Collecting method, device and duplex voice character collecting system of specific format character

Abstract

This disclosure presents a method, apparatus, and full-duplex speech character collection system for acquiring characters of a specific format. The method for acquiring characters of a specific format includes: acquiring speech information; processing the speech information to identify at least one dialogue turn; determining the intent of the speech information based on the identified dialogue turn; parsing strings included in the speech information based on the identified dialogue turn; determining whether the parsed strings meet the character conditions required for a target format; and when the determination result is that the target format is not met, generating response information based on historical dialogue turns including the identified dialogue turns, the intent of the speech information, and the parsed strings. This allows for automatic generation of response information to quickly and accurately acquire characters of a specific format.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computers, and particularly relates to a method and device for collecting characters in a specific format and a duplex voice character collection system. BACKGROUND

[0002] With the increasing application of intelligent online conversation technologies such as intelligent customer service in the society, a voice dialogue system (robot) completes information collection through voice dialogue with a dialogue end and a user, which is an important intelligent voice conversation scenario. Unlike online task robots in the related art, voice interaction does not follow strict dialogue turn switching, but there is a large amount of duplex behavior of listening and speaking at the same time, and the collection of a slot often spans multiple dialogue turns or micro-turns. The dialogue turn refers to a message sent by a user as a turn, and there is a silent time interval between a turn and the next turn. The micro-turn refers to a short sentence obtained by cutting according to a slight pause in the dialogue turn in the voice dialogue. The slot is a specific information extracted by the robot through understanding the input content of the user in the online task robot. In the existing intelligent voice conversation scenario, common extraction information can include a telephone number, a bank card number, etc., and is commonly used in various business consultation and customer service conversations. However, in the duplex voice interaction process of the intelligent voice conversation, it is difficult to quickly and accurately collect characters in a specific format. SUMMARY

[0003] To solve the problems in the related art, the embodiments of the present disclosure provide a method and device for collecting characters in a specific format, a duplex voice character collection system, and corresponding electronic devices, readable storage media, and computer program products.

[0004] In a first aspect, the embodiments of the present disclosure provide a method for collecting characters in a specific format, comprising:

[0005] obtaining voice information;

[0006] processing the voice information to identify at least one dialogue turn;

[0007] determining an intent of the voice information based on the identified dialogue turn;

[0008] parsing a character string included in the voice information based on the identified dialogue turn;

[0009] generating a response information based on historical dialogue turns including the identified dialogue turn, the intent of the voice information, and the parsed character string, when the judgment result is that the target format is not met.

[0010] With reference to the first aspect, in a first implementation form of the first aspect, the method further includes: generating a confirmation response information based on the parsed string when the determination result is that the target format is met.

[0011] With reference to the first aspect, in a second implementation form of the first aspect, the character condition includes at least one of a character number, a character type, and a character arrangement manner of the target format.

[0012] With reference to the first aspect, in a third implementation form of the first aspect, the method further includes:

[0013] determining, based on the historical dialogue turn including the identified dialogue turn, the intent of the voice information, and the parsed string, a dialogue party action of a dialogue party that sends the voice information and a historical character collection result including the parsed string,

[0014] The determining whether the character condition required by the target format is met according to the parsed string includes:

[0015] determining whether the character condition required by the target format is met according to the historical character collection result including the parsed string.

[0016] With reference to any one of the first aspect, the first implementation form to the third implementation form of the first aspect, in a fourth implementation form of the first aspect, the intent of the voice information is an affirmative or negative intent, or the intent of the voice information is a global intent.

[0017] With reference to the fourth implementation form of the first aspect, in a fifth implementation form of the first aspect, the determining the intent of the voice information based on the identified dialogue turn includes:

[0018] determining, by using a keyword and a classification model trained and supervised by learning, the intent of the voice information based on the identified dialogue turn to be an affirmative or negative intent, or determining the intent of the voice information to be a global intent.

[0019] With reference to the fourth implementation form of the first aspect, in a sixth implementation form of the first aspect, the parsing the string included in the voice information based on the identified dialogue turn includes:

[0020] extracting a character fragment based on the identified dialogue turn;

[0021] performing a regularization processing on the character fragment to obtain a regularized string;

[0022] performing a normalization processing on the regularized string to obtain a normalized string.

[0023] With reference to the sixth implementation manner of the first aspect, in a seventh implementation manner of the first aspect, the regularizing the character segment to obtain the regularized string comprises:

[0024] detecting whether multiple character segments exist for the same dialogue turn and whether time intervals between the multiple character segments are less than a preset threshold value;

[0025] in response to detecting the multiple character segments and the time intervals between the multiple character segments being less than the preset threshold value, determining whether multiple candidate strings obtained based on regularizing the multiple character segments are repeated;

[0026] based on determining that the multiple candidate strings are repeated, retaining one candidate regularized string as the regularized string.

[0027] With reference to the third implementation manner of the first aspect, in an eighth implementation manner of the first aspect, the determining the dialogue party action of the dialogue party sending the speech information and the historical character collection result including the parsed string based on the historical dialogue turns including the identified dialogue turn, the intent of the speech information and the parsed string comprises:

[0028] determining the dialogue party action of the dialogue party sending the speech information and the historical character collection result including the parsed string based on the historical dialogue turns including the identified dialogue turn, the intent of the speech information and the parsed string by using a keyword to combine a classification model of supervised learning training.

[0029] With reference to the third implementation manner of the first aspect, in a ninth implementation manner of the first aspect, the method further comprises:

[0030] decoding the string to be collected based on the current character segment extracted from the current dialogue turn, the previous character segment extracted from the dialogue turn before the current dialogue turn and the determined dialogue party action of the dialogue party sending the speech information by using an algorithm based on beam search, wherein a coefficient less than 1 is assigned to a repeated candidate path of the current character segment and the previous character segment in the algorithm based on beam search, and / or a decoding result of a string corresponding to a previous dialogue turn of the current dialogue turn is revoked based on a determination result that the intent of the speech information is a negative intent.

[0031] With reference to the third implementation manner of the first aspect, in a tenth implementation manner of the first aspect, the generating the response information based on the historical dialogue turns including the identified dialogue turn, the intent of the speech information and the parsed string when the determination result is that the target format is not met comprises:

[0032] When the judgment result is that the target format is not met, a supervised model based on a deep neural network is used to generate response information based on historical dialogue turns including the recognized dialogue turns, the intent of the voice information, the parsed string, and the dialogue party action of the dialogue party sending the voice information.

[0033] With reference to any one of the first aspect, the first implementation manner of the first aspect, and the second implementation manner of the first aspect, in a eleventh implementation manner of the first aspect, the disclosure further includes:

[0034] Based on the response information or the confirmation response information, the response information or the confirmation response information of the natural language text is generated.

[0035] In a second aspect, the disclosure provides a specific format character collection device, including:

[0036] The voice information acquisition module is configured to acquire voice information.

[0037] The voice information processing module is configured to process the voice information to recognize at least one dialogue turn.

[0038] The intent recognition module is configured to determine the intent of the voice information based on the recognized dialogue turn.

[0039] The character parsing module is configured to parse a string included in the voice information based on the recognized dialogue turn.

[0040] The target format judgment module is configured to determine whether the target format required character condition is met according to the parsed string.

[0041] The response generation module is configured to generate response information based on historical dialogue turns including the recognized dialogue turns, the intent of the voice information, and the parsed string when the judgment result is that the target format is not met.

[0042] In a third aspect, the disclosure provides a duplex voice character collection system, including a duplex conversation manager and a specific format character collection device,

[0043] The duplex conversation manager includes:

[0044] The voice information acquisition module is configured to acquire voice information.

[0045] The voice information processing module is configured to process the voice information to recognize at least one dialogue turn.

[0046] The specific format character collection device includes:

[0047] An intent recognition module configured to determine an intent of the voice information based on the recognized dialogue turn;

[0048] A character parsing module configured to parse a character string included in the voice information based on the recognized dialogue turn;

[0049] An answer generation module configured to, when the determination result is that the target format is not met, generate answer information based on historical dialogue turns including the recognized dialogue turn, the intent of the voice information, and the parsed character string.

[0050] In a fourth aspect, an electronic device is provided in the embodiments of the present disclosure, including a memory and a processor; wherein

[0051] The memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method in any one of the first aspect, the first implementation to the eleventh implementation of the first aspect.

[0052] In a fifth aspect, a readable storage medium is provided in the embodiments of the present disclosure, and the readable storage medium has stored thereon computer instructions, which are executed by a processor to implement the method in any one of the first aspect, the first implementation to the eleventh implementation of the first aspect.

[0053] In a sixth aspect, a computer program product is provided in the embodiments of the present disclosure, and the computer program product includes computer instructions, which are executed by a processor to implement the method in any one of the first aspect, the first implementation to the eleventh implementation of the first aspect.

[0054] The technical solutions provided in the embodiments of the present disclosure can include the following beneficial effects:

[0055] According to the technical solutions provided in the embodiments of the present disclosure, by obtaining voice information, processing the voice information to recognize at least one dialogue turn, determining an intent of the voice information based on the recognized dialogue turn, parsing a character string included in the voice information based on the recognized dialogue turn, and when the determination result is that the target format is not met, generating answer information based on historical dialogue turns including the recognized dialogue turn, the intent of the voice information, and the parsed character string, the answer information can be automatically generated to quickly and accurately collect characters in a specific format.

[0056] According to the technical solutions provided in the embodiments of the present disclosure, by when the determination result is that the target format is met, generating confirmation answer information based on the parsed character string, the answer information can be automatically generated to quickly and accurately collect characters in a specific format.

[0057] According to the technical scheme provided by the embodiment of the present disclosure, by means of the character condition including at least one of the character number, character type and character arrangement mode of the target format, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0058] According to the technical scheme provided by the embodiment of the present disclosure, by means of the character condition including at least one of the character number, character type and character arrangement mode of the target format, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0059] According to the technical scheme provided by the embodiment of the present disclosure, by means of the character condition including at least one of the character number, character type and character arrangement mode of the target format, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0060] According to the technical scheme provided by the embodiment of the present disclosure, by means of the character condition including at least one of the character number, character type and character arrangement mode of the target format, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0061] According to the technical scheme provided by the embodiment of the present disclosure, by means of the character condition including at least one of the character number, character type and character arrangement mode of the target format, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0062] According to the technical scheme provided by the embodiment of the present disclosure, the regularized string is obtained by regularizing the character segment, including: detecting whether there are multiple character segments for the same dialogue turn and whether the time interval between the multiple character segments is less than a preset threshold; in response to detecting multiple character segments and a time interval between the multiple character segments being less than a preset threshold, determining whether multiple candidate strings obtained by regularizing the multiple character segments are repeated; based on the determination that the multiple candidate strings are repeated, retaining one candidate regularized string as the regularized string, which can automatically remove unnecessary repeated strings, thereby quickly and accurately collecting characters of a specific format.

[0063] According to the technical scheme provided by the embodiment of the present disclosure, the dialogue party action of the dialogue party issuing the voice information and the historical character collection result including the parsed string are determined based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string, including: using a keyword combined classification model trained by supervised learning to determine the dialogue party action of the dialogue party issuing the voice information and the historical character collection result including the parsed string based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string, which can accurately track the dialogue state, thereby automatically generating a response information to quickly and accurately collect characters of a specific format.

[0064] According to the technical scheme provided by the embodiment of the present disclosure, the string to be collected is decoded based on the current character segment extracted from the current dialogue turn, the previous character segment extracted from the dialogue turn before the current dialogue turn and the determined dialogue party action of the dialogue party issuing the voice information by using a beam search based algorithm, wherein the coefficient of the repeated candidate path of the current character segment and the previous character segment in the beam search based algorithm is less than 1, and / or the decoding result of the string corresponding to the previous dialogue turn of the current dialogue turn is revoked based on the determination result that the intent of the voice information is a negative intent, which can accurately track the dialogue state, thereby automatically generating a response information to quickly and accurately collect characters of a specific format.

[0065] According to the technical scheme provided by the embodiment of the present disclosure, when the judgment result is that the target format is not met, the response information is generated based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string, including: when the judgment result is that the target format is not met, the response information is generated based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information, the parsed string and the dialogue party action of the dialogue party sending the voice information by using the deep neural network-based supervised model, the dialogue state can be accurately tracked, and the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0066] According to the technical scheme provided by the embodiment of the present disclosure, the response information or the confirmation response information of the natural language text is generated based on the response information or the confirmation response information, the dialogue state can be accurately tracked, and the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0067] According to the technical scheme provided by the embodiment of the present disclosure, the voice information is obtained by the voice information obtaining module; the voice information is processed to recognize at least one dialogue turn by the voice information processing module; the intent of the voice information is determined based on the recognized dialogue turn by the intent recognition module; the string included in the voice information is parsed based on the recognized dialogue turn by the character parsing module; whether the target format required character condition is met is determined according to the parsed string by the target format judgment module; and the response information is generated based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string when the judgment result of the target format judgment module is that the target format is not met by the response generation module, the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0068] According to the technical scheme provided by the embodiment of the present disclosure, the technical scheme comprises a duplex conversation manager and a specific format character collection device, wherein the duplex conversation manager comprises a voice information acquisition module configured to acquire voice information, and a voice information processing module configured to process the voice information to identify at least one dialogue turn, and the specific format character collection device comprises an intention identification module configured to determine the intention of the voice information based on the identified dialogue turn, a character analysis module configured to analyze the character string included in the voice information based on the identified dialogue turn, a target format judgment module configured to determine whether the character string meets the character condition required by the target format, and a response generation module configured to generate response information based on the historical dialogue turns including the identified dialogue turn, the intention of the voice information and the analyzed character string when the determination result is that the target format is not met, so that the response information can be automatically generated in the duplex conversation scene to quickly and accurately collect characters in a specific format.

[0069] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0070] Other features, objects, and advantages of the present disclosure will become more apparent from the following detailed description of the non-limiting embodiments, taken in conjunction with the accompanying drawings. In the drawings:

[0071] Figure 1 A flow chart of a specific format character collection method according to an embodiment of the present disclosure is shown;

[0072] Figure 2 An example of applying the specific format character collection method shown in Figure 1 FIG. 1 to an intelligent customer service application scenario is shown;

[0073] Figure 3 Another example of applying the specific format character collection method shown in Figure 1 FIG. 1 to an intelligent customer service application scenario is shown;

[0074] Figure 4 A timing diagram of a conversation example of applying the specific format character collection method shown in Figure 1 FIG. 1 to an intelligent customer service application scenario is shown;

[0075] Figure 5 A structural block diagram of a specific format character collection device according to an embodiment of the present disclosure is shown;

[0076] Figure 6 A structural block diagram of a duplex voice character collection system according to an embodiment of the present disclosure is shown;

[0077] Figure 7 FIG. 1 shows a structural block diagram of an electronic device according to an embodiment of the disclosure;

[0078] Figure 8 FIG. 1 shows a structural block diagram of an electronic device according to an embodiment of the disclosure; DETAILED DESCRIPTION

[0079] Hereinafter, exemplary embodiments of the disclosure will be described in detail with reference to the accompanying drawings so as to be easily carried out by one of ordinary skill in the art. Also, portions unrelated to the description of the exemplary embodiments are omitted in the drawings for the sake of clarity.

[0080] In the disclosure, it should be understood that terms such as "include" or "have" are intended to indicate that there are labels, numbers, steps, actions, components, parts or combinations thereof disclosed in the specification, and do not exclude the possibility of existence or addition of one or more other labels, numbers, steps, actions, components, parts or combinations thereof.

[0081] In addition, it should be further noted that the embodiments in the disclosure and the labels in the embodiments can be combined with each other without conflict. The disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0082] In the process of duplex voice interaction in the intelligent voice conversation scene such as intelligent customer service, it is difficult to accurately collect character information of a specific format.

[0083] To solve the above problems, the disclosure proposes a specific format character collection method and device, a duplex voice character collection system, and corresponding electronic equipment, readable storage medium and computer program product. Through various conversation strategies such as acceptance and clarification, automatic and rapid and accurate identification and collection of target characters such as mobile phone number, bank card number, ID number, license plate number, and specific ID identification can be realized. That is, the character to be collected in the embodiment of the disclosure refers to a mobile phone number, or a bank card number, or an ID number, a license plate number, a specific ID identification, etc. which is expected to be collected as a whole from the conversation party providing voice information. In an embodiment of the disclosure, the character to be collected can include part or all of various characters such as numbers, letters, Chinese characters, symbols, etc.

[0084] According to the technical scheme provided by the embodiment of the present disclosure, the voice information is acquired, the voice information is processed to identify at least one dialogue turn, the intention of the voice information is determined based on the identified dialogue turn, the string included in the voice information is parsed based on the identified dialogue turn, and when the result of the judgment is that the target format is not met, the response information is generated based on the historical dialogue turns including the identified dialogue turn, the intention of the voice information, and the parsed string, so that the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0085] Figure 1 A flowchart of a specific format character collection method according to an embodiment of the present disclosure is shown. As shown in the specific format character collection method includes steps S101, S102, S103, S104, and S106. Figure 1 The specific format character collection method includes steps S101, S102, S103, S104, and S106.

[0086] In step S101, the voice information is acquired.

[0087] In step S102, the voice information is processed to identify at least one dialogue turn.

[0088] In step S103, the intention of the voice information is determined based on the identified dialogue turn.

[0089] In step S104, the string included in the voice information is parsed based on the identified dialogue turn.

[0090] In step S105, it is judged whether the character condition required by the target format is met according to the parsed string.

[0091] In step S106, when the result of the judgment is that the target format is not met, the response information is generated based on the historical dialogue turns including the identified dialogue turn, the intention of the voice information, and the parsed string.

[0092] In an embodiment of the present disclosure, the dialogue party sending the voice information (non-system response information) refers to a user having a voice call with the system performing character collection in the embodiment of the present disclosure.

[0093] In an embodiment of the present disclosure, the intention of the voice information (of the dialogue party) can refer to whether the voice information of the dialogue party is to provide content related to the character or unrelated content, for example, the intention of the voice information can be to input a character string, or to correct the information provided before, or to ask the system implementing the collection method of a specific format character a question, etc. Specifically, the intention of the voice information can be divided into two categories: (1) positive or negative intention, for example, if the recognized dialogue turn information is "no", the output result of determining the intention of the voice information is negative "False"; (2) global intention, which refers to an intention that is not related to the current character collection process, for example, the recognized dialogue turn information is "are you a robot?", "is it my own mobile phone number?", "what did you just say?" etc. The intention recognition can be regarded as a classification problem, which can be processed by a method of using keywords in combination with training a supervised learning classification model in related technologies, and the present disclosure will not be repeated here.

[0094] In an embodiment of the present disclosure, the intention of the voice information is a positive or negative intention, or the intention of the voice information is a global intention.

[0095] According to the technical scheme provided by the embodiment of the present disclosure, by determining that the intention of the voice information is a positive or negative intention, or the intention of the voice information is a global intention, the intention of the dialogue party can be more accurately understood, so that the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0096] In an embodiment of the present disclosure, Figure 1 The step S102 includes: training a supervised learning classification model in combination with keywords, determining that the intention of the voice information is a positive or negative intention based on the recognized dialogue turn, or determining that the intention of the voice information is a global intention.

[0097] According to the technical scheme provided by the embodiment of the present disclosure, by determining the intention of the voice information based on the recognized dialogue turn, including: training a supervised learning classification model in combination with keywords, determining that the intention of the voice information is a positive or negative intention based on the recognized dialogue turn, or determining that the intention of the voice information is a global intention, the intention of the dialogue party can be more accurately understood, so that the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0098] In an embodiment of the present disclosure, a string refers to a string of characters (which can only contain 1-bit character) that can be extracted from the recognized dialogue turn. The parsed string can be all or part of the target character that is the final target of character collection. In addition, the parsed string can be correct, that is, consistent with the string that the dialogue party should provide, or incorrect, that is, inconsistent with the string that the dialogue party should provide. It should be noted that the reason for the error can be that the dialogue party provides an incorrect string, or it can be due to an error in parsing the correct string provided by the dialogue party. Since both the dialogue party and the system executing the character collection method of a specific format can make mistakes in the process of character collection, the dialogue party may need to be confirmed for the string through various conversation strategies during the collection of the character. In the case where the judgment result does not meet the target format, the system executing the character collection method of a specific format can generate a response message based on the historical dialogue turns including the recognized dialogue turn, the intent of the voice information, and the parsed string, or can generate a response message to the dialogue party based on the historical dialogue turns including the recognized dialogue turn, the intent of the voice information, and the historical character collection result including the parsed string.

[0099] In an embodiment of the present disclosure, the character condition required by the specific format can be referred to as a system judgment rule for determining whether the collected character is legal. That is, the string meeting the character condition required by the target format can be considered to comply with the legality rule of the system, and the string not meeting the character condition required by the target format can be considered to not comply with the legality rule of the system. In an embodiment of the present disclosure, the response message converted into natural language and provided to the dialogue party can prompt the dialogue party to further provide voice information to facilitate further collection of the target character, or the response message indicates that the character collection is complete. The character collection complete includes two cases, one case is that the character collection is correct, confirmed by the dialogue party, and complies with the legality rule of the system itself executing the character collection method of a specific format, and the other case is that the character collection is incorrect, for example, does not comply with the legality rule of the system itself executing the character collection method of a specific format, at which time the character collection can be ended. For example, the legality rule can refer to whether the target character has a standard bit number, a character type, a character arrangement manner, etc. that complies with the rule, such as a standard bit number, a character type, a character arrangement manner of a telephone number, an ID number, or a license plate number. However, it should be understood that the legality rule of the embodiment of the present disclosure is not limited thereto, but can be various other rules.

[0100] In an embodiment of the present disclosure, the character collection method of a specific format can further include: when the judgment result meets the target format, generating a confirmation response message based on the parsed string.

[0101] In one embodiment of this disclosure, the determination result of satisfying the target format refers to the collected phone number, ID card number, or license plate number meeting the character conditions required by the target format, such as the number of digits, character type, and character arrangement of the collected characters. Alternatively, the determination result of satisfying the target format can also be understood as satisfying the legality rules.

[0102] According to the technical solution provided in this disclosure, when the judgment result is that the target format is met, confirmation response information is generated based on the parsed string, and response information can be automatically generated to quickly and accurately collect characters of a specific format.

[0103] In one embodiment of this disclosure, the character conditions include at least one of the number of characters, character type, and character arrangement of the target format. In one embodiment of this disclosure, the confirmation response information refers to the response information confirming that the parsed string meets the character conditions required by the target format, for example, "Just to confirm, is the phone number you mentioned 12300045678?".

[0104] In one embodiment of this disclosure, the number of characters in the target format can refer to a specified number of digits for telephone numbers, ID card numbers, license plate numbers, etc. The character type in the target format can refer to a specified character type for telephone numbers, ID card numbers, license plate numbers, etc., such as numbers, letters, Chinese characters, etc. The character arrangement in the target format can refer to a specified character arrangement for telephone numbers, ID card numbers, license plate numbers, etc., such as character arrangements stipulated by various laws.

[0105] According to the technical solution provided in the embodiments of this disclosure, by using the character conditions including at least one of the number of characters, character type, and character arrangement of the target format, response information can be automatically generated to quickly and accurately collect characters of a specific format.

[0106] The following is for reference Figures 2 to 4 Description of execution Figure 1 This is a schematic diagram illustrating a scenario for capturing characters in a specific format. It should be noted that, although... Figures 2 to 4 The scenario described is an intelligent customer service scenario, but the method for collecting specific format characters in this embodiment is not limited to intelligent customer service scenarios, but can be applied to any online dialogue scenario.

[0107] Figure 2 Showing will Figure 1 This diagram illustrates an example of how a method for collecting characters in a specific format is applied in an intelligent customer service application scenario. Figure 2 In the scenario shown, characters in specific formats can be collected, such as phone numbers, ID card numbers, bank account numbers, and license plate numbers.

[0108] like Figure 2 As shown, user 201 engages in a dialogue with intelligent customer service system 200 via mobile terminal 2011. Intelligent customer service system 200 includes a full-duplex session manager 210 and a character acquisition device 220 for specific formats. In one embodiment of this disclosure, the full-duplex session manager 210 and the character acquisition device 220 for specific formats can be implemented using hardware, software, or a combination of both. In one embodiment of this disclosure, the full-duplex session manager 210 and the character acquisition device 220 for specific formats can be constructed as a single unit, rather than two separate devices.

[0109] In one embodiment of this disclosure, a full-duplex conversation refers to a voice dialogue where both parties do not block each other, meaning both can listen and speak simultaneously. The full-duplex conversation manager 210 is an upstream device of the character acquisition device 220. The full-duplex conversation manager 210 can acquire voice information and perform voice processing on the voice information provided by the user 201 to identify dialogue turns. Throughout the full-duplex voice interaction process, the user 201 can speak at any time, and the continuous voice stream can be processed by the full-duplex conversation manager 210 using VAD (Voice Activity Detection) to break it down into dialogue turns and micro-dialogue turns. For example, it is temporarily assumed that the silence duration threshold between dialogue turns is 800ms, and the silence duration threshold between micro-dialogue turns is 400ms. Unlike related technologies that strictly control dialogue turn transitions, where the intelligent voice conversation system only responds after a dialogue turn ends, resulting in response times that severely impact user experience, the duplex conversation manager uses the streaming results of the current ASR (Automatic Speech Recognition) processing as input parameters to request the character acquisition device 220 with specific format characters in each dialogue turn or micro-dialogue turn. The character acquisition device 220 then returns the specific natural language response text. The duplex conversation manager 210 can convert the natural language response text into speech information through TTS (Text-to-speech) processing and send it to the mobile terminal 2011 so that the user 201 can obtain the response speech information. In one embodiment of this disclosure, the duplex conversation manager 210 can be implemented using methods from related technologies, which will not be elaborated upon here.

[0110] like Figure 2As shown, the specific format character collection device 220 includes an acquisition module 2201, an intent recognition module 2202, a global intent processing submodule 2203, a character parsing module 2204, a character regularization processing submodule 2205, a response generation module 2206, and a natural language generation module 2207. The acquisition module 2201 is configured to acquire the identified dialogue turn recognized by the duplex conversation manager 210. The intent recognition module 2202 determines the intent of the user 201 based on the identified dialogue turn acquired by the acquisition module 2201. The classification of the intent has been described in the above section and will not be repeated here. In an embodiment of the present disclosure, the intent recognition module 2202 can further include the global intent processing submodule 2203. In the case where the intent recognition module 2202 identifies a global intent, the identified global intent needs to be processed because the global intent can involve various aspects, for example, classification, judgment, etc. of the user intent for the identified global intent. The global intent processing can be implemented based on related technical means, which will not be repeated here. Figure 2 The specific format character collection device 220 shown is merely an example, and a person skilled in the art can implement the specific format character collection device 220 with more or fewer modules according to the teachings of the embodiments of the present disclosure.

[0111] In Figure 2 In the embodiment shown, the character to be collected can be a phone number. In Figure 2 In the specific format character collection device 220 shown, the character parsing module 2204 parses the character string included in the voice information based on the identified dialogue turn. The input of the character parsing module 2204 is mainly the information of the user dialogue turn, and the output is the result of the normalization of the character segment contained in the current dialogue turn. This module will first extract the character segment contained in the current dialogue turn, for example, the input dialogue turn is “then three zeros”, and the character segment “three zeros” is extracted. Then the extracted character segment is input to the character regularization processing submodule 2205, and after the regularization processing, the regularized character string can be obtained, and the normalized output “000” is output as the output of the entire character parsing module 2204 based on the rule method. In an embodiment of the present disclosure, the character parsing module 2204 can be implemented by using related technologies according to the teachings of the present disclosure. In an embodiment of the present disclosure, the character segment refers to a language segment that represents a character string by natural language, and its expression form can include a language segment that directly represents a character string, for example, “555”, and can also include a language segment that represents the composition relationship of characters in a character string, for example, “four fives”, “three twos and three threes”, etc.

[0112] In Figure 2In the collection device 220 of the specific format character shown, there is a case that needs special processing, for example, the user may have a very short pause in the dialogue turn when saying the character information, the duration is less than 400 ms, that is, within the same micro-dialogue turn. For example, the user says "000, oh, three zeros", here the user actually repeats the same string twice, therefore, it is necessary to determine whether to perform the de-duplication processing according to the regularization result, which is beneficial to reduce the decoding space in the next step and finally improve the efficiency of character collection.

[0113] Based on the description of the reference, Figure 2 Figure 1 The step S104 shown includes: extracting a character segment based on the identified dialogue turn; performing regularization processing on the character segment to obtain a regularized string; and performing normalization processing on the regularized string to obtain a normalized string.

[0114] According to the technical scheme provided by the embodiments of the present disclosure, the string included in the voice information is parsed based on the identified dialogue turn, including: extracting a character segment based on the identified dialogue turn; performing regularization processing on the character segment to obtain a regularized string; and performing normalization processing on the regularized string to obtain a normalized string, which can quickly and accurately collect characters of a specific format and improve the efficiency of character collection.

[0115] Further, the regularization processing on the character segment to obtain a regularized string includes: detecting whether there are multiple character segments for the same dialogue turn and whether the time interval between the multiple character segments is less than a preset threshold; in response to detecting multiple character segments and that the time interval between the multiple character segments is less than the preset threshold, determining whether multiple candidate strings obtained based on the regularization processing of the multiple character segments are repeated; and based on determining that the multiple candidate strings are repeated, retaining one candidate regularized string as the regularized string.

[0116] According to the technical scheme provided by the embodiments of the present disclosure, the regularization processing on the character segment to obtain a regularized string includes: detecting whether there are multiple character segments for the same dialogue turn and whether the time interval between the multiple character segments is less than a preset threshold; in response to detecting multiple character segments and that the time interval between the multiple character segments is less than the preset threshold, determining whether multiple candidate strings obtained based on the regularization processing of the multiple character segments are repeated; and based on determining that the multiple candidate strings are repeated, retaining one candidate regularized string as the regularized string, which can automatically remove unnecessary repeated strings and quickly and accurately collect characters of a specific format.

[0117] As​Figure 2 As shown, the response generation module 2206 generates the response information for the user 201 based on the dialogue turn obtained by the obtaining module 2201, the intent of the user 201 identified by the intent identification module 2202, and the historical character collection result including the parsed string. In an embodiment of the present disclosure, the currently collected string includes not only the string parsed by the current character parsing module 2204 based on the current dialogue turn, but also the historical string collected based on the previous dialogue turn. Generating the response information for the user 201 based on the historical character collection result including the parsed string can help to more accurately prompt the user 201. In an embodiment of the present disclosure, the “historical” data such as the historical string only means the data calculated, obtained, and identified in the operation step before the current operation step in the current target character collection process, and does not include the data in other target character collection processes other than the current target character collection process. For example, in the process of collecting the mobile phone number of the user 201 as the target character, the first three digits 123 of the mobile phone number have been collected before, and the string parsed by the current character parsing module 2204 is “000”, then the response generation module 2206 currently generates the response information for the user 201 based on the dialogue turn obtained by the obtaining module 2201, the intent of the user 201 identified by the intent identification module 2202, the first three digits “123” of the target character collected before, and the string “000” parsed by the character parsing module 2204.

[0118] In an embodiment of the present disclosure, in the character parsing module 2204 or after the character parsing module 2204, a target format judgment module (or a module for performing legality verification) for judging whether the parsed string meets the character condition required by the target format can also be included. The related description of the target format judgment module can be referred to the following description of the legality verification module 3203. Figure 3 The related description of the legality verification module 3203 described above.

[0119] In one embodiment of the disclosure, the response generation module 2206 can be considered as a module that makes character session decisions, the input of which can be the historical user intent understanding results (which can include the determined user actions) and historical response information (i.e., system actions that perform the collection method of a specific format character) used in the context in the current target character collection process, the information of the current dialogue turn and the historical character collection results up to the current turn, and the output is the response information (i.e., system actions). In one embodiment of the disclosure, when the flag format judgment module (or the module that performs legality verification) judges that the parsed string does not meet the target format, the response generation module 2206 can generate response information based on the historical dialogue turns including the recognized dialogue turn, the intent of the speech information and the parsed string. It should be understood that the basis for the response generation module 2206 to generate the response information can or can not include historical data other than the data obtained in the current dialogue turn. In one embodiment of the disclosure, a set of exemplary response information (system actions) is shown as shown in Table 1.

[0120] Table 1

[0121]

[0122] It should be understood that the names recorded in the response information (system action) column shown in Table 1 are the names of a type of response information (system action), not the specific response information to be sent to the dialogue party. In one embodiment of the disclosure, the response information "EMPTY" means that no specific response information is generated and no feedback is given to the user. However, the response information "EMPTY" can be considered as the default reply to the current dialogue turn of the user, but no longer prompts the user 201 to provide further information. For example, the current dialogue turn of the user 201 is the reply "Thank you" to the previous response information "Thank you for your confirmation" of the system, and this response information "EMPTY" indicates that the character collection is ended.

[0123] The response generation module 2206 can have various implementations. In one embodiment, a deep neural network based supervised model can be employed. Based on the annotated training data, the following features are defined: current dialog turn of the user; current global intent; length of character phrase contained in the current dialog turn; length of target characters decoded so far; current dialog party (user) action; last dialog party action; last response information (system action). Except for the first dialog turn feature which needs to be processed by word embedding and then output a sentence vector by LSTM (Long Short-Term Memory) or CNN (Convolutional Neural Network), the other features are one-hot processed and then get corresponding embedding. After that, the embedding of each feature is concatenated and sent to a 3-layer MLP network (Multi-Layer Perceptron), and finally a softmax layer is used to get the probability distribution over each response information (system action), and the response information (system action) with the maximum probability is selected as the final decision result. The above embodiment is only an example, and the present disclosure is not limited thereto. Those skilled in the art can implement the response generation module 2206 in various ways according to the teachings of the disclosed embodiments.

[0124] As shown in Figure 2 The natural language generation module 2207 generates natural language text response information based on the response information of the user 201 generated by the response generation module 2206 to prompt the user 201 to provide voice information or indicate the end of character collection. In one embodiment of the present disclosure, for the sake of simplicity, a template based approach can be employed to generate natural language text response information based on the response information of the user 201 generated by the response generation module 2206. The natural language text response information can be provided to the duplex conversation manager 210, converted into voice by TTS and finally returned to the user 201. In the natural language generation module 2207, some special processing can also be done. For example, using SSML (Speech Synthesis Markup Language), special pronunciation and playback pause processing of characters can be supported, for example, the "1" in the phone number will be read as "one". For another example, there will be a slight pause between "your mobile number is" and "123".

[0125] The natural language generation module 2207 generates natural language text response information based on the response information to user 201 generated by the response generation module 2206. For example, it generates the natural language text response "Yes, please tell me your mobile phone number" to prompt the user to read their mobile phone number. Another example is generating the natural language text response "Did you say '123'?" to prompt the user to clarify the character information from the previous dialogue round. Yet another example is not generating any natural language text response information to indicate the end of this character collection.

[0126] In one embodiment of this disclosure, the method for collecting specific format characters may further include: generating natural language text response information or confirmation response information based on the response information or the confirmation response information.

[0127] According to the technical solution provided in this disclosure, by generating natural language text response information or confirmation response information based on the response information or confirmation response information, the dialogue state can be accurately tracked, thereby automatically generating response information to quickly and accurately collect characters in a specific format.

[0128] The following is for reference Figure 3 This section further describes another example of how a method for collecting characters in a specific format is applied in an intelligent customer service application scenario. Figure 3 Showing will Figure 1 This diagram illustrates another example of how a method for collecting characters in a specific format is applied in an intelligent customer service application scenario. Figure 3 In the scenario shown, characters in specific formats can be collected, such as phone numbers, ID card numbers, bank account numbers, and license plate numbers.

[0129] Figure 3 The scene shown is Figure 2 The difference between the scenarios shown is that Figure 3 The intelligent customer service system 300 shown is different. Figure 2 The intelligent customer service system 200 shown is an example. Specifically, the specific format character acquisition device 320 in the intelligent customer service system 300 differs from the specific format character acquisition device 220 in the intelligent customer service system 200. Besides including... Figure 2 In addition to the acquisition module 2201, intent recognition module 2202, global intent processing submodule 2203, character parsing module 2204, character regularization processing submodule 2205, response generation module 2206, and natural language generation module 2207 in the specific format character acquisition device 220 shown, Figure 3 The specific format character acquisition device 320 shown also includes a dialogue party action determination module 3201, a character acquisition and decoding module 3202, and a legality verification module 3203.

[0130] existFigure 3 In the collection device 320 of the specific format character shown, after the character parsing module 2204 parses the character string included in the speech information based on the identified dialogue turn, the dialogue party action determination module 3201 determines the user action of the user 201 and the historical character collection result including the parsed character string based on the historical dialogue turns including the identified dialogue turn, the intent of the user 201 (of the speech information) determined by the intent recognition module 2202, and the parsed character string by the character parsing module 2204. The historical character collection result includes the character string collected based on the dialogue turns before the current dialogue turn. The dialogue party action determination module 3201 can be considered as a module performing character session state tracking, the input of which is all previous session contexts (historical dialogue turns and response information), and previous intent recognition results, parsed character strings, and the output of which is the user action of the user 201 and the character collection result up to the current dialogue turn. The dialogue party action determination module 3201 is similar in function to the intent recognition module 2202, both of which understand the intent or action of the user. The difference between them is that the dialogue party action determination module 3201 is context-sensitive, i.e., not only dependent on the current dialogue turn. For example, the intent recognition module 2202 judges that the intent of the user 201 in the current dialogue turn is a negative intent. Based on this judgment, if the previous turn response information (system action) is "clarification", the dialogue party action determination module 3201 judges that the user is negating the clarification content of the previous turn response information (system action), otherwise it may be negating the content (previous dialogue turn) said by the user 201 himself in the previous step. In an embodiment of the present disclosure, a set of user actions is exemplarily shown as shown in Table 2.

[0131] Table 2

[0132]

[0133]

[0134] Determining the user action can be considered as a classification problem, which can be implemented by using keywords to train a supervised learning classification model, and the specific details thereof will not be described here again according to the teaching of the embodiments of the present disclosure.

[0135] In an embodiment of the present disclosure, the collected character string refers to the character string determined by the dialogue party action determination module 3201, which can include the character string parsed by the character parsing module 2204, and can also include the user action of the user 201 to correct the previously parsed character string based on the current dialogue turn, the character collection result after the character string correction in the historical character collection result. For example, the character string determined based on the previous dialogue turn in the historical character collection result is "123", and the user action of the user 201 based on the current dialogue turn is "not, is 124", then the historical character collection result including the parsed character string will no longer include the previous character string "123", but will be corrected to include the character string "124".

[0136] As shown in Figure 3 The dialogue party action determination module 3201 provides the user action of the user 201 and the historical character collection result including the parsed character string to the response generation module 2206. The response generation module 2206 can generate response information for the user 201 based on the dialogue turn obtained by the obtaining module 2201, the intent of the user 201 identified by the intent identification module 2202, the historical character collection result including the parsed character string, and the historical user action of the user 201 and the historical character collection result including the parsed character string determined by the dialogue party action determination module 3201. The form of the response information can refer to the content described above.

[0137] In an embodiment of the present disclosure, the character collection method in the specific format can also include: determining the dialogue party action of the dialogue party sending the voice information and the historical character collection result including the parsed character string based on the historical dialogue turn including the identified dialogue turn, the intent of the voice information and the parsed character string. Wherein, step S105 includes: judging whether the target format required character condition is met according to the historical character collection result including the parsed character string.

[0138] According to the technical scheme provided by the embodiments of the present disclosure, by determining the dialogue party action of the dialogue party sending the voice information and the historical character collection result including the parsed character string based on the historical dialogue turn including the identified dialogue turn, the intent of the voice information and the parsed character string, and judging whether the target format required character condition is met according to the historical character collection result including the parsed character string, the dialogue state can be accurately tracked, and the response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0139] In an embodiment of the present disclosure, the determining the dialog party action of the dialog party sending the voice information and the historical character collection result including the parsed string based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string comprises: determining the dialog party action of the dialog party sending the voice information and the historical character collection result including the parsed string based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string by using a classification model of supervised learning trained by combining keywords.

[0140] According to the technical scheme provided by the embodiments of the present disclosure, the determining the dialog party action of the dialog party sending the voice information and the historical character collection result including the parsed string based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string comprises: determining the dialog party action of the dialog party sending the voice information and the historical character collection result including the parsed string based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string by using a classification model of supervised learning trained by combining keywords, which can accurately track the dialog state, thereby automatically generating the response information to quickly and accurately collect characters in a specific format.

[0141] In an embodiment of the present disclosure, the generating the response information based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string when the judgment result is that the target format is not met comprises: generating the response information based on the historical dialog turn including the identified dialog turn, the intent of the voice information, the parsed string and the dialog party action of the dialog party sending the voice information by using a supervised model based on a deep neural network when the judgment result is that the target format is not met.

[0142] According to the technical scheme provided by the embodiments of the present disclosure, the generating the response information based on the historical dialog turn including the identified dialog turn, the intent of the voice information and the parsed string when the judgment result is that the target format is not met comprises: generating the response information based on the historical dialog turn including the identified dialog turn, the intent of the voice information, the parsed string and the dialog party action of the dialog party sending the voice information by using a supervised model based on a deep neural network when the judgment result is that the target format is not met, which can accurately track the dialog state, thereby automatically generating the response information to quickly and accurately collect characters in a specific format.

[0143] As Figure 3As shown, the character collection device 320 of the special format character further includes a character collection and decoding module 3202. The character collection and decoding module 3202 can employ a beam search based algorithm. The beam search based algorithm employed in the embodiments of the present disclosure differs from the beam search algorithm known in the related art in that the scoring function is special, different from the scoring manner in the language model of the beam search algorithm in the related art, and a rule based method is employed here, such as penalizing the repeated candidate path for the current character fragment and the previous character fragment. Specifically, when a repeated character string appears, the score of the path is multiplied by a coefficient less than 1 (for example, 0.9 or other values) as a penalty term. This scoring strategy can very well solve the situation where the user repeatedly reports a certain character fragment. The beam search based algorithm employed in the embodiments of the present disclosure can also differ from the beam search algorithm known in the related art in that it supports modifying the reported character, when the user intent (user action) of the user 201 is identified as modifying (correcting) the reported character of the previous step, such as the user saying "not right", the decoding result of the previous step needs to be revoked based on the beam search based algorithm, and the decoding result of the earlier step is rolled back. Therefore, the decoding process based on the beam search based algorithm according to the embodiments of the present disclosure is not a single forward decoding process. The other processes performed by the beam search based algorithm in the embodiments of the present disclosure are the same as the beam search algorithm known in the related art, which will not be described here.

[0144] In an embodiment of the present disclosure, the special format character collection method can further include: decoding the character string to be collected based on the current character fragment extracted from the current dialogue turn, the previous character fragment extracted from the dialogue turn before the current dialogue turn, and the determined dialogue party action of the dialogue party issuing the voice information, by using a beam search based algorithm, wherein a coefficient less than 1 is assigned to the candidate path that repeats the current character fragment and the previous character fragment in the beam search based algorithm, and / or the decoding result of the character string corresponding to the previous dialogue turn of the current dialogue turn is revoked based on the determination result that the intent of the voice information is a negative intent.

[0145] According to the technical scheme provided by the embodiment of the present disclosure, the string to be collected is decoded by using the beam search based algorithm, based on the current character segment extracted from the current dialogue turn, the previous character segment extracted from the dialogue turn before the current dialogue turn, and the dialogue party action of the dialogue party sending the voice information. In the beam search based algorithm, the coefficient of the repeated candidate path of the current character segment and the previous character segment is less than 1, and / or the decoding result of the string corresponding to the previous dialogue turn of the current dialogue turn is revoked based on the determination result that the intention of the voice information is a negative intention. The dialogue state can be accurately tracked, so that the response information is automatically generated to quickly and accurately collect the character of a specific format.

[0146] As shown in Figure 4 The legality verification module 3203 determines the legality (whether it conforms to the target format) of the historical character collection result based on the preset character legality determination rule (character condition required by the target format). The input of the legality verification module 3203 can be the historical character collection result, for example, the optimal candidate result output by the character collection decoding module 3202, and its output is whether the candidate is legal. The output result of the legality verification module 3203 will directly affect the dialogue process of the character collection, for example, when the character bit number, character type, character arrangement mode and the like of the target character such as mobile phone number or ID number of the current collection are detected to be incorrect based on the character legality determination rule, the user 201 can be required to re-input the target character. Those skilled in the art can understand that various character legality determination rules can be used according to the teaching of the embodiment of the present disclosure, and the specific details are not described here.

[0147] In an embodiment of the present disclosure, the specific format character collection device 320 can not include the specific format character collection device 3202. In this case, the legality verification module 3203 is connected to the dialogue party action determination module 3201, and verifies the legality of the historical character collection result provided by the dialogue party action determination module 3201, including the parsed string.

[0148] In an embodiment of the present disclosure, the method for collecting characters in a specific format can further include determining the legality of the historical character collection result based on a preset character legality determination rule. In an embodiment of the present disclosure, based on the determination result that the historical character collection result is legal, the identified dialogue turn, the determined historical dialogue party action of the dialogue party, the historical response information of the dialogue party, and the historical character collection result including the parsed string, the response information of the dialogue party is generated to prompt the dialogue party to provide voice information or indicate the end of character collection. In this way, the response information can be automatically generated to quickly and accurately collect characters in a specific format, and the collected characters are ensured to be compliant.

[0149] In an embodiment of the present disclosure, the method for collecting characters in a specific format can further include, based on the determination result that the historical character collection result is not legal, generating preset response information of the dialogue party to indicate the end of character collection. In this way, the non-compliant string can be automatically excluded, so that the response information can be automatically generated to quickly and accurately collect characters in a specific format, and the collected characters are ensured to be compliant.

[0150] The following refers to Figure 4 a timing diagram of a conversation example of the intelligent customer service application scenario. Figure 1 The timing diagram shows that Figure 4 the method for collecting characters in a specific format shown in FIG. 6 is applied to the timing diagram of the conversation example of the intelligent customer service application scenario. It should be understood that Figure 2 the timing diagram shown in FIG. 6 shows that Figure 3 or Figure 4 The intelligent customer service system shown in FIG. 6 receives information provided by the user 201 through the duplex conversation manager 210 and provides system response information to the user 201. In Figure 4 In the first dialogue turn, two micro dialogue turns are included. The intelligent customer service system in the embodiment of the present disclosure does not wait until the end of the dialogue turn to respond as in the common non-duplex system, but returns an acceptance speech at the first micro dialogue turn, effectively shortening the response time and ultimately optimizing the user experience.

[0151] As Figure 5As shown, at the earliest time, the duplex conversation manager 210 sends the response information "Could you please tell me your phone number" to the user 201, and the information provided by the user 201 in the first dialogue turn is composed of two micro dialogue turns "123" and "then three zeros". The intelligent customer service system provides the response information "uh-huh" for taking over between the two micro dialogue turns, and provides the response information "three zeros" for implicit clarification after the first dialogue turn. Next, the user 201 provides the information "33869" in the second dialogue turn, and the intelligent customer service system provides the response information "33869" for implicit clarification after the second dialogue turn.

[0152] In order to more clearly illustrate the processing flow of the character collection system, through the following examples in Table 3, it can be seen that each step of the character collection system input and output. The first sentence of the character collection system is omitted in the following table because this is a fixed dialogue that does not depend on any dialogue strategy, the first column is the name of each module, and the subsequent ith column is the output of the corresponding submodule at the ith micro dialogue turn.

[0153] Table 3

[0154]

[0155] As shown in Table 3, at each micro dialogue turn, each module in the character collection system works in turn to process the user's voice information, thereby collecting character strings. Although Table 3 only shows 3 micro dialogue turns, the present disclosure is not limited thereto, and those skilled in the art can understand that through more micro dialogue turns, the character collection system can collect target characters such as the user's phone number.

[0156] The following describes how to complete the collection of target characters in one embodiment.

[0157] In one embodiment of the present disclosure, the collection method of the specific format character further includes: based on the generated response information of the dialogue party clarifying the target character, generating natural language text response information to prompt the dialogue party to provide voice information of feedback for the target character. Wherein, the determination of the dialogue party action of the dialogue party sending the voice information based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string, and the historical character collection result including the parsed string includes: determining which one of confirming the target character, correcting the target character and replying irrelevant to the target character is the dialogue party action of the dialogue party sending the voice information based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string. Wherein, when the judgment result is not satisfied with the target format, the response information is generated based on the historical dialogue turns including the recognized dialogue turns, the intent of the voice information and the parsed string, including: when the judgment result is not satisfied with the target format, in response to the dialogue party action of the dialogue party sending the voice information is to confirm the target character, based on the recognized dialogue turns, the determined historical dialogue party action of the dialogue party, the historical response information of the dialogue party and the historical character collection result including the parsed string, the target character is determined as the result of character collection, and the response information of the dialogue party is generated. This can accurately track the dialogue state, thereby automatically generating response information to quickly and accurately collect characters of a specific format.

[0158] The embodiments of the present disclosure can be used for collection of various target characters. In duplex voice interaction, various conversation strategies such as acceptance and clarification can be used to automatically and quickly identify, normalize and collect target characters.

[0159] The following refers to Figure 5 An example of a device for collecting information of a specific format character is described. Figure 5 A structural block diagram of a device for collecting a specific format character according to an embodiment of the present disclosure is shown. As shown in Figure 5 The device for collecting a specific format character 500 includes a voice information acquisition module 501, a voice information processing module 502, an intent recognition module 503, a character parsing module 504, a target format judgment module 505 and a response generation module 506.

[0160] The voice information acquisition module 501 is configured to acquire voice information.

[0161] The voice information processing module 502 is configured to process the voice information to identify at least one dialogue turn.

[0162] The intent recognition module 503 is configured to determine the intent of the voice information based on the recognized dialogue turn.

[0163] The character parsing module 504 is configured to parse a character string included in the voice information based on the identified dialogue turn.

[0164] The target format judgment module 505 is configured to determine whether a character condition required by a target format is met according to the parsed character string.

[0165] The response generation module 506 is configured to generate response information based on historical dialogue turns including the identified dialogue turn, the intent of the voice information, and the parsed character string when the determination result of the target format judgment module 505 is that the target format is not met.

[0166] According to the technical scheme provided by the embodiment of the present disclosure, the voice information acquisition module is configured to acquire voice information; the voice information processing module is configured to process the voice information to identify at least one dialogue turn; the intent identification module is configured to determine the intent of the voice information based on the identified dialogue turn; the character parsing module is configured to parse a character string included in the voice information based on the identified dialogue turn; the target format judgment module is configured to determine whether a character condition required by a target format is met according to the parsed character string; and the response generation module is configured to generate response information based on historical dialogue turns including the identified dialogue turn, the intent of the voice information, and the parsed character string when the determination result of the target format judgment module is that the target format is not met. The response information can be automatically generated to quickly and accurately collect characters in a specific format.

[0167] In one embodiment of the present disclosure, the response generation module 506 is configured to generate confirmation response information based on the parsed character string when the determination result of the target format judgment module 505 is that the target format is met.

[0168] In one embodiment of the present disclosure, the character condition includes at least one of the number of characters, the type of characters, and the arrangement of characters of the target format.

[0169] In one embodiment of the present disclosure, the character collection device 500 of the specific format can further include a dialogue party action determination module configured to determine a dialogue party action of a dialogue party sending the voice information and a historical character collection result including the parsed character string based on historical dialogue turns including the identified dialogue turn, the intent of the voice information, and the parsed character string, wherein the target format judgment module 505 is configured to determine whether a character condition required by a target format is met according to the historical character collection result including the parsed character string.

[0170] In an embodiment of the present disclosure, the intent of the voice information is an affirmative or negative intent, or the intent of the voice information is a global intent.

[0171] In an embodiment of the present disclosure, the intent recognition module 503 is configured to:

[0172] The classification model trained by keyword combination and supervised learning is used to determine the intent of the voice information as an affirmative or negative intent based on the identified dialogue turn, or determine the intent of the voice information as a global intent.

[0173] In an embodiment of the present disclosure, the character parsing module 504 includes:

[0174] A character fragment extraction sub-module configured to extract character fragments based on the identified dialogue turn;

[0175] A regularization sub-module configured to perform regularization processing on the character fragments to obtain a regularized string;

[0176] A normalization sub-module configured to perform normalization processing on the regularized string to obtain a normalized string.

[0177] In an embodiment of the present disclosure, the regularization sub-module includes:

[0178] A detection component configured to detect whether multiple character fragments exist for the same dialogue turn and whether the time interval between the multiple character fragments is less than a preset threshold;

[0179] A repetition determination component configured to, in response to detecting multiple character fragments and that the time interval between the multiple character fragments is less than the preset threshold, determine whether multiple candidate strings obtained based on the multiple character fragments for regularization processing are repeated;

[0180] A retention component configured to, based on a determination that the multiple candidate strings are repeated, retain one candidate regularized string as the regularized string.

[0181] In an embodiment of the present disclosure, the dialogue party action determination module is further configured to:

[0182] The classification model trained by keyword combination and supervised learning is used to determine the dialogue party action of the dialogue party issuing the voice information and the historical character collection result including the parsed string based on the historical dialogue turn including the identified dialogue turn, the intent of the voice information, and the parsed string.

[0183] In an embodiment of the present disclosure, the specific format character collection device 500 can further include:

[0184] The character acquisition and decoding module is configured to use a beam search-based algorithm to decode the string to be acquired based on the current character fragment extracted from the current dialogue turn, the prior character fragment extracted from the dialogue turn before the current dialogue turn, and the determined dialogue party action of the party that issued the voice information. In the beam search-based algorithm, candidate paths where the current character fragment and the prior character fragment are repeated are assigned a coefficient less than 1, and / or the decoding result of the string corresponding to the previous dialogue turn is revoked based on the determination result that the intent of the voice information is a negative intent.

[0185] In one embodiment of this disclosure, the response generation module 506 is further configured to:

[0186] If the judgment result is that the target format is not met, a response information is generated using a supervised model based on a deep neural network, based on the historical dialogue rounds identified in the dialogue rounds, the intent of the voice information, the parsed string, and the dialogue actions of the party that sent the voice information.

[0187] In one embodiment of this disclosure, the acquisition device 500 for specific format characters may further include a natural language generation module, which is further configured to generate natural language text response information or confirmation response information based on the response information or the confirmation response information.

[0188] In one embodiment of this disclosure, the natural language generation module generates natural language text response information based on the response information generated by the response generation module to clarify the target character of the dialogue party, prompting the dialogue party to provide voice information for feedback on the clarified target character. In one embodiment of this disclosure, the dialogue party action determination module is further configured to: determine, based on the historical dialogue rounds including the identified dialogue rounds, the intent of the voice information, and the parsed string, whether the dialogue party sending the voice information's action is to confirm the target character, correct the target character, or provide a response unrelated to the target character; wherein, the response generation module is further configured to: when the determination result is that the target format is not met, in response to the dialogue party sending the voice information confirming the target character, determine the target character as the result of character acquisition based on the identified dialogue rounds, the determined historical dialogue party actions, the historical response information of the dialogue party, and the historical character acquisition results including the parsed string, and generate response information for the dialogue party.

[0189] Those skilled in the art will understand that, with reference to Figures 1 to 4 The described technical solution can be compared with the reference. Figures 1 to 4 The described embodiments are combined to provide reference. Figures 1 to 4Technical effects achieved by the described embodiments. Details can be referred to the above description according to Figure 6 which are not repeated here.

[0190] Examples of a duplex voice character collection system are described below. Figure 6 Examples of a duplex voice character collection system are described below. Figure 6 A structural block diagram of a duplex voice character collection system according to an embodiment of the present disclosure is shown. As shown in Figure 6 , the duplex voice character collection system 600 includes a duplex conversation manager 601 and a specific format character collection device 602. The duplex conversation manager 601 includes a voice information acquisition module 6011 and a voice information processing module. The specific format character collection device 602 includes an intention recognition module 6021, a character analysis module 6022, a target format judgment module 6023, and a response generation module 6024. Figure 5 The specific format character collection device 602 shown in Figure 6 is slightly different from the specific format character collection device 500 shown in . It should be understood that those skilled in the art can construct duplex conversation managers and specific format character collection devices that are the same as or different from the foregoing embodiments according to the teachings of the embodiments of the present disclosure.

[0191] In the duplex conversation manager 601, the voice information acquisition module 6011 is configured to acquire voice information. The voice information processing module 6012 is configured to process the voice information to identify at least one dialogue turn.

[0192] In the specific format character collection device 602, the intention recognition module 6021 is configured to determine the intention of the voice information based on the identified dialogue turn. The character analysis module 6022 is configured to analyze the character string included in the voice information based on the identified dialogue turn. The target format judgment module 6023 is configured to determine whether the character condition required by the target format is met according to the analyzed character string. The response generation module 6024 is configured to generate response information based on the historical dialogue turns including the identified dialogue turn, the intention of the voice information, and the analyzed character string when the determination result is that the target format is not met.

[0193] According to the technical scheme provided by the embodiment of the present disclosure, the technical scheme comprises a duplex conversation manager and a specific format character collection device, wherein the duplex conversation manager comprises: a voice information acquisition module configured to acquire voice information; and a voice information processing module configured to process the voice information to identify at least one dialogue turn, and the specific format character collection device comprises: an intention identification module configured to determine the intention of the voice information based on the identified dialogue turn; a character analysis module configured to analyze a character string included in the voice information based on the identified dialogue turn; a target format judgment module configured to determine whether the character string meets the character condition required by a target format according to the analyzed character string; and a response generation module configured to generate response information based on historical dialogue turns including the identified dialogue turn, the intention of the voice information and the analyzed character string when the determination result is that the target format is not met, so that the response information can be automatically generated in a duplex conversation scene to quickly and accurately collect characters in a specific format.

[0194] Those skilled in the art can understand that the technical scheme described above can be combined with the embodiments described with reference to Figures 1 to 5 the drawings, so as to have the technical effects realized by the embodiments described with reference to Figures 1 to 5 the drawings. The specific content can be referred to the description made above according to Figures 1 to 5 the drawings, and the specific content will not be repeated here. Figure 7

[0195] Figure 7 A structural block diagram of an electronic device according to an embodiment of the present disclosure is shown.

[0196] The embodiment of the present disclosure further provides an electronic device, as shown in the drawings, comprising at least one processor 701; and a memory 702 connected with the at least one processor 701 in communication; wherein the memory 702 stores instructions executable by the at least one processor 701, and the instructions are executed by the at least one processor 701 to implement the following steps: Figure 8 acquiring voice information;

[0197] processing the voice information to identify at least one dialogue turn;

[0198] determining the intention of the voice information based on the identified dialogue turn;

[0199] analyzing a character string included in the voice information based on the identified dialogue turn;

[0200] determining whether the character string meets the character condition required by a target format according to the analyzed character string;

[0201]

[0202] ​​When the determination result is that the target format is not met, the response information is generated based on the historical dialogue turn including the recognized dialogue turn, the intent of the voice information, and the parsed string.

[0203] In an embodiment of the present disclosure, the memory 702 stores instructions executable by the at least one processor 701 to implement the following steps:

[0204] When the determination result is that the target format is met, the confirmation response information is generated based on the parsed string.

[0205] In an embodiment of the present disclosure, the character condition includes at least one of a character number, a character type, and a character arrangement manner of the target format.

[0206] In an embodiment of the present disclosure, the memory 702 stores instructions executable by the at least one processor 701 to implement the following steps:

[0207] Based on the historical dialogue turn including the recognized dialogue turn, the intent of the voice information, and the parsed string, the dialogue party action of a dialogue party sending the voice information and the historical character collection result including the parsed string are determined,

[0208] The determination whether the character condition required by the target format is met according to the parsed string includes:

[0209] The determination whether the character condition required by the target format is met according to the historical character collection result including the parsed string.

[0210] In an embodiment of the present disclosure, the intent of the voice information is an affirmative or negative intent, or the intent of the voice information is a global intent.

[0211] In an embodiment of the present disclosure, the determination of the intent of the voice information based on the recognized dialogue turn includes:

[0212] The intent of the voice information is determined to be an affirmative or negative intent, or the intent of the voice information is determined to be a global intent based on the recognized dialogue turn by using a keyword combined with a classification model of supervised learning.

[0213] In an embodiment of the present disclosure, the parsing of the string included in the voice information based on the recognized dialogue turn includes:

[0214] The character fragment is extracted based on the recognized dialogue turn;

[0215] The character fragment is regularized to obtain a regularized string.

[0216] normalize the regularized string to obtain a normalized string.

[0217] In an embodiment of the present disclosure, the regularizing the character segment to obtain the regularized string comprises:

[0218] detecting whether multiple character segments exist for the same dialogue turn and whether a time interval between the multiple character segments is less than a preset threshold value;

[0219] in response to detecting the multiple character segments and the time interval between the multiple character segments being less than the preset threshold value, determining whether multiple candidate strings obtained based on the regularizing the multiple character segments are repeated;

[0220] based on a determination that the multiple candidate strings are repeated, retaining one candidate regularized string as the regularized string.

[0221] In an embodiment of the present disclosure, the determining the dialogue party action of the dialogue party sending the speech information based on the historical dialogue turns including the recognized dialogue turn, the intent of the speech information and the parsed string, and the historical character collection result including the parsed string comprises:

[0222] using a keyword combined training supervised learning classification model to determine the dialogue party action of the dialogue party sending the speech information based on the historical dialogue turns including the recognized dialogue turn, the intent of the speech information and the parsed string, and the historical character collection result including the parsed string.

[0223] In an embodiment of the present disclosure, the memory 702 stores instructions executable by the at least one processor 701 to implement the following steps:

[0224] using a beam search based algorithm to decode the to-be-collected string based on the current character segment extracted from the current dialogue turn, the prior character segment extracted from the dialogue turn before the current dialogue turn and the determined dialogue party action of the dialogue party sending the speech information, wherein in the beam search based algorithm, a coefficient less than 1 is assigned to a repeated candidate path of the current character segment and the prior character segment, and / or based on a determination result that the intent of the speech information is a negative intent, a decoding result of a string corresponding to a previous dialogue turn of the current dialogue turn is revoked.

[0225] In an embodiment of the present disclosure, when the determination result is that the target format is not met, the response information is generated based on the historical dialogue turns including the recognized dialogue turn, the intent of the voice information, and the parsed string.

[0226] When the determination result is that the target format is not met, the response information is generated based on the historical dialogue turns including the recognized dialogue turn, the intent of the voice information, the parsed string, and the dialogue party action of the dialogue party sending the voice information, by using a supervised model based on a deep neural network.

[0227] In an embodiment of the present disclosure, the memory 702 stores instructions executable by the at least one processor 701 to implement the following steps:

[0228] The response information or the confirmation response information in the natural language text is generated based on the response information or the confirmation response information.

[0229] Figure 8 is a structural schematic diagram of a computer system suitable for implementing the method according to the embodiments of the present disclosure. As shown in ​ The computer system 800 includes a processing unit 801, which can perform various processes in the above-mentioned embodiments shown in the accompanying drawings according to programs stored in a read-only memory (ROM) 802 or loaded into a random access memory (RAM) 803 from a storage portion 808. Various programs and data required for the operation of the system 800 are also stored in the RAM 803. The CPU 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0230] The following components are connected to the I / O interface 805: an input portion 806 including a keyboard, a mouse, and the like; an output portion 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker, and the like; a storage portion 808 including a hard disk, and the like; and a communication portion 809 including a network interface card such as a LAN card, a modem, and the like. The communication portion 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as needed. A removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is mounted on the drive 810 as needed, so that a computer program read therefrom is installed in the storage portion 808 as needed. Among them, the processing unit 801 can be implemented as a CPU, a GPU, a TPU, a FPGA, an NPU, and the like.

[0231] In particular, the method described above with reference to the accompanying drawings can be implemented as a computer software program according to embodiments of the present disclosure. For example, embodiments of the present disclosure include a computer program product comprising a computer program tangibly embodied on a non-transitory computer readable medium, the computer program containing program code for executing the methods of the accompanying drawings. In such embodiments, the computer program can be downloaded and installed from a network via the communication section 809, and / or installed from the removable medium 811. For example, embodiments of the present disclosure include a readable storage medium having stored thereon computer instructions which, when executed by a processor, implement program code for executing the methods of the accompanying drawings.

[0232] The flow diagrams and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow diagrams and block diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0233] The units or modules described in the embodiments of the present disclosure can be implemented by software, or by hardware, or by a combination of software and hardware. The described units or modules can also be provided in a processor, and the names of the units or modules do not constitute a limitation on the units or modules themselves in some cases.

[0234] As another aspect, the present disclosure also provides a computer readable storage medium, which can be the computer readable storage medium contained in the node described in the above embodiments, or can exist separately from the device and not be assembled into the device. The computer readable storage medium stores one or more programs, which are executed by one or more processors to perform the methods described in the present disclosure.

[0235] The above description is merely that of the preferred embodiments of the present disclosure and a description of the technical principles of the present disclosure. It should be understood by those skilled in the art that the inventive scope involved in the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by the combinations of the above technical features or equivalent features without departing from the inventive concept. For example, the technical solutions formed by the mutual replacement of the above features and the technical features with similar functions disclosed in the present disclosure (but not limited to) without departing from the inventive concept.

Claims

1. A method for collecting characters of a specific format, comprising: obtaining speech information; processing the speech information to identify at least one dialogue turn; determining an intent of the speech information based on the identified dialogue turn; parsing a character string included in the speech information based on the identified dialogue turn; determining whether the parsed character string meets a character condition required by a target format; generating a response message based on historical dialogue turns including the identified dialogue turn, the intent of the speech information, and the parsed character string, when a result of the determination is that the parsed character string does not meet the target format. The method further comprises determining a dialogue party action of a dialogue party that utters the speech information and a historical character collection result including the parsed character string based on historical dialogue turns including the identified dialogue turn, the intent of the speech information, and the parsed character string. The determination of whether the parsed character string meets the character condition required by the target format comprises determining whether the parsed character string meets the character condition required by the target format according to a historical character collection result including the parsed character string. The method further comprises decoding a character string to be collected based on a current character segment extracted from a current dialogue turn, a previous character segment extracted from a dialogue turn before the current dialogue turn, and the determined dialogue party action of the dialogue party that utters the speech information using a beam search based algorithm, wherein a coefficient less than 1 is assigned to a candidate path in which the current character segment and the previous character segment are repeated in the beam search based algorithm, and / or a decoding result of a character string corresponding to a dialogue turn before the current dialogue turn is revoked based on a determination result that the intent of the speech information is a negative intent. 2.The method of claim 1, further comprising: generating a confirmation response message based on the parsed character string when a result of the determination is that the parsed character string meets the target format. 3.The method of claim 1, wherein the character condition comprises at least one of a number of characters, a type of character, and an arrangement of characters of the target format. 4.The method of any one of claims 1 to 3, wherein the intent of the speech information is a positive intent or a negative intent, or an intent unrelated to the current character collection process is a global intent. 5.The method of claim 4, wherein the determination of the intent of the speech information based on the identified dialogue turn comprises: determining the intent of the speech information to be the positive intent or the negative intent, or determining the intent of the speech information to be the global intent based on the identified dialogue turn using a classification model of supervised learning combined with keywords. 6.The method of claim 4, wherein the parsing of the character string included in the speech information based on the identified dialogue turn comprises: extracting a character segment based on the identified dialogue turn; performing a regularization process on the character segment to obtain a regularized character string; performing a normalization process on the regularized character string to obtain a normalized character string.

7. The method of claim 6, wherein the normalizing the character segment to obtain a normalized string comprises: detecting whether there are multiple character segments for a same dialog turn and whether time intervals between the multiple character segments are less than a preset threshold; in response to detecting the multiple character segments and the time intervals between the multiple character segments being less than the preset threshold, determining whether multiple candidate strings obtained based on the normalizing the multiple character segments are duplicated; and based on a determination that the multiple candidate strings are duplicated, retaining one candidate normalized string as the normalized string.

8. The method of claim 1, wherein the determining, based on the historical dialog turns including the identified dialog turn, the intent of the voice information, and the parsed string, a dialog party action of a dialog party that utters the voice information and historical character collection results including the parsed string comprises: utilizing a classification model of supervised learning trained with keywords to determine, based on the historical dialog turns including the identified dialog turn, the intent of the voice information, and the parsed string, the dialog party action of the dialog party that utters the voice information and the historical character collection results including the parsed string.

9. The method of claim 1, wherein the generating, based on the historical dialog turns including the identified dialog turn, the intent of the voice information, and the parsed string, a response information when the determination result is that the parsed string does not satisfy the target format comprises: utilizing a supervised model based on deep neural network to generate, based on the historical dialog turns including the identified dialog turn, the intent of the voice information, the parsed string, and the dialog party action of the dialog party that utters the voice information, the response information when the determination result is that the parsed string does not satisfy the target format.

10. The method of claim 2, further comprising: generating, based on the response information or the confirmation response information, a response information or a confirmation response information of natural language text.

11. An apparatus for collecting a character of a specific format, comprising: a voice information obtaining module configured to obtain voice information; a voice information processing module configured to process the voice information to identify at least one dialog turn; an intent identifying module configured to determine an intent of the voice information based on the identified dialog turn; a character parsing module configured to parse a string included in the voice information based on the identified dialog turn; a target format judging module configured to determine whether the parsed string satisfies a character condition required by a target format; and a response generating module configured to generate, based on the historical dialog turns including the identified dialog turn, the intent of the voice information, and the parsed string, a response information when the determination result is that the parsed string does not satisfy the target format. ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ The device is further configured to perform the following steps: determining a dialog party action of a dialog party sending the voice information and a historical character collection result including the parsed string based on historical dialog turns including the recognized dialog turn, the intent of the voice information, and the parsed string. The target format judging module is further configured to judge whether the parsed string meets a character condition required by a target format by performing the following steps: judging whether the parsed string meets a character condition required by a target format according to a historical character collection result including the parsed string. The device is further configured to perform the following steps: decoding a to-be-collected string based on a current character segment extracted from a current dialog turn, a previous character segment extracted from a dialog turn before the current dialog turn, and the determined dialog party action of the dialog party sending the voice information by using a beam search-based algorithm, wherein a coefficient less than 1 is assigned to a candidate path in which the current character segment and the previous character segment are repeated in the beam search-based algorithm, and / or a decoding result of a string corresponding to a previous dialog turn of the current dialog turn is revoked based on a determination result that the intent of the voice information is a negative intent.

12. A duplex voice character collection system, comprising a duplex session manager and a specific format character collection device, wherein The duplex session manager comprises: a voice information obtaining module configured to obtain voice information; a voice information processing module configured to process the voice information to recognize at least one dialog turn, The specific format character collection device comprises: an intent recognition module configured to determine an intent of the voice information based on the recognized dialog turn; a character parsing module configured to parse a string included in the voice information based on the recognized dialog turn; a target format judging module configured to judge whether the parsed string meets a character condition required by a target format; a response generating module configured to generate a response message based on historical dialog turns including the recognized dialog turn, the intent of the voice information, and the parsed string when the judgment result is that the parsed string does not meet the target format; The system is further configured to perform the following steps: determining a dialog party action of a dialog party sending the voice information and a historical character collection result including the parsed string based on historical dialog turns including the recognized dialog turn, the intent of the voice information, and the parsed string. The target format judging module is further configured to judge whether the parsed string meets a character condition required by a target format by performing the following steps: judging whether the parsed string meets a character condition required by a target format according to a historical character collection result including the parsed string. The device is further configured to perform the following steps: decoding a to-be-collected string based on a current character segment extracted from a current dialog turn, a previous character segment extracted from a dialog turn before the current dialog turn, and the determined dialog party action of the dialog party sending the voice information by using a beam search-based algorithm, wherein a coefficient less than 1 is assigned to a candidate path in which the current character segment and the previous character segment are repeated in the beam search-based algorithm, and / or a decoding result of a string corresponding to a previous dialog turn of the current dialog turn is revoked based on a determination result that the intent of the voice information is a negative intent. The system is further configured to perform the following steps: decoding, by using a beam search based algorithm, a string to be collected based on the current character segment extracted from the current dialogue turn, the prior character segment extracted from the dialogue turn before the current dialogue turn, and the determined dialogue party action of the dialogue party that utters the voice information, wherein a coefficient less than 1 is assigned to a repeated candidate path of the current character segment and the prior character segment in the beam search based algorithm, and / or a decoding result of a string corresponding to a dialogue turn before the current dialogue turn is revoked based on a determination result that an intent of the voice information is a negative intent. 13.An electronic device comprising a memory and a processor; wherein, The memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the method according to any one of claims 1-10. 14.A readable storage medium having computer instructions stored thereon, wherein the computer instructions are executed by a processor to implement the method according to any one of claims 1-10. 15.A computer program product comprising computer instructions, wherein the computer instructions are executed by a processor to implement the method according to any one of claims 1-10.

Images