software combined with a hardware platform. Based on this understanding, all or part of the technical solutions of the present invention can be embodied in the form of software products, and the computer software products can be stored in storage media, such as ROM/RAM, magnetic disks, optical disks, etc. , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform the methods described in various embodiments of the present invention or some parts of the embodiments.

Therefore, embodiments of the present application also provide a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and the computer program is executed by a processor to implement any one of the above-mentioned method embodiments based on Some or all of the steps of the random state's discourse recommendation method. For example, the storage medium may include a hard disk, a floppy disk, an optical disk, a magnetic tape, a magnetic disk, a USB flash drive, a flash memory, and the like.

Embodiments of the present application also provide a computer program product, the computer program product comprising a non-transitory computer-readable storage medium storing a computer program, the computer program being operable to cause a computer to execute the method described in the foregoing method embodiments Part or all of the steps of any random state-based discourse recommendation method.

It should be noted that, for the sake of simple description, the foregoing method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described action sequence. Because in accordance with the present application, certain steps may be performed in other orders or concurrently. Secondly, those skilled in the art should also know that the implementation manners described in the specification are all optional implementation manners, and the actions and modules involved are not necessarily required by the present application.

In the above-mentioned embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

In the several embodiments provided in this application, it should be understood that the disclosed apparatus may be implemented in other manners. For example, the device implementations described above are only illustrative, for example, the division of the units is only a logical function division, and other divisions may be used in actual implementation, for example, multiple units or components may be combined or Integration into another system, or some features can be ignored, or not implemented. On the other hand, the shown or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be in electrical or other forms.

The units described as separate components may or may not be physically separated, and components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution in this implementation manner.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware, and can also be implemented in the form of software program modules.

The integrated unit, if implemented in the form of a software program module and sold or used as a stand-alone product, may be stored in a computer readable memory. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence, or the part that contributes to the prior art, or all or part of the technical solution, and the computer software product is stored in a memory, Several instructions are included to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and the program can be stored in a computer-readable memory, and the memory can include: a flash disk , Read-only memory (English: Read-Only Memory, referred to as: ROM), random access device (English: Random Access Memory, referred to as: RAM), magnetic disk or optical disk, etc.

The embodiments of the present application have been introduced in detail above, and specific examples are used to illustrate the principles and implementations of the present application. The descriptions of the above embodiments are only used to help understand the methods and core ideas of the present application; at the same time, for Persons of ordinary skill in the art, based on the idea of the present application, will have changes in the specific implementation manner and application scope. In summary, the contents of this specification should not be construed as limitations on the present application.

Claims (10)

1. A method for random state based conversational recommendation, the method comprising:

performing text conversion processing on current voice information of a user to obtain a first sentence corresponding to the voice information;

Querying historical dialogue data according to the first statement to obtain a second statement, wherein the historical dialogue data is used for recording dialogue data generated before the current moment by a dialogue event to which the first statement belongs, the occurrence time of the second statement is earlier than that of the first statement, and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is minimum;

extracting the will of the second sentence to obtain a first will characteristic;

generating a state tracking label of the first statement according to the first intention characteristic, wherein the state tracking label is used for identifying own intention direction and demand strength of the user when the user speaks the voice information;

inputting the first sentence, the state tracking tag and at least one dialect into a scoring model to obtain at least one first score, wherein the scoring model is used for scoring each of the at least one dialect, and the at least one first score is in one-to-one correspondence with the at least one dialect;

multiplying each first score in the at least one first score by a random function to obtain at least one second score, wherein the at least one second score is in one-to-one correspondence with the at least one first score;

Recommending the dialect corresponding to the maximum second score in the at least one second score to a reply device, so that the reply device generates a reply sentence according to the recommended dialect to reply to the current voice information of the user.

2. The method of claim 1, wherein generating the state tracking label for the first sentence according to the first will characteristics comprises:

acquiring a first demand score corresponding to the second statement;

extracting the will of the first sentence to obtain a second will characteristic;

updating the first demand score according to the second will characteristic to obtain a second demand score corresponding to the first statement;

and combining the second demand score and the first will characteristic to obtain the state tracking label.

3. The method of claim 2, wherein the updating the first demand score according to the second will characteristics to obtain a second demand score corresponding to the first sentence, comprises:

coding the second will characteristics to obtain a first character string;

acquiring call parameters of a call to which the first statement belongs, and completing missing values in the call parameters to obtain target parameters;

Splicing the target parameter, the first demand score and the first character string to obtain a second character string;

and inputting the second character string into a logistic regression model to obtain the second demand score.

4. The method of claim 3, wherein completing missing values in the call parameters to obtain target parameters comprises:

replacing the missing value with a preset replacing value to obtain the target parameter;

or acquiring a historical data set corresponding to the data type according to the data type of the missing value, and replacing the missing value with a median of at least one data value included in the historical data set to obtain the target parameter;

or replacing the missing value with a mean value of at least one data value included in the historical data set to obtain the target parameter.

5. The method of claim 1, wherein the random function is as follows:

wherein x represents a conversational type to which the first sentence belongs, C_xRepresenting the number of successful singletons by said type of dialogs in the historical dialog data, S_xRepresenting the number of dialogs in the historical dialog data using the type of dialogs, a representing the weight coefficient, random ([1, n) _x]) Is a and n_xAssociated random number, n_xThe number of available dialogs matched for the intended feature of the first sentence.

6. The method of claim 1, wherein performing text conversion processing on the current speech information of the user to obtain a first sentence corresponding to the speech information comprises:

acquiring acoustic features of the voice information;

determining dialect types of the voice information according to the acoustic features;

acquiring an audio transposition formula corresponding to the dialect category, and converting the voice information into standard voice through the audio transposition formula, wherein the audio transposition formula is used for identifying conversion characteristics between corresponding dialect pronunciation and mandarin pronunciation;

acquiring a pinyin text of the voice information according to the standard voice;

and matching the pinyin text in a preset vocabulary library to obtain the first sentence.

7. The method of claim 6, wherein determining the dialect class of the speech information based on the acoustic features comprises:

determining the energy distribution, rhythm distribution, fundamental frequency and average voice power of the voice information according to the acoustic features;

Determining a dialect area corresponding to the voice information according to the fundamental frequency and the average voice power;

respectively coding the energy distribution and the rhythm distribution to obtain an energy distribution vector and a rhythm distribution vector;

longitudinally splicing the energy distribution vector and the rhythm distribution vector to obtain a tone distribution vector;

and matching in a tone library corresponding to the dialect area according to the tone distribution vector, and determining the dialect category of the voice information according to a matching result.

8. A stochastic state based dialog recommendation device, the device comprising:

the analysis module is used for performing text conversion processing on the current voice information of the user to obtain a first sentence corresponding to the voice information;

the query module is used for querying historical dialogue data according to the first statement to obtain a second statement, wherein the historical dialogue data is used for recording dialogue data generated before the current moment of a dialogue event to which the first statement belongs, the occurrence time of the second statement is earlier than that of the first statement, and the absolute value of the difference between the occurrence time of the second statement and the occurrence time of the first statement is minimum;

The processing module is used for extracting the will of the second sentence to obtain a first will feature, and generating a state tracking label of the first sentence according to the first will feature, wherein the state tracking label is used for identifying the own direction of the will and the required strength of the user when the user speaks the voice information;

a scoring module, configured to input the first sentence, the state tracking tag, and at least one utterance into a scoring model to obtain at least one first score, where the scoring model is configured to score each of the at least one utterance, the at least one first score is in one-to-one correspondence with the at least one utterance, and multiply each of the at least one first score by a random function to obtain at least one second score, where the at least one second score is in one-to-one correspondence with the at least one first score;

and the recommending module is used for recommending the dialect corresponding to the maximum second score in the at least one second score to the answering equipment so as to enable the answering equipment to generate an answering sentence according to the recommended dialect and answer the current voice information of the user.

9. An electronic device comprising a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, the one or more programs including instructions for performing the steps in the method of any of claims 1-7.

10. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program which is executed by a processor to implement the method according to any one of claims 1-7.

Images