Digital consecutive interpretation method, device and equipment based on large language model

Abstract

The embodiments of the present application disclose a method, device and equipment for digital consecutive interpretation based on a large language model, which belongs to the field of speech and text processing technology. The embodiments of the present application collect multiple groups of speech information to be translated in the form of audio texts through a voice input terminal, and generate a first interactive voice information containing digital interactive texts, thereby realizing the orderly integration of information. When the digital consecutive interpretation mechanism is triggered, semantic vector mining is performed with the help of a large language model system, which greatly reduces the cognitive load of the interpreter. The interpreter does not need to deeply analyze complex semantic relationships by himself. The use of pre-stored benchmark semantic interaction vectors and digital annotation indication strategies for digital consecutive interpretation improves accurate characterization and avoids possible errors in digital conversion and semantic understanding caused by manual labor. The entire process has a high degree of automation, which improves the efficiency of interpretation and can meet the needs of rapid progress in business communication in a timely manner.

Description

Technical Field

[0001] The embodiments of the present application belong to the technical field of speech and text processing, and specifically relate to a digital consecutive interpretation method, device and equipment based on a large language model. Background Art

[0002] In the frequent conversational communication scenarios of multinational companies, such as business meetings and project negotiations, digital consecutive interpretation faces many challenges. Traditional digital consecutive interpretation mainly relies on interpreters to manually process a large amount of digital information. Interpreters need to listen to and process the voice information of each party separately in different consecutive interpretation cycles, including the extraction, understanding and conversion of digital information. This puts interpreters under a huge cognitive load, which is prone to fatigue and distraction. At the same time, when manually processing digital consecutive interpretation, errors are prone to occur in semantic understanding, digital conversion, and compliance with digital annotation habits in the target language, making it difficult to ensure the accuracy of interpretation. Moreover, the speed of manual processing is limited, and it is difficult to efficiently keep up with the fast pace of business communication, which leads to the problem of low interpretation efficiency. Summary of the invention

[0003] The embodiments of the present application provide a method, device and equipment for digital consecutive interpretation based on a large language model, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] The embodiment of the present application provides a digital consecutive interpretation method based on a large language model, which is applied to a digital consecutive interpretation device. The method comprises: collecting first voice information to be translated corresponding to a first conversation scenario through a voice input terminal, wherein the first voice information to be translated includes N groups of audio texts of a target conversation event, and the first conversation scenario represents the interpretation task state between the target conversation event and the voice input terminal in a first consecutive interpretation cycle, and N is an integer not less than 1; collecting second voice information to be translated corresponding to a second conversation scenario through the voice input terminal, wherein the second voice information to be translated includes M groups of audio texts of the target conversation event, and the second conversation scenario represents the interpretation task state between the target conversation event and the voice input terminal in a second consecutive interpretation cycle, and the The second consecutive interpretation cycle is the next consecutive interpretation cycle of the first consecutive interpretation cycle, and M is an integer not less than 1; based on the first voice information to be translated and the second voice information to be translated, first interactive voice information is generated, wherein the first interactive voice information includes digital interactive texts of the N groups of audio texts and the M groups of audio texts; if the first interactive voice information triggers the digital consecutive interpretation mechanism, semantic vector mining is performed on the first interactive voice information to obtain a first semantic interaction vector; wherein the semantic vector mining is implemented through a large language model system; using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector, digital consecutive interpretation is performed on the first semantic interaction vector to obtain a consecutive interpretation annotated text.

[0005] The present application provides a digital consecutive interpretation device, the digital consecutive interpretation device comprising:

[0006] A voice collection module, used for collecting first voice information to be translated corresponding to a first conversation scenario through a voice input terminal, wherein the first voice information to be translated includes N groups of audio texts of a target conversation event, the first conversation scenario represents the state of an interpretation task between the target conversation event and the voice input terminal in a first consecutive interpretation cycle, and N is an integer not less than 1; collecting second voice information to be translated corresponding to a second conversation scenario through the voice input terminal, wherein the second voice information to be translated includes M groups of audio texts of the target conversation event, the second conversation scenario represents the state of an interpretation task between the target conversation event and the voice input terminal in a second consecutive interpretation cycle, the second consecutive interpretation cycle is the next consecutive interpretation cycle of the first consecutive interpretation cycle, and M is an integer not less than 1;

[0007] A speech generation module, configured to generate first interactive speech information based on the first speech information to be translated and the second speech information to be translated, wherein the first interactive speech information includes digital interactive texts of the N groups of audio texts and the M groups of audio texts;

[0008] A semantic mining module, configured to perform semantic vector mining on the first interactive voice information to obtain a first semantic interaction vector if the first interactive voice information triggers a digital consecutive interpretation mechanism; wherein the semantic vector mining is implemented by a large language model system;

[0009] The consecutive interpretation module is used to perform digital consecutive interpretation on the first semantic interaction vector by using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector to obtain a consecutive interpretation annotated text.

[0010] An embodiment of the present application provides a digital consecutive interpretation device, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the above method.

[0011] An embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0012] In an embodiment of the present application, multiple groups of voice information to be translated in the form of audio text are collected through the voice input terminal, and a first interactive voice information containing digital interactive text is generated, thereby realizing orderly integration of information. When the digital consecutive interpretation mechanism is triggered, semantic vector mining is performed with the help of a large language model system, which greatly reduces the cognitive load of the interpreter. The interpreter does not need to deeply analyze complex semantic relationships by himself. The use of pre-stored benchmark semantic interaction vectors and digital annotation indication strategies for digital consecutive interpretation improves accurate characterization and avoids possible manual errors in digital conversion and semantic understanding. The entire process has a high degree of automation, which improves the efficiency of interpretation and can meet the needs of rapid progress in business communication in a timely manner. The interpretation method supports mutual translation of multiple languages, and the annotation symbols of notes can be customized or personalized. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flowchart of a digital consecutive interpretation method based on a large language model provided in an embodiment of the present application.

[0014] Figure 2 A schematic diagram of the structure of a digital consecutive interpretation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of them. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0016] The terms "first", "second", etc. in the embodiments of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the embodiments of the present application can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally a class, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the embodiments of the present application, "and / or" represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0017] Figure 1 A digital consecutive interpretation method based on a large language model is shown, which is applied to a digital consecutive interpretation device. The method includes the following steps 110 to 150.

[0018] Step 110: The digital consecutive interpretation device collects first speech information to be translated corresponding to the first conversation scenario through the speech input terminal.

[0019] The first voice information to be translated includes N groups of audio texts of a target conversation event, the first conversation scenario represents the state of the interpretation task between the target conversation event and the voice input end in a first consecutive interpretation cycle, and N is an integer not less than 1.

[0020] Step 120: The digital consecutive interpretation device collects second voice information to be translated corresponding to the second conversation scenario through the voice input terminal.

[0021] The second voice information to be translated includes M groups of audio texts of the target conversation event, the second conversation scenario represents the interpretation task state between the target conversation event and the voice input end in a second consecutive interpretation cycle, the second consecutive interpretation cycle is the next consecutive interpretation cycle of the first consecutive interpretation cycle, and M is an integer not less than 1.

[0022] Step 130: The digital consecutive interpretation device generates first interactive voice information based on the first voice information to be translated and the second voice information to be translated.

[0023] The first interactive voice information includes digital interactive texts of the N groups of audio texts and the M groups of audio texts.

[0024] Step 140: If the first interactive voice information triggers the digital consecutive interpretation mechanism, the digital consecutive interpretation device performs semantic vector mining on the first interactive voice information to obtain a first semantic interaction vector.

[0025] Wherein, the semantic vector mining is implemented through a large language model system.

[0026] Step 150: The digital consecutive interpretation device uses a plurality of pre-stored reference semantic interaction vectors and a digital annotation indication strategy corresponding to each reference semantic interaction vector to perform digital consecutive interpretation on the first semantic interaction vector to obtain a consecutive interpretation annotated text.

[0027] In step 110 to step 150 , the digital consecutive interpretation device (digital consecutive interpretation equipment) is respectively connected to the speech input terminal and the large language model system for communication.

[0028] In complex communication scenarios such as business meetings between multinational companies, there are numerous conversation events. The digital consecutive interpretation device plays a key role in the whole process. First, in step 110, the digital consecutive interpretation device collects the first voice information to be translated corresponding to the first conversation scene through the voice input terminal. For example, in a business meeting of a multinational company on a large-scale project cooperation, the first voice information to be translated may contain multiple groups of audio texts of the target conversation event. For example, the target conversation event is a discussion about the project budget and market share estimation, and the first conversation scene represents the state of the interpretation task between the target conversation event and the voice input terminal in the first consecutive interpretation cycle. The N groups of audio texts (N is an integer not less than 1) contained in the first voice information to be translated here may be various contents involved in the speech of a representative of one enterprise, such as the preliminary budget of the enterprise for the project investment, which may mention "We plan to invest about five million US dollars, of which R&D expenses account for about one third" and other similar words containing digital information.

[0029] Then, in step 120, the digital consecutive interpretation device continues to collect the second voice information to be translated corresponding to the second conversation scene through the voice input terminal. This process is similar to collecting the first voice information to be translated, but it is in the next consecutive interpretation cycle. For example, in a business meeting, the representative of the other company responds to the previous speech and may mention his own resource investment and expected returns. The second voice information to be translated includes M groups of audio texts of the target conversation event (M is an integer not less than 1), such as "We can provide equipment worth two million US dollars, and we expect to reach a market share of 20% within three years."

[0030] Then, in step 130, the digital consecutive interpretation device generates a first interactive voice message based on the first voice message to be translated and the second voice message to be translated. This first interactive voice message contains the digital interactive text of N groups of audio texts and M groups of audio texts. Taking the previous business meeting as an example, the first interactive voice message may contain interactive content related to numbers such as the project investment budget (US$5 million, US$2 million), the proportion of R&D expenses (one-third), and the expected market share (20 percent) mentioned by both parties. These numbers are of key significance in business decision-making and communication, and accurate interactive processing helps both parties better understand each other's intentions and resource investment.

[0031] In step 140, if the first interactive voice information triggers the digital consecutive interpretation mechanism, the digital consecutive interpretation device will perform semantic vector mining on the first interactive voice information, and this process is achieved through a large language model system. The large language model system has unique advantages in processing numbers in these voice messages. For example, the large language model system can recognize numbers in different forms based on its pre-training results. For fractional forms such as "one third", percentage forms such as "twenty percent", and specific values ​​such as "five million US dollars" and "two million US dollars", the large language model system can accurately understand their meaning in the entire semantics. Through mining, the first semantic interaction vector is obtained. For example, in this business meeting scenario, the first semantic interaction vector contains digital vector representations of the semantic level such as the comparison of the investment budgets of both parties, the proportion relationship, and the expected relationship of market share. For example, the comparison of investment budgets may be reflected in the vector as a component representing the difference (three million US dollars), the proportion relationship is reflected as a component representing the proportion difference (one sixth), and the expected relationship of market share is reflected as a component representing the gap (a few percent), etc.

[0032] Finally, in step 150, the digital consecutive interpretation device uses the pre-stored multiple benchmark semantic interaction vectors and the digital annotation indication strategy corresponding to each benchmark semantic interaction vector to perform digital consecutive interpretation on the first semantic interaction vector to obtain the consecutive interpretation annotated text. The pre-stored benchmark semantic interaction vectors can be established based on typical digital semantic relationships in a large number of business meeting scenarios. For example, for the comparison of investment budgets, there is a benchmark semantic interaction vector that indicates the increase in investment budget, and its corresponding digital annotation indication strategy may be to mark the specific amount of increase with a specific symbol or format; for the change in market share, another benchmark semantic interaction vector indicates the increase in market share, and its corresponding digital annotation indication strategy may be to use a certain color mark or a special format to reflect it in the consecutive interpretation annotated text. In this way, the digital consecutive interpretation device can accurately convert the digital information in the first semantic interaction vector into a consecutive interpretation annotated text that conforms to the target language and business communication habits. For example, when converting digital information involving dollar amounts into the target language (such as Chinese), it is marked according to the Chinese digital reading and writing rules, such as "five million US dollars" is marked as "five million US dollars", and when it comes to digital comparison, it is marked according to the requirements of accurately expressing data relationships in business communication. For example, "the investment budget is three million US dollars more than the other party" and other content are accurately reflected in the consecutive interpretation annotated text, thereby ensuring that both parties of multinational companies can clearly and accurately understand each other's digital information in business meetings.

[0033] Throughout the entire process, the role of digital consecutive interpretation devices in multinational business meetings cannot be ignored. Starting from collecting the voice information of both parties, after a series of processing steps, accurate consecutive interpretation annotation text is finally generated, providing efficient and accurate digital information processing guarantee for business exchanges between multinational companies.

[0034] In terms of speech recognition, as mentioned above, when collecting the first speech information to be translated and the second speech information to be translated, the speech recognition module at the speech input end connected to the digital consecutive interpretation device plays a fundamental role. It processes speech signals based on acoustic models and language models. The acoustic model extracts acoustic features from speech signals. For example, for the accents and speech speeds of different representatives in a business meeting, the acoustic model can accurately extract their acoustic features and convert them into feature vectors through training with a large amount of speech data. The language model constrains and optimizes the results output by the acoustic model according to the grammar, semantics and other rules of the language. Taking English speeches in business meetings as an example, when someone mentions "One million five hundred thousand dollars", the acoustic model first converts the speech into feature vectors. The language model determines that the most likely word combination is this accurate expression based on these vectors and the grammar and semantic rules of English, rather than other semantically incomprehensible combinations. This accurate speech-to-text conversion lays the foundation for the subsequent processing of digital information.

[0035] In the digital information processing link, the advantages of the large language model are fully reflected. After the first interactive voice information is generated, the large language model can recognize various forms of numbers in it. For example, the large language model can accurately recognize different forms of numbers such as "about 0.5 probability" and "35% return rate" mentioned in business meetings. And it is of great significance to convert numbers in different forms into Arabic numerals. For example, the Chinese character number "three million" is converted to "3000000", and the uppercase Chinese character number "one million" is also converted to "1000000". This unified Arabic numeral form is more convenient in subsequent processing. When marking numbers for the target language, the large language model will operate according to the digital note-taking habits of the target language. For example, when the target language is English, the number 10000 will be marked as 10,000; when the target language is Chinese, 12345 will be understood and marked according to the reading rules of four groups.

[0036] The application of this application based on digital consecutive interpretation device in business meetings of multinational enterprises has obvious advantages over traditional manual digital consecutive interpretation. When processing numbers in business meetings in traditional manual methods, interpreters need to quickly process numbers in their brains while listening to speeches, including identifying numbers in different forms, converting them, and considering the digital expression habits of the target language. For example, various digital information that frequently appears in business meetings is prone to fatigue or distraction when processed manually. The digital consecutive interpretation device in the embodiment of the present application reduces this cognitive load through an automated process. In terms of accuracy, manual processing is prone to errors due to various factors, such as misjudgment of different digital representations or misunderstanding of target language habits. The large language model is trained based on a large amount of data and has high accuracy in digital recognition, conversion and annotation, thereby effectively avoiding digital translation errors caused by human factors. In terms of efficiency, automated processing saves interpreters time to process numbers, can better keep up with the fast pace of speeches in business meetings, and reduces the correction time after possible errors, significantly improving the overall efficiency of consecutive interpretation.

[0037] Further analysis shows that in a business meeting scenario, the speeches of different enterprise representatives may contain complex digital relationships and logic. For example, when discussing project costs and benefits, cost breakdown may be involved, such as raw material costs, labor costs, equipment costs, etc., each accounting for a certain proportion of the total cost, and these proportions may change as the project progresses. In the first voice information to be translated and the second voice information to be translated, these complex digital relationships need to be accurately captured and processed. When processing this information, the digital consecutive interpretation device accurately extracts these digital relationships from the voice information collected from the voice input end, and performs reasonable processing in subsequent steps. For example, when generating the first interactive voice information in step 130, it is not just a simple combination of the numbers in the two groups of voice information, but a reasonable interactive combination based on the semantic relationship. If the first group of voice information mentions that the raw material cost accounts for 30% of the total cost, and the second group of voice information mentions that the raw material cost is expected to decrease by 5% under the new plan, then this digital interactive relationship will be accurately reflected in the first interactive voice information, providing an accurate basis for subsequent semantic vector mining and digital consecutive interpretation.

[0038] In the semantic vector mining step 140, the large language model system processes the first interactive voice information based on its pre-training results for a large amount of business-related text data. The neural network structure inside the large language model system can deeply mine the digital semantic relationship in the first interactive voice information. Taking the discussion on market share competition in a business meeting as an example, if the first interactive voice information contains market share data of different companies in different regions and information such as the growth trend of market share, the large language model system can convert this information into a semantic interaction vector. This vector can accurately represent important business digital semantic information such as the comparative relationship between market shares of different companies and the difference in growth rate. For example, the market share of a company in a certain region increases from 10% to 15%, and the market share of another company in the same region decreases from 15% to 10%. These digital relationships will be represented in the semantic interaction vector in a way that can accurately reflect this change relationship.

[0039] Finally, in step 150, when the pre-stored multiple benchmark semantic interaction vectors and the corresponding digital annotation indication strategies are used for digital consecutive interpretation, these pre-stored information is carefully constructed. For common digital relationship types in business meetings, there are corresponding benchmark semantic interaction vectors. For example, for different business situations such as cost reduction, market share growth, and profit increase, there are corresponding benchmark semantic interaction vectors. The digital annotation indication strategy corresponding to each benchmark semantic interaction vector is also formulated according to the needs of business communication. For example, for the case of cost reduction, the digital annotation indication strategy may be to use specific words (such as "reduced" and "reduced") plus specific values ​​in the consecutive interpretation annotation text to accurately express the cost reduction. For the case of market share growth, it may be marked with words such as "rising to" and "increased" plus specific values ​​and percentages. In this way, the digital consecutive interpretation device can accurately convert the digital information in the business meeting into a consecutive interpretation annotation text that conforms to the business communication habits and the requirements of the target language, thereby ensuring the accuracy and efficiency of business communication between multinational companies in the transmission of digital information.

[0040] In the complex scenario of multinational business meetings, the digital consecutive interpretation device in the embodiment of the present application is closely linked and interconnected from the collection of voice information, the generation of digital interactive voice information, semantic vector mining to the final generation of digital consecutive interpretation annotated text. Each step provides a guarantee for accurate and efficient processing of digital information in business meetings, thereby improving the efficiency and accuracy of consecutive interpretation.

[0041] In actual application, when the priority of the first conversation scenario is higher than that of the second conversation scenario; the generation of the first interactive voice information based on the first voice information to be translated and the second voice information to be translated in step 130 can be implemented by any one of the following embodiments 1 to 3.

[0042] Embodiment 1: Based on the content load corresponding to the second voice information to be translated, a first feature adjustment is performed on the first voice information to be translated to obtain first voice adjustment information, wherein an audio text in the first voice adjustment information is mapped to an audio text in the second voice information to be translated; the second voice information to be translated is used as the second voice adjustment information; content interaction is performed on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information.

[0043] Embodiment 2: The first voice information to be translated is used as the first voice adjustment information; based on the content load corresponding to the first voice information to be translated, the second feature adjustment is performed on the second voice information to be translated to obtain the second voice adjustment information, wherein the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated; the first voice adjustment information and the second voice adjustment information are content-interacted to obtain the first interactive voice information.

[0044] Embodiment 3: Based on the set feature fine-grained weights, the first voice information to be translated is optimized in feature fine-grained manner to obtain first voice adjustment information; based on the set feature fine-grained weights, the second voice information to be translated is optimized in feature fine-grained manner to obtain second voice adjustment information; the first voice adjustment information and the second voice adjustment information are interacted with each other in content to obtain the first interactive voice information.

[0045] In a business meeting scenario between multinational companies, the technical solution of generating the first interactive voice information in step 130 when the priority of the first conversation scenario is higher than that of the second conversation scenario has far-reaching and complex connotations and important significance.

[0046] In Example 1, the process of performing a first feature adjustment on the first voice information to be translated to obtain the first voice adjustment information based on the content load corresponding to the second voice information to be translated is very delicate and critical. Imagine a business meeting scene of a multinational company, where the first voice information to be translated often contains rich and diverse content. For example, when a representative of a company explains the investment plan for a large project, he may mention a lot of detailed digital information, such as "We initially plan to make a large-scale capital investment in this project, with a total budget set at US$80 million. Among them, the initial market research and project preparation stage is expected to invest US$10 million, and this stage will last about 6 months. In the first year after the project is officially launched, we plan to invest US$30 million in infrastructure construction and equipment procurement, and it is expected to recruit 200 professionals to participate in this stage of work. As the project progresses, US$20 million and US$10 million will be invested in the second and third years respectively for the expansion and improvement of the project. It is expected to achieve profitability by the end of the third year of the project operation, with an expected profit of US$50 million." This is a fairly comprehensive and complex investment plan, which includes a lot of numerical information such as investment amounts, time periods, and expected profits at different stages.

[0047] The corresponding second voice message to be translated may be a brief response from the representative of the other enterprise to some of the content, such as "We believe that the investment of 30 million US dollars in infrastructure and equipment procurement in the first year after the project is launched is too high. Based on our experience and market analysis, 25 million US dollars may be enough." Here, the content load of the second voice message to be translated is relatively small. It only responds to part of the content in the first voice message to be translated (the investment amount in the first year). In this case, when adjusting the features of the first voice message to be translated, it is necessary to highlight the part related to the investment amount in the first year based on the content load of the second voice message to be translated. This may mean elaborating on the composition and necessity of the investment amount of 30 million US dollars in the first year, such as 15 million US dollars for the procurement of specific high-end equipment, which plays a vital role in the quality and efficiency of the project; 10 million US dollars for the construction of infrastructure that meets international standards, including the construction of factory buildings and office areas; and another 5 million US dollars for related technology introduction and training. Through such adjustment, the first voice adjustment information is obtained. At this time, the audio text in the first voice adjustment information and the audio text in the second voice information to be translated are mapped, that is, they are all about the rationality of the first year's investment amount. After that, the second voice information to be translated is used as the second voice adjustment information, and the first voice adjustment information and the second voice adjustment information are interactive. This content interaction is not a simple combination, but a deep integration. It will make detailed associations and comparisons between the basis behind the different views of both parties on the first year's investment amount. For example, in the first interactive voice message, it will clearly show the basis for one party to believe that 30 million US dollars (the high-end nature of the equipment, the high standards of the infrastructure, and the need for technology introduction) and the other party to believe that 25 million US dollars is sufficient (may be based on different equipment procurement channels, cost differences in infrastructure construction, and different views on technology introduction), so as to fully show all the relevant information of both parties on this digital issue, including the investment amount, investment composition, and the considerations behind it.

[0048] In Example 2, the first voice information to be translated is used as the first voice adjustment information, and then the second voice information to be translated is adjusted according to the content load corresponding to the first voice information to be translated to obtain the second voice adjustment information. This process also has unique significance and value. Taking a business meeting as an example again, for example, the first voice information to be translated is about the overall plan of a multinational cooperation project, and this plan involves detailed digital information in multiple aspects. For example, when explaining a new product promotion project involving multiple countries, the company mentioned, "We plan to promote in three main target markets, namely Country A, Country B, and Country C. In the Country A market, we plan to invest $15 million in marketing expenses, including $8 million for advertising, $5 million for market channel development, and $2 million for sales promotion. In this market, we expect to gain a 10% market share in the first year, and increase the market share to 20% in the third year as the market promotion deepens. For the Country B market, our marketing budget is $12 million, which is allocated according to $6 million for advertising, $4 million for channel development, and $2 million for sales promotion. The goal is to gain an 8% market share in the first year and 15% in the third year. In the Country C market, we plan to invest $10 million, with $5 million for advertising, $3 million for channel development, and $2 million for sales promotion. We hope to achieve a 6% market share in the first year and increase it to 12% in the third year." This is an overall plan that includes many numerical details such as investment allocation in different markets, market share expectations, and time periods.

[0049] The second voice message to be translated is a brief reminder from the other party's enterprise representative about the project risks, such as "I think that in the market of Country A, due to the uncertainty of local policies and fierce market competition, this project has a great risk, which may lead to an overspending of marketing expenses of 3 million US dollars, and the growth of market share may not meet expectations, and can only reach 15% at most." Since the content load of the first voice message to be translated is large, when adjusting the features of the second voice message to be translated, it is necessary to adjust the second voice message to the risk analysis related to these contents based on the investment and market share expectations of each market in the first voice message to be translated. For the market of Country A, a detailed analysis is made of the possible changes in risks under different investment allocation situations (such as fluctuations in advertising expenses, changes in the difficulty of market channel development, etc.), and how these changes affect the growth of market share. For example, if the advertising expenses increase by 2 million US dollars due to local policy restrictions, it may lead to a reduction in funds for channel development and promotion activities, which in turn affects the acquisition of market share, and may only reach 12% of the market share in the end, and the overall marketing expenses will exceed 3 million US dollars. Through such adjustment, the second voice adjustment information is obtained. At this time, the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated, and both are about the relationship between investment, market share and risk in the A country market. After the content of the two voice adjustment information is interacted, the first interactive voice information can be obtained. This first interactive voice information contains various digital relationships involved in the planning of projects in different markets and the risk analysis for a specific market (A country), such as the relationship between the investment amount in different markets and the possible overspending amount in the A country market, and the relationship between the expected market share in different markets and the possible failure to reach the expected share in the A country market. The accurate presentation of these digital relationships is of vital importance for multinational companies to comprehensively evaluate the feasibility and risks of projects in business meetings.

[0050] In Example 3, based on setting feature fine-grained weights, the first voice information to be translated is optimized in feature fine-grained manner to obtain the first voice adjustment information, and at the same time, based on setting feature fine-grained weights, the second voice information to be translated is optimized in feature fine-grained manner to obtain the second voice adjustment information. This process is highly targeted and directional. For example, in a business meeting, setting feature fine-grained weights may be closely related to important factors in business decision-making, including core elements such as market share, profit margin, and cost control. If the first voice information to be translated contains the company's multi-faceted planning for an international project, including market share, the company expects to achieve a 35% share in a specific international market (a regional market composed of multiple countries); in terms of profit margin, the overall profit margin of the project is estimated to reach 25%, of which product sales profit accounts for 15% and service value-added profit accounts for 10%; in terms of cost control, it is planned to reduce costs by 15% by optimizing the supply chain and improving production efficiency, and specifically points out that the cost of raw material procurement will be reduced by 8%, labor costs will be reduced by 4%, and other operating costs will be reduced by 3%. According to the set feature fine-grained weights, if the weight of the market share is high, then when the first voice information to be translated is optimized at the feature fine-grained level, the content related to the market share will be processed more deeply and in more detail. For example, the specific market strategy for achieving a 35% market share will be further elaborated, and differentiated marketing strategies for different national markets may be mentioned. In countries with a large market share, brand promotion and customer relationship maintenance will be focused on, and more marketing resources will be invested. It is expected that the market share in these countries will be able to reach 40%-50%; while in countries with relatively small market shares, market cultivation will be emphasized, and the market share will be gradually expanded through cooperation with local companies. It is expected that the market share in these countries will gradually increase from the initial 5%-10% to 15%-20%, etc. The first voice adjustment information is obtained through such optimization.

[0051] Similarly, for the second voice information to be translated, if it contains doubts or supplements to the first enterprise plan, for example, it is believed that in this international market, due to local cultural differences and the strong strength of competitors, the market share is difficult to reach 35%, and can only reach 25% at most, and there are different views on profit margins and cost control, and it is believed that the overall profit margin may only reach 20%, of which product sales profit accounts for 12%, service value-added profit accounts for 8%, and cost control can only be reduced by 10% at most, 5% reduction in raw material procurement costs, 3% reduction in labor costs, and 2% reduction in other operating costs. When performing feature fine-grained optimization based on the set feature fine-grained weights, the market share part will be emphasized, and a detailed analysis will be made on why it is believed that the market share is difficult to reach 35%. Factors such as the brand advantages, market share basis, and effective marketing strategies for local cultural differences of competitors in the local market may be pointed out to obtain the second voice adjustment information. Then, the first voice adjustment information and the second voice adjustment information are interacted with each other to obtain the first interactive voice information. This first interactive voice information will very accurately present the digital interactive content of the two parties on important business factors such as market share, profit margin, and cost control. For example, it will clearly show the difference in market share expectations between the two parties (35% vs. 25%) and the basis behind them (different market strategies, competitor situations, cultural differences, etc.), different estimates of profit margins (25% vs. 20%) and the differences in their respective composition ratios, different views on cost control (reducing 15% vs. reducing 10%) and the specific differences in different cost items. This precise digital interactive information plays an irreplaceable role in helping multinational companies to deeply discuss project plans, negotiate cooperation terms, and evaluate the comprehensive benefits of projects in business meetings.

[0052] In the scenario of a multinational business meeting, these embodiments provide an effective method for accurately processing the voice information of both parties and generating the first interactive voice information. In Example 1, by adjusting the first voice information to be translated according to the content load of the second voice information to be translated, it is possible to focus on the key digital related parts in the voice content of both parties, thereby improving the efficiency and accuracy of subsequent processing. This helps to quickly grasp the key digital content of the differences or concerns of both parties in business exchanges, avoid wasting energy on irrelevant digital information, and thus more efficiently promote the progress of business discussions. Example 2 adjusts the second voice information to be translated based on the first voice information to be translated, so that the voice content of both parties can better focus on the important digital content in the first voice information to be translated during the interaction, which is conducive to the complete presentation of the digital relationship in business exchanges. This method can ensure that in a business meeting, for the comprehensive and complex project planning proposed by one party, the response of the other party can be closely associated with the key digital information, so that both parties can comprehensively and systematically evaluate various digital factors in business cooperation. Example 3 is optimized based on setting feature fine-grained weights, and can accurately process the digital content in the voices of both parties according to the key factors in business decisions, so that the first interactive voice information accurately reflects the digital interaction between the two parties on important business factors.

[0053] In addition, when the priority of the first conversation scenario is lower than that of the second conversation scenario, the generation of the first interactive voice information based on the first voice information to be translated and the second voice information to be translated in step 130 can be implemented by any one of the following embodiments 4 to 6.

[0054] Embodiment 4: The first voice information to be translated is used as the first voice adjustment information; based on the content load corresponding to the first voice information to be translated, the second voice information to be translated is subjected to a first feature adjustment to obtain the second voice adjustment information, wherein the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated; the first voice adjustment information and the second voice adjustment information are subjected to content interaction to obtain the first interactive voice information.

[0055] Embodiment 5: Based on the content load corresponding to the second voice information to be translated, a second feature adjustment is performed on the first voice information to be translated to obtain first voice adjustment information, wherein there is a mapping relationship between the audio text in the first voice adjustment information and the audio text in the second voice information to be translated; the second voice information to be translated is used as the second voice adjustment information; content interaction is performed on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information.

[0056] Embodiment 6: Based on the set feature fine-grained weights, the first voice information to be translated is optimized in feature fine-grained manner to obtain first voice adjustment information; based on the set feature fine-grained weights, the second voice information to be translated is optimized in feature fine-grained manner to obtain second voice adjustment information; the first voice adjustment information and the second voice adjustment information are content-interacted to obtain the first interactive voice information.

[0057] In a business meeting scenario between multinational companies, when the priority of the first conversation scenario is lower than that of the second conversation scenario, the technical solution of generating the first interactive voice information in step 130 is of great significance.

[0058] In Example 4, the first voice information to be translated is used as the first voice adjustment information, and then the second voice information to be translated is adjusted by the first feature according to the content load corresponding to the first voice information to be translated to obtain the second voice adjustment information. For example, in a business meeting, the first voice information to be translated may be a preliminary and relatively concise expression of the views of a representative of one enterprise on a certain project, such as "We think that in terms of marketing promotion of this project, an initial investment of 3 million US dollars in advertising is sufficient, and the market share is expected to increase by 5% within half a year." The second voice information to be translated is a more detailed project plan of the representative of the other enterprise, containing a lot of digital information, such as "We have formulated a comprehensive plan for this project, with a total budget of 10 million US dollars, including a marketing budget of 5 million US dollars. Of the 5 million US dollars, 2 million US dollars are for advertising, 1.5 million US dollars for offline activities, 1 million US dollars for online marketing, and 500,000 US dollars for market research. In terms of personnel arrangement, it is planned to recruit 30 professionals to participate in project promotion, and it is expected to increase the market share from the current 10% to 20% within one year." Since the content load of the first voice information to be translated is relatively small, when the first feature adjustment is made to the second voice information to be translated, it will be adjusted according to the amount of advertising and the expected increase in market share in the first voice information to be translated. For example, focus on the comparative relationship between the advertising part and 3 million US dollars in the second voice information to be translated, and the correlation between the increase in market share and 5%, so as to obtain the second voice adjustment information. At this time, the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated, and then the content of the first voice adjustment information and the second voice adjustment information is interacted. This interaction will integrate the digital relationship between the two parties' capital investment and market share increase in project promotion, and present it in detail in the first interactive voice information, such as the comparison between the initial investment of one party and the advertising budget in the comprehensive plan of the other party, and the relationship between the increase in market share under different expectations.

[0059] In Example 5, the first voice information to be translated is adjusted by the second feature according to the content load corresponding to the second voice information to be translated to obtain the first voice adjustment information, and then the second voice information to be translated is used as the second voice adjustment information, and then the content interaction is performed to obtain the first interactive voice information. Take a business meeting as an example, for example, the second voice information to be translated is a detailed report on a new business development project by a representative of one enterprise, and the content includes "This new business development project is expected to have a total investment of 15 million US dollars and will be carried out in three phases. The first phase will invest 3 million US dollars in market research and team building, which is expected to last for 3 months; the second phase will invest 8 million US dollars in business development and marketing, which is expected to last for 6 months, of which marketing includes 4 million US dollars in advertising, 2 million US dollars in channel expansion and 2 million US dollars in promotional activities; the third phase will invest 4 million US dollars in project optimization and consolidation, which is expected to last for 3 months. And we expect that within one year after the entire project is launched, the market share of the new business will reach 15% and the profit will reach 3 million US dollars." The first voice information to be translated is a brief feedback from the representative of the other enterprise, such as "I think the investment amount is a bit high, and the profit expectation may be too optimistic." According to the content load of the second voice information to be translated, when the second feature adjustment is made to the first voice information to be translated, it will be combined with the specific investment amount and profit expectations in the second voice information to be translated. For example, for a total investment of 15 million US dollars, the first voice information to be translated is adjusted to explain in detail the specific reasons why the investment is too high. It may be based on the average investment level of similar projects in the market of 10 million US dollars, or based on the assessment of its own resources and market risks, it is felt that 12 million US dollars is a more reasonable upper limit; for a profit expectation of 3 million US dollars, it may be pointed out that according to the current market situation and competition, it is more reasonable to expect a profit of about 2 million US dollars. After obtaining the first voice adjustment information, its audio text is mapped to the audio text in the second voice information to be translated. Then, the content of the two voice adjustment information is interacted, and the first interactive voice information will fully present the numerical relationship between the two parties regarding the investment amount and profit expectations, such as the difference in the investment amount of the two parties and their respective basis, the difference in profit expectations and the market considerations behind them.

[0060] In Example 6, based on setting feature fine-grained weights, the first voice information to be translated is optimized in feature fine-grained manner to obtain first voice adjustment information, and at the same time, the second voice information to be translated is optimized in feature fine-grained manner according to the set feature fine-grained weights to obtain second voice adjustment information, and then content interaction is performed to obtain first interactive voice information. For example, in a business meeting, setting feature fine-grained weights is related to key factors in business decision-making, such as cost-effectiveness, market potential, and risk control. For example, the first voice information to be translated is some preliminary ideas of a corporate representative on a cooperation project, including "We are considering cooperating on this project and hope that the cost will be controlled within 8 million US dollars. Through this project, we can tap at least 30% of new market potential, and the risk factor will be kept below 20%." The second voice message to be translated is a more detailed project analysis by the representative of the other company. The content is as follows: "After a detailed analysis of this project, we believe that the cost may reach 10 million US dollars, because the raw material cost is estimated to be 5 million US dollars, the labor cost is 3 million US dollars, and other costs are 2 million US dollars; in terms of market potential, although there are certain opportunities, according to market research and competitor conditions, it is expected that about 20% of new market potential can be tapped; in terms of risk, due to market fluctuations and policy uncertainties, the risk factor may reach 30%." If the cost-effectiveness weight is set higher in the feature fine-grained weight, then when the first voice message to be translated is optimized, the specific measures to control the cost within 8 million US dollars will be deeply analyzed, such as reducing the cost of raw materials by 1 million US dollars by finding new suppliers, and reducing the labor cost by 500,000 US dollars by optimizing the process, etc., to obtain the first voice adjustment information. When the second voice message to be translated is optimized, the rationality of the cost composition of 10 million US dollars will also be elaborated in detail around the cost part, such as why the raw material cost is estimated to be 5 million US dollars based on the current market price and quality requirements, etc., to obtain the second voice adjustment information. Then, the two voice adjustment messages will interact with each other. The first interactive voice message will accurately present the digital interactive content of both parties on important business factors such as cost-effectiveness, market potential and risk control, such as the differences and reasons for the cost forecasts of both parties, the basis for different expectations of market potential and different assessments of risk factors.

[0061] It can be seen that Example 4 adjusts the second voice information to be translated based on the first voice information to be translated, and can focus on the key content in the second voice information based on the relatively concise first voice information, so that the first interactive voice information after the interaction accurately presents the digital relationship between the two parties, thereby improving the pertinence of the communication. Example 5 adjusts the first voice information to be translated based on the second voice information to be translated, and can effectively combine brief feedback with detailed reports, and fully present the differences and considerations of the two parties in terms of investment, profit and other numbers. Example 6 is based on weight optimization, and can accurately process the key business digital content in the voices of both parties, so that the first interactive voice information fully reflects the digital interaction between the two parties on important factors.

[0062] On the basis of Example 2 to Example 6, the content interaction of the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information includes: obtaining a first entity relationship network corresponding to the first voice adjustment information; obtaining a second entity relationship network corresponding to the second voice adjustment information; performing mean processing on entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain a global entity feature value corresponding to each entity distribution label; and generating the first interactive voice information based on the global entity feature value corresponding to each entity distribution label.

[0063] In some other embodiments, on the basis of Embodiment 2 to Embodiment 6, the content interaction of the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information can also be achieved through the following technical steps: obtaining the first entity relationship network corresponding to the first voice adjustment information; obtaining the second entity relationship network corresponding to the second voice adjustment information; mapping the entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain the entity mapping feature value corresponding to each entity distribution label; generating the first interactive voice information based on the entity mapping feature value corresponding to each entity distribution label.

[0064] In the business meeting scenario between multinational companies, the technical solution for content interaction further described based on the above embodiments 2-6 to obtain the first interactive voice information is of great significance.

[0065] First, according to the technical solution mentioned above, in the process of obtaining the first entity relationship network corresponding to the first voice adjustment information and the second entity relationship network corresponding to the second voice adjustment information, the entity relationship network contains many important elements related to the business meeting. Taking a business meeting of a multinational company on product R&D cooperation as an example, the first voice adjustment information may be the adjusted information of one company on the product R&D investment and expected benefits. For example, the content of the first voice adjustment information contains information about R&D investment, a planned investment of US$5 million in R&D equipment procurement, an estimated expenditure of US$2 million on human resources, and an expected first-year revenue of US$10 million after the product is launched. Then the corresponding first entity relationship network will be built around these elements, such as R&D equipment procurement, human resources investment and expected benefits will become the key entities in this relationship network. There are interrelated relationships between these entities, such as the sum of R&D equipment procurement and human resources investment will affect the expected benefits, and this influence relationship will also be reflected in the entity relationship network.

[0066] Similarly, for the second voice adjustment information, if it is the adjustment information of the other party's enterprise on the product development cycle and market share, for example, the expected development cycle is 2 years, the market share target is 20% in the first year after the product is launched, and it will be increased to 30% in the second year, etc. The corresponding second entity relationship network will be constructed with the development cycle, market share, etc. as key entities, and the length of the development cycle may affect the growth rate of the market share, and this relationship will also be reflected in the relationship network.

[0067] Next, when the entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network are averaged to obtain the global entity feature value corresponding to each entity distribution label, the above example is continued to be used for explanation. For example, there are the same entity distribution labels in the two entity relationship networks, such as the entity distribution label "R&D investment". In the first entity relationship network, the R&D investment (entity feature value) is the sum of 5 million US dollars for equipment purchase and 2 million US dollars for manpower investment, which is 7 million US dollars; in the second entity relationship network, the R&D investment may be measured from different angles, such as the division according to the R&D stage, the initial R&D investment is 3 million US dollars, the mid-term investment is 3 million US dollars, and the total is 6 million US dollars. The two entity feature values ​​of "R&D investment" are averaged, that is, (700+600) / 2=6.5 million US dollars, and this 6.5 million US dollars is the global entity feature value corresponding to the entity distribution label "R&D investment". For other possible identical entity distribution labels, such as "expected benefits", etc., they are also processed in the same way. Then, based on the global entity feature value corresponding to each entity distribution label, the first interactive voice information is generated. This first interactive voice message will include entity information that has been processed by averaging, such as "R&D investment is comprehensively assessed to be US$6.5 million, and the expected return (the value after averaging with other relevant information) is...", etc. This will integrate the different information about product development of both parties from their respective entity relationship networks to form a first interactive voice message that integrates the views and data of both parties.

[0068] In some other embodiments, after obtaining the first entity relationship network corresponding to the first voice adjustment information and the second entity relationship network corresponding to the second voice adjustment information, the entity mapping feature value corresponding to each entity distribution label is obtained by mapping the entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network, and then the technical solution of generating the first interactive voice information is based on these values. Taking the business meeting of multinational enterprise product R&D cooperation as an example, for the same entity distribution label "R&D investment", the R&D investment in the first entity relationship network is expressed as 7 million US dollars in terms of equipment procurement and manpower investment, and the R&D investment in the second entity relationship network is expressed as 6 million US dollars in terms of R&D stage. The mapping process here may be based on a pre-set rule, such as mapping according to the investment ratio. If the R&D investment ratio in the first entity relationship network is set to 70% (US$3.5 million) for equipment procurement and 30% (US$1.5 million) for manpower investment, the early R&D investment in the second entity relationship network accounts for 50% (US$3 million) and the mid-term investment accounts for 50% (US$3 million). Then after the mapping, the mapping of the entity feature value of "equipment procurement" may be adjusted according to its proportional relationship in the respective inputs to obtain a mapped feature value. Similar mapping operations are also performed on the entity feature values ​​under the same entity distribution labels such as "human resource investment" and "R&D stage investment" to obtain the entity mapping feature value corresponding to each entity distribution label. The first interactive voice information is generated based on these entity mapping feature values, such as "The mapped value of equipment procurement in R&D investment is..., and the mapped value of human resource investment is...", etc., so as to integrate the R&D investment information of both parties into the first interactive voice information in a mapped manner.

[0069] This design can effectively integrate the different views and data of both parties in the business meeting by constructing an entity relationship network and performing mean processing or mapping processing on the entity feature values ​​of the same entity distribution label. Mean processing can combine the values ​​of both parties in various entities to provide a relatively balanced comprehensive result, which helps the two parties reach a consensus on cooperation and other aspects. Mapping processing converts and associates the values ​​of both parties according to specific rules, so that the entity feature values ​​under different representations can be integrated in a more reasonable way. Whether it is R&D investment, market share or expected returns, they can be accurately and reasonably reflected in the first interactive voice information, thereby improving the accuracy and efficiency of consecutive interpretation.

[0070] In an optional embodiment, when the priority of the first conversation scenario is the same as the priority of the second conversation scenario; the generating of the first interactive voice information based on the first voice information to be translated and the second voice information to be translated in step 130 includes: obtaining each audio text unit that is associated with the first voice information to be translated and the second voice information to be translated; sorting the associated audio text units, and retaining each audio text unit that is not associated with the first voice information to be translated and the second voice information to be translated, to obtain the first interactive voice information.

[0071] In a business meeting scenario between multinational companies, when the priority of the first conversation scenario is the same as the priority of the second conversation scenario, the technical solution of generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated in step 130 is of great significance.

[0072] In a business meeting, the first voice message to be translated and the second voice message to be translated contain rich and diverse content. For example, the first voice message to be translated may be an explanation of a cooperation project by a representative of one enterprise, which mentions the investment budget, market positioning, expected returns and other aspects of the project. For example, it is mentioned that the investment budget is 8 million US dollars, the market positioning is for high-end consumer groups, and the expected return is 15 million US dollars within two years after the project is launched. The second voice message to be translated is the response of the representative of the other enterprise to the project, which may include an assessment of project risks, plans for its own resource investment, and expectations for market share. For example, it is believed that the project risk mainly lies in the fierce market competition, which may lead to a 20% increase in costs. The resources that can be invested include equipment worth 3 million US dollars and human resources worth 1 million US dollars, and the market share in the project is expected to reach 10% in the first year and increase to 15% in the second year.

[0073] The process of obtaining each audio text unit that is associated with the first voice information to be translated and the second voice information to be translated is very critical. For the above example, the investment budget of 8 million US dollars in the first voice information to be translated is associated with the possible cost increase of 20% in the second voice information to be translated, because the cost increase will affect the initial investment budget; the expected return in the first voice information to be translated is also associated with the market share in the second voice information to be translated, because the size of the market share will directly affect the realization of the expected return. These associated audio text units need to be accurately obtained. At the same time, the market positioning of the first voice information to be translated, which is oriented towards high-end consumer groups, may also be associated with the own resource investment (such as high-end equipment) in the second voice information to be translated, which shows the relationship between resource investment and market positioning.

[0074] The operation of organizing the related audio text units also has its own unique connotations. In the context of a business meeting, we will continue to use the above example to illustrate. For the related part of investment budget and cost increase, the investment budget of $8 million should be organized with the possible increase of 20% in cost (i.e., the increase of $1.6 million). This may mean re-evaluating the allocation of investment budget under the condition of cost increase. For example, the budget originally used for marketing promotion may need to be cut to cope with the increase in cost. For the related part of expected revenue and market share, the rationality of the expected revenue of $15 million should be re-examined based on the situation that the market share reaches 10% in the first year and increases to 15% in the second year. If the market share growth rate is not as expected, the expected revenue may need to be adjusted accordingly, which requires reasonable organization of this part of the audio text units. For the related part of market positioning and resource investment, based on the market positioning as a high-end consumer group and the high-end equipment investment worth $3 million, it is necessary to further clarify how the resource investment can better serve the market positioning, such as whether it is necessary to increase the investment in high-end equipment maintenance and upgrades.

[0075] Moreover, in this process, the various audio text units that are not related to the first voice information to be translated and the second voice information to be translated must be retained. For example, although the market positioning for high-end consumer groups in the first voice information to be translated is related to part of the content in the second voice information to be translated, it is an independent and important information that needs to be retained; similarly, the human resources invested in the second voice information to be translated, which is $1 million, also needs to be retained if it is not related to other content in the first voice information to be translated. These retained audio text units also play an important role in constructing the first interactive voice information. They can fully reflect the content of the speeches of both parties in the business meeting.

[0076] The first interactive voice message finally obtained is a complete message that integrates the related and non-related parts of the voice messages of both parties. This first interactive voice message includes not only the adjustment information of related parts such as investment budget and cost increase, the information after the expected return and market share are re-evaluated, and the information after the market positioning and resource investment optimization, but also includes independent information such as market positioning itself and specific human resource investment. This enables the first interactive voice message to fully and accurately reflect the content of the communication between the two parties in the business meeting, and provides a complete basis for further communication, cooperation decision-making, etc.

[0077] In this way, by accurately acquiring and organizing the associated audio text units, it is possible to deeply analyze the internal connections between the various contents of the business meeting between the two parties, such as the relationship between investment and cost, market share and revenue, etc., which helps both parties to better understand each other's views and intentions. At the same time, retaining non-associated audio text units ensures the integrity of the information, so that all important information of the two parties in the communication process can be included in the first interactive voice information, thereby improving the accuracy and comprehensiveness of consecutive interpretation.

[0078] In a preferred embodiment, performing semantic vector mining on the first interactive voice information to obtain a first semantic interaction vector includes: sending the first interactive voice information to a large language model system so that the large language model system performs semantic vector mining on the first interactive voice information to obtain the first semantic interaction vector.

[0079] In detail, the sending of the first interactive voice information to the large language model system so that the large language model system performs semantic vector mining on the first interactive voice information to obtain the first semantic interaction vector includes: performing voice enhancement and context feature annotation on the first interactive voice information to obtain target interactive voice information; performing semantic upsampling on the target interactive voice information to obtain upsampled semantic information; sending the upsampled semantic information to the large language model system so that the large language model system recognizes the upsampled semantic information to obtain the target interactive voice information, and performing semantic vector mining on the target interactive voice information to obtain the first semantic interaction vector.

[0080] Furthermore, the method of using the pre-stored multiple benchmark semantic interaction vectors and the digital annotation indication strategy corresponding to each benchmark semantic interaction vector to perform digital consecutive interpretation on the first semantic interaction vector to obtain the consecutive interpretation annotated text includes: receiving the target digital annotation indication strategy corresponding to the first semantic interaction vector sent by the large language model system, wherein the target digital annotation indication strategy is obtained by the large language model system after performing digital consecutive interpretation on the first semantic interaction vector using the pre-stored multiple benchmark semantic interaction vectors and the digital annotation indication strategy corresponding to each benchmark semantic interaction vector.

[0081] In the business meeting scenario between multinational companies, the technical solutions in the above embodiments have profound significance and importance.

[0082] In business meetings, the first interactive voice message carries the key content of the communication between the two parties. This first interactive voice message contains many elements related to business decisions, such as information about project investment scale, market share, cost budget, expected returns, etc. This information often contains a large amount of digital content. For digital consecutive interpreters, if they only rely on traditional methods to process it, they will face a huge cognitive load.

[0083] First of all, it is very critical to perform voice enhancement and context feature annotation on the first interactive voice information to obtain the target interactive voice information. In the business meeting scenario, each part of the first interactive voice information has a specific meaning and association. Taking a business meeting about multinational companies cooperating to develop new products as an example, the first interactive voice information mentioned that one party plans to invest 50 million US dollars in the research and development of new products, and the market share is expected to reach 15% in the first year after the product is launched on the market. The cost budget is controlled within 30 million US dollars, and the expected profit is 80 million US dollars. When performing voice enhancement, these key figures and related expressions will be highlighted, such as emphasizing the investment of 50 million US dollars, 15% market share, 30 million US dollars cost budget and 80 million US dollars expected profit. At the same time, contextual feature annotation is performed, for example, the R&D investment of 50 million US dollars is annotated as contextual information related to the innovation degree and R&D cycle of new products; the market share of 15% is annotated as related to market competition conditions, product advantages, etc.; the cost budget of 30 million US dollars is annotated as related to raw material procurement, labor costs, etc.; the expected profit of 80 million US dollars is annotated as related to market pricing, sales volume, etc. Through such speech enhancement and contextual feature annotation, the target interactive speech information is obtained, so that this information can be more accurately identified and processed in subsequent processing.

[0084] Next, the target interactive voice information is semantically upsampled to obtain upsampled semantic information. In the complex context of a business meeting, semantic upsampling helps to more comprehensively explore the semantic connotations in the information. Continuing with the above example, the target interactive voice information after voice enhancement and context feature annotation may be linked to the long-term strategic planning of the enterprise through semantic upsampling. For example, the enterprise may hope to improve its technological leadership in related fields in the next five years through this $50 million R&D investment. The semantic information related to this long-term strategic planning may be implicit in the original target interactive voice information, and it is made clear through semantic upsampling. For market share, the 15% market share may be linked to the development goals of a specific regional market in the enterprise's global market layout. For example, in a specific regional market, this 15% market share is a key step to open up a larger market. This broader semantic connection is reflected through upsampling. Similar semantic upsampling is also performed for cost budget and expected benefits, linking them with the overall financial planning, market competition strategy and other macro levels of the enterprise. The upsampled semantic information obtained in this way contains richer and more comprehensive semantic content, providing more sufficient information for the accurate recognition and processing of the large language model system.

[0085] Then, the upsampled semantic information is sent to the large language model system, which identifies the target interactive voice information and mines the semantic vector of the target interactive voice information to obtain the first semantic interaction vector. The large language model system can accurately process these upsampled semantic information by virtue of its ability to learn and analyze large amounts of data. In the business meeting scenario, the large language model system can identify various complex relationships in the upsampled semantic information. For example, it can identify the causal relationship between R&D investment of $50 million and the expected improvement of technological leadership, the progressive relationship between 15% market share and opening up a specific regional market, the constraint relationship between the cost budget of $30 million and the overall financial planning, and the target relationship between the expected profit of $80 million and the market competition strategy. By analyzing and mining these relationships, the large language model system can convert the target interactive voice information into a semantic vector form, that is, obtain the first semantic interaction vector. This first semantic interaction vector can accurately represent the semantic relationship between the communication content of all parties in the business meeting, especially the semantic relationship related to numbers, such as the proportional relationship between the R&D investment amount and the expected profit amount, the correlation between the market share and the cost budget, etc.

[0086] Finally, the first semantic interaction vector is digitally interpreted consecutively using the pre-stored multiple benchmark semantic interaction vectors and the digital annotation indication strategy corresponding to each benchmark semantic interaction vector to obtain the consecutive interpretation annotated text. The large language model system receives the first semantic interaction vector and processes it according to the pre-stored multiple benchmark semantic interaction vectors and the corresponding digital annotation indication strategy. In the business meeting scenario, these benchmark semantic interaction vectors and digital annotation indication strategies are constructed based on a large number of business communication examples. For example, there is a benchmark semantic interaction vector representing the relationship between investment and return. When the relationship between the R&D investment of 50 million US dollars and the expected return of 80 million US dollars in the first semantic interaction vector matches it, the corresponding digital annotation indication strategy may be to annotate this relationship as "investment of 50 million US dollars and expected return of 80 million US dollars, and the return is 1.6 times the investment". For the relationship between market share and cost budget, if there is a corresponding benchmark semantic interaction vector and digital annotation indication strategy, it may be annotated as "the cost budget corresponding to 15% of the market share is 30 million US dollars, and there is a certain balance between the cost budget and the market share", etc. In this way, a consecutive interpretation annotated text is obtained, which accurately reflects the numerical relationship in the communication content of the parties in the business meeting and is annotated in a way that conforms to the target language and business practices.

[0087] The above-mentioned embodiments have a significant effect on reducing the cognitive load of digital consecutive interpreters facing the translation of large amounts of digital information and improving the accuracy and efficiency of digital consecutive interpretation. During the processing, by performing voice enhancement, context feature annotation, and semantic upsampling on the first interactive voice information, the large language model system can more comprehensively and accurately process the semantics and digital relationships therein, reducing the burden on the interpreter to understand complex semantics and digital relationships. With its powerful processing capabilities, the large language model system accurately performs semantic vector mining and digital consecutive interpretation, avoiding the possible digital understanding errors and semantic relationship misjudgments that may occur during manual processing by the interpreter, and improving accuracy. At the same time, the automated processing flow saves the interpreter a lot of time and improves the efficiency of digital consecutive interpretation, thereby contributing to the smooth progress of business meetings of multinational companies.

[0088] In an alternative design idea, after generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated, the method further includes: if the first interactive voice information does not trigger the digital consecutive interpretation mechanism, collecting the third voice information to be translated corresponding to the third conversation scene through the voice input end, wherein the third voice information to be translated includes G groups of audio texts of the target conversation event, the third conversation scene represents the interpretation task state between the target conversation event and the voice input end in the third consecutive interpretation cycle, and the third consecutive interpretation cycle is the next consecutive interpretation cycle of the second consecutive interpretation cycle. , G is an integer not less than 1; based on the first interactive voice information and the third voice information to be translated, second interactive voice information is generated, wherein the second interactive voice information includes the digital interactive text of the N groups of audio texts, the M groups of audio texts and the G groups of audio texts; if the second interactive voice information triggers the digital consecutive interpretation mechanism, semantic vector mining is performed on the second interactive voice information to obtain a second semantic interaction vector; using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector, digital consecutive interpretation is performed on the second semantic interaction vector to obtain a consecutive interpretation annotation text.

[0089] This technical solution is of great significance in business meeting scenarios between multinational companies.

[0090] During the business meeting, the first voice information to be translated and the second voice information to be translated are used to generate the first interactive voice information. However, if the first interactive voice information does not trigger the digital consecutive interpretation mechanism, it means that the content of the first interactive voice information may not meet the requirements for digital consecutive interpretation at this time, which may be due to unclear digital relationships or missing key digital information.

[0091] For example, in a business meeting, the first voice message to be translated is the preliminary project plan mentioned by a corporate representative on one side, which contains some unclear numerical information, such as "the project will probably require a certain amount of capital investment, there will be several stages, and the manpower investment in each stage will probably be some people." The expressions such as "certain," "several," and "some" here make the numbers unclear. The second voice message to be translated is the response of the other corporate representative, which is also similarly vague, such as "I think the amount of investment should have a range, and the manpower investment also needs to be adjusted according to market conditions." The first interactive voice messages generated by the two may not have a clear enough numerical relationship to trigger the digital consecutive interpretation mechanism.

[0092] At this time, the third voice information to be translated corresponding to the third conversation scenario is collected through the voice input terminal. For example, the third voice information to be translated is another participant's more specific plan for the project, which mentions that "the total investment of the project is expected to be around 8 million US dollars, divided into three phases, the first phase will invest 2 million US dollars, 20 people; the second phase will invest 3 million US dollars, 30 people; the third phase will invest 3 million US dollars, 25 people, and the expected market share will reach 10% in the first year and 15% in the second year" and other content containing clear numbers.

[0093] The second interactive voice message is generated based on the first interactive voice message and the third voice message to be translated. This process is to integrate the previous relatively vague interactive information with the new information containing clear numbers. For example, in the second interactive voice message, the previous vague discussion about project planning will be integrated with the new $8 million investment, investment and manpower input in each stage, market share expectations and other digital information to form a digital interactive text containing N groups of audio texts, M groups of audio texts and G groups of audio texts.

[0094] If the second interactive voice information triggers the digital consecutive interpretation mechanism, the semantic vector mining of the second interactive voice information is performed to obtain the second semantic interactive vector. In the context of a business meeting, the large language model system will mine the semantic relationship in the second interactive voice information. For example, the relationship between the investment of 8 million US dollars and the investment, manpower input, and market share expectations at each stage is analyzed. The total investment of 8 million US dollars and the investment allocation of 2 million, 3 million, and 3 million at each stage are the relationship between the whole and the part; there is a resource allocation relationship between the investment amount and the amount of manpower input; and investment and manpower input have an impact relationship with market share expectations. Through such mining, these relationships are converted into semantic vectors to obtain the second semantic interaction vector.

[0095] Then, the second semantic interaction vector is digitally interpreted consecutively using the pre-stored multiple benchmark semantic interaction vectors and the digital annotation indication strategy corresponding to each benchmark semantic interaction vector to obtain the consecutive interpretation annotation text. For example, there is a vector representing the relationship between investment and market share in the pre-stored benchmark semantic interaction vector. When the $8 million investment in the second semantic interaction vector matches the expected relationship of 10%-15% market share, the corresponding digital annotation indication strategy may annotate this relationship as "under an investment of $8 million, the expected market share is 10% in the first year and 15% in the second year, and the investment and market share growth are positively correlated", etc., thereby obtaining accurate consecutive interpretation annotation text.

[0096] This helps to reduce the cognitive load of digital consecutive interpreters who are faced with translating large amounts of digital information. By introducing the third voice information to be translated to generate the second interactive voice information when the first interactive voice information does not trigger the mechanism, the interpreter's burden of analyzing complex and ambiguous digital information is reduced. The accuracy of digital consecutive interpretation is improved, and the semantic vectors are mined with the help of a large language model system and annotated according to pre-stored strategies, avoiding errors in manual processing of digital relationships. At the same time, efficiency is improved and the time it takes for interpreters to process uncertain digital information is reduced.

[0097] In an independent embodiment, after generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated, the method also includes: obtaining a target entity relationship network corresponding to the first interactive voice information; obtaining entity feature values ​​corresponding to each audio text unit in the target entity relationship network; determining a global entity feature value of the target entity relationship network based on the entity feature values ​​corresponding to each audio text unit; calculating an entity feature discreteness based on the entity feature values ​​corresponding to each audio text unit and the global entity feature value; if the entity feature discreteness is not less than a preset entity feature discreteness, determining that the first interactive voice information triggers the digital consecutive interpretation mechanism.

[0098] In another independent embodiment, after generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated, the method further includes: obtaining a target entity relationship network corresponding to the first interactive voice information; obtaining entity feature values ​​corresponding to each audio text unit in the target entity relationship network; determining a global entity feature value of the target entity relationship network based on the entity feature values ​​corresponding to each audio text unit; calculating an entity feature discreteness based on the entity feature values ​​corresponding to each audio text unit and the global entity feature value; if the entity feature discreteness is not less than a preset entity feature discreteness, and the key digital audio text and conversation event keywords in the first interactive voice information are identified, then it is determined that the first interactive voice information triggers the digital consecutive interpretation mechanism.

[0099] Based on the above two independent embodiments, the method also includes: if the first interactive voice information triggers the digital consecutive interpretation mechanism, outputting a first judgment result, wherein the first judgment result is used to prompt the triggering of the digital consecutive interpretation mechanism; if the first interactive voice information does not trigger the digital consecutive interpretation mechanism, outputting a second judgment result, wherein the second judgment result is used to prompt that the digital consecutive interpretation mechanism is not triggered.

[0100] In the business meeting scenario between multinational companies, the technical solutions involved in these embodiments play a vital role in the entire digital consecutive interpretation process, which will be explained in detail below.

[0101] In business meetings, the information exchanged by representatives of all parties is complex and includes a lot of project planning, market analysis, financial budget, etc. These contents will be reflected in the first voice message to be translated and the second voice message to be translated. After the first interactive voice message is generated based on the two, the subsequent processing mechanism begins to play a role.

[0102] First, the process of obtaining the target entity relationship network corresponding to the first interactive voice information requires a deep understanding of the various information elements and their interrelationships in the business meeting. For example, in a business meeting about multinational companies cooperating to develop new products, the first interactive voice information may include one company's explanation of product R&D investment, such as the source, allocation and expected return of R&D funds; it also includes the other company's analysis of market positioning, target customer groups and potential market size. These different aspects of information constitute the nodes in the target entity relationship network. For example, there are potential connections between the various parts of R&D funds (equipment procurement funds, labor costs, R&D site rental, etc.) and the different elements of market positioning (consumption capacity of target customer groups, geographical scope of the market, distribution of competitors, etc.), and these connections weave the entire target entity relationship network.

[0103] Then, obtaining the entity feature value corresponding to each audio text unit in the target entity relationship network is a key step. In the above business meeting example, for R&D investment, if the equipment procurement funds are mentioned as 3 million US dollars, then this 3 million is an entity feature value; the labor cost is 1.5 million US dollars, and this 1.5 million is also an entity feature value; if the first-year sales of the new product after its launch are expected to be 10 million US dollars, this 10 million is also an entity feature value. For market positioning, if the target customer group is mainly middle and high-end consumers, the quantitative assessment of their consumption capacity (for example, the average consumption amount is 500 US dollars / time) can also be regarded as an entity feature value; if the potential market size is expected to be 10 million potential customers, this 10 million is also an entity feature value. These entity feature values ​​are quantitative representations of various key information in business meetings, and they reflect the importance and influence of different elements in the entire business activities.

[0104] Determining the global entity feature value of the target entity network based on the entity feature values ​​corresponding to these individual audio text units is a more complex process. This requires comprehensive consideration of the weights and interrelationships of the various entity feature values ​​in the entire business relationship. For example, when evaluating the overall feasibility of a new product development project, R&D investment may have a relatively large weight because it is directly related to whether the product can be successfully developed and launched on the market. For example, the weight of R&D investment is 0.4, the weight of market positioning is 0.3, and the weight of other related factors (such as marketing promotion, etc.) is 0.3. For the R&D investment part, if the weight of equipment procurement funds of $3 million accounts for 0.5 in R&D investment, the weight of labor costs of $1.5 million accounts for 0.3, and the weight of other R&D costs of $500,000 accounts for 0.2, then the contribution of the R&D investment part to the global entity feature value is (300×0.5+150×0.3+50×0.2)×0.4. Similarly, the contribution of the market positioning part and the other related factors part to the global entity feature value is calculated, and then these contributions are added together to obtain the global entity feature value of the target entity network.

[0105] Next, based on the entity feature values ​​corresponding to each audio text unit and the global entity feature value, the entity feature dispersion is calculated. The calculation of dispersion is intended to measure the degree of dispersion of each entity feature value relative to the global entity feature value. In the context of a business meeting, this degree of dispersion reflects the imbalance or complexity between different elements in business activities. Taking the new product development project just now as an example, if the numerical differences in equipment procurement funds, labor costs and other R&D costs in R&D investment are large, then a relatively large value will be obtained when calculating the dispersion. For example, the equipment procurement funds are 3 million US dollars, the labor costs are 1.5 million US dollars, and the other R&D costs are 500,000 US dollars. The global entity feature value is calculated to be 2 million US dollars. Through specific dispersion calculation methods (such as variance calculation), it will be found that these values ​​have a large dispersion relative to the global value of 2 million US dollars.

[0106] In the first independent embodiment, if the entity feature discreteness is not less than the preset entity feature discreteness, it is determined that the first interactive voice information triggers the digital consecutive interpretation mechanism. The preset entity feature discreteness is a threshold value pre-set according to factors such as the nature of the business meeting, the needs of the participants, and past experience. For example, in a business cooperation negotiation meeting that requires extremely high digital accuracy and involves complex financial analysis, this preset discreteness threshold may be relatively low, because even small digital differences or imbalances may have a significant impact on business decisions. In a more preliminary project negotiation meeting, this threshold may be relatively high. When the entity feature discreteness exceeds this threshold, it indicates that the digital relationship in the first interactive voice information is sufficiently complex or unbalanced, and digital consecutive interpretation is required to accurately handle these relationships. For example, if the investment budgets for different regional markets in a business meeting vary greatly, or the ratio between expected returns and input costs varies greatly in different business segments, this large discreteness requires digital consecutive interpretation to sort out and accurately convey the digital information therein.

[0107] In the second independent embodiment, in addition to the condition that the entity feature discreteness is not less than the preset entity feature discreteness, it is also necessary to identify the key digital audio text and the conversation event theme words in the first interactive voice information to determine that the first interactive voice information triggers the digital consecutive interpretation mechanism. In the business meeting scenario, the key digital audio text is often the digital information that has a core influence on business decisions. For example, in corporate mergers and acquisitions negotiations, the target company's valuation, debt-to-asset ratio, and expected merger premium are key digital audio texts; in market cooperation projects, market share targets, expected market growth rates, and investment ratios of both parties are also key digital audio texts. Conversation event theme words reflect the core topics of business meetings, such as "corporate mergers and acquisitions", "market expansion cooperation", "new product research and development cooperation", etc. Only when there is a large entity feature discreteness and these key digital audio texts and conversation event theme words can be clearly identified, can it be determined that digital consecutive interpretation is required. This is because the existence of these key numbers and theme words further demonstrates the importance and complexity of digital relationships, and they, together with the entity feature discreteness, constitute the complete basis for judging whether digital consecutive interpretation is needed. For example, in a business meeting about market expansion cooperation, if only some general market research data are mentioned, although there may be a certain degree of entity feature dispersion, if there are no clear key digital audio texts such as target market share, cooperation investment amount, and conversation event keywords such as "market expansion cooperation", digital simultaneous interpretation may not be required.

[0108] If the first interactive voice information triggers the digital consecutive interpretation mechanism, the first discrimination result is output, wherein the first discrimination result is used to prompt the triggering of the digital consecutive interpretation mechanism. This discrimination result plays an important guiding role in the digital consecutive interpretation process of the entire business meeting. It can notify the relevant digital processing system or interpreter, indicating that the digital relationship in the first interactive voice information needs to be processed by in-depth consecutive interpretation. For example, in a multi-party multinational business meeting, this discrimination result can be transmitted to a centralized digital processing center or the work terminal of each interpreter to start the digital consecutive interpretation process in time. If the first interactive voice information does not trigger the digital consecutive interpretation mechanism, the second discrimination result is output, wherein the second discrimination result is used to prompt that the digital consecutive interpretation mechanism is not triggered. This helps to avoid unnecessary digital processing operations and reduce the cognitive load of digital consecutive interpreters facing the translation of a large amount of digital information. For example, in the early greetings of a business meeting or the general project background introduction stage, complex digital relationships or key numbers may not be involved. At this time, outputting the second discrimination result can avoid the interpreter from performing unnecessary digital consecutive interpretation processing on these simple information, thereby saving the interpreter's energy and time.

[0109] Designed in this way, the above technical scheme provides an effective solution for reducing the cognitive load of digital consecutive interpreters facing the translation of a large amount of digital information and improving the accuracy and efficiency of digital consecutive interpretation. By judging whether to trigger the digital consecutive interpretation mechanism through the calculation of entity relationship network, entity feature value, discreteness, and identification of key numbers and keywords, it is possible to accurately screen out situations that really require digital consecutive interpretation. This avoids the interpreter's processing of unnecessary digital information and reduces the interpreter's cognitive burden. In terms of accuracy, digital consecutive interpretation is only performed when specific conditions are met, ensuring the accurate processing of complex digital relationships and reducing the possibility of misjudgment. In terms of efficiency, unnecessary operating procedures are avoided, saving the interpreter's time, allowing the interpreter to focus more on the content that really requires digital consecutive interpretation, thereby improving the efficiency of digital consecutive interpretation for the entire business meeting.

[0110] In summary, multiple groups of voice information to be translated in the form of audio texts are collected through the voice input terminal, and the first interactive voice information containing digital interactive texts is generated, thereby realizing the orderly integration of information. When the digital consecutive interpretation mechanism is triggered, semantic vector mining is carried out with the help of a large language model system, which greatly reduces the cognitive load of the interpreter. The interpreter does not need to deeply analyze complex semantic relationships by himself. The use of pre-stored benchmark semantic interaction vectors and digital annotation indication strategies for digital consecutive interpretation improves accuracy and avoids possible errors in digital conversion and semantic understanding caused by manual work. The entire process has a high degree of automation, which improves the efficiency of interpretation and can meet the needs of rapid progress in business communication in a timely manner.

[0111] Based on the same inventive concept, an embodiment of the present application provides a digital consecutive interpretation device, which includes: a voice collection module, which is used to collect first voice information to be translated corresponding to a first conversation scenario through a voice input terminal, wherein the first voice information to be translated includes N groups of audio texts of a target conversation event, and the first conversation scenario represents the interpretation task state between the target conversation event and the voice input terminal in a first consecutive interpretation cycle, and N is an integer not less than 1; collect second voice information to be translated corresponding to a second conversation scenario through the voice input terminal, wherein the second voice information to be translated includes M groups of audio texts of the target conversation event, and the second conversation scenario represents the interpretation task state between the target conversation event and the voice input terminal in a second consecutive interpretation cycle, and the second consecutive interpretation cycle is The next consecutive interpretation cycle of the first consecutive interpretation cycle, wherein M is an integer not less than 1; a speech generation module, used to generate first interactive speech information based on the first speech information to be translated and the second speech information to be translated, wherein the first interactive speech information includes a digital interactive text of the N groups of audio texts and the M groups of audio texts; a semantic mining module, used to perform semantic vector mining on the first interactive speech information to obtain a first semantic interaction vector if the first interactive speech information triggers a digital consecutive interpretation mechanism; wherein the semantic vector mining is implemented by a large language model system; a consecutive interpretation module, used to perform digital consecutive interpretation on the first semantic interaction vector using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector, to obtain a consecutive interpretation annotation text.

[0112] Further, Figure 2 Schematic diagram of the structure of a digital consecutive interpretation device 200 provided in an embodiment of the present application. Figure 2 The digital consecutive interpretation device 200 shown includes a processor 210. The processor 210 can call and run a computer program from a memory to implement the method in the embodiment of the present application.

[0113] Alternatively, if Figure 2 As shown, the digital consecutive interpretation device 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present application.

[0114] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0115] Alternatively, if Figure 2As shown, the digital consecutive interpretation device 200 may further include a transceiver 220. The processor 210 may control the transceiver 220 to interact with other devices. Specifically, the transceiver 220 may send information or data to other devices, or receive information or data sent by other devices.

[0116] Optionally, the digital consecutive interpretation device 200 can implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present application. For the sake of brevity, they are not repeated here.

[0117] It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities.

[0118] It is understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0119] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0120] It should be noted that, in this article, the term "includes", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be noted that the scope of the method and device in the embodiment of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0121] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), including a number of instructions for a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the embodiment of the present application.

[0122] The embodiments of the embodiments of the present application are described above in conjunction with the accompanying drawings, but the embodiments of the present application are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present application and the scope of protection of the embodiments of the present application, all of which are within the protection of the embodiments of the present application.

Claims

1. A digital consecutive interpretation method based on a large language model, characterized in that: The method is applied to a digital consecutive interpretation device, and the method comprises: Collecting first voice information to be translated corresponding to a first conversation scenario through a voice input terminal, wherein the first voice information to be translated includes N groups of audio texts of a target conversation event, the first conversation scenario represents an interpretation task state between the target conversation event and the voice input terminal in a first consecutive interpretation cycle, and N is an integer not less than 1; Collecting second voice information to be translated corresponding to a second conversation scenario through the voice input terminal, wherein the second voice information to be translated includes M groups of audio texts of the target conversation event, the second conversation scenario represents the state of the interpretation task between the target conversation event and the voice input terminal in a second consecutive interpretation cycle, the second consecutive interpretation cycle is the next consecutive interpretation cycle of the first consecutive interpretation cycle, and M is an integer not less than 1; Based on the first voice information to be translated and the second voice information to be translated, first interactive voice information is generated, wherein the first interactive voice information includes digital interactive texts of the N groups of audio texts and the M groups of audio texts; In a case where the priority of the first conversation scenario is higher than that of the second conversation scenario, generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: Based on the set feature fine-grained weights, the first speech information to be translated and the second speech information to be translated are optimized in feature fine-grained manner to obtain first speech adjustment information and second speech adjustment information; Performing content interaction on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information: obtaining a first entity relationship network corresponding to the first voice adjustment information; obtaining a second entity relationship network corresponding to the second voice adjustment information; performing mean processing on entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain a global entity feature value corresponding to each entity distribution label; generating the first interactive voice information based on the global entity feature value corresponding to each entity distribution label, or obtaining a first entity relationship network corresponding to the first voice adjustment information; obtaining a second entity relationship network corresponding to the second voice adjustment information; mapping entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain an entity mapping feature value corresponding to each entity distribution label; generating the first interactive voice information based on the entity mapping feature value corresponding to each entity distribution label; If the first interactive voice information triggers the digital consecutive interpretation mechanism, semantic vector mining is performed on the first interactive voice information to obtain a first semantic interaction vector; wherein the semantic vector mining is implemented by a large language model system; By using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector, digital consecutive interpretation is performed on the first semantic interaction vector to obtain a consecutive interpretation annotated text.

2. The method according to claim 1, characterized in that In a case where the priority of the first conversation scenario is higher than that of the second conversation scenario, generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: According to the content load corresponding to the second voice information to be translated, the first voice information to be translated is subjected to a first feature adjustment to obtain first voice adjustment information, wherein the audio text in the first voice adjustment information is mapped to the audio text in the second voice information to be translated; using the second voice information to be translated as second voice adjustment information; Content interaction is performed on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information.

3. The method according to claim 1, characterized in that In a case where the priority of the first conversation scenario is higher than that of the second conversation scenario, generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: using the first speech information to be translated as first speech adjustment information; According to the content load corresponding to the first voice information to be translated, the second voice information to be translated is subjected to a second feature adjustment to obtain second voice adjustment information, wherein the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated; Content interaction is performed on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information.

4. The method according to claim 1, characterized in that In a case where the priority of the first conversation scenario is lower than that of the second conversation scenario, generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: using the first speech information to be translated as first speech adjustment information; According to the content load corresponding to the first voice information to be translated, the second voice information to be translated is subjected to a first feature adjustment to obtain second voice adjustment information, wherein the audio text in the second voice adjustment information is mapped to the audio text in the first voice information to be translated; Performing content interaction on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information; Alternatively, when the priority of the first conversation scenario is lower than that of the second conversation scenario, generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: According to the content load corresponding to the second voice information to be translated, the first voice information to be translated is subjected to a second feature adjustment to obtain first voice adjustment information, wherein the audio text in the first voice adjustment information is mapped to the audio text in the second voice information to be translated; using the second voice information to be translated as second voice adjustment information; Performing content interaction on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information; Alternatively, when the priority of the first conversation scenario is lower than that of the second conversation scenario, generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: Based on the set feature fine-grained weights, the first speech information to be translated is optimized in feature fine-grained manner to obtain first speech adjustment information; According to the set feature fine-grained weight, the feature fine-grained optimization is performed on the second speech information to be translated to obtain second speech adjustment information; Content interaction is performed on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information.

5. The method according to claim 1, characterized in that In a case where the priority of the first conversation scenario is the same as the priority of the second conversation scenario; generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: Acquire each audio text unit associated with the first voice information to be translated and the second voice information to be translated; The associated audio text units are sorted, and the unassociated audio text units in the first voice information to be translated and the second voice information to be translated are retained to obtain the first interactive voice information.

6. The method according to claim 1, characterized in that The performing semantic vector mining on the first interactive voice information to obtain the first semantic interaction vector includes: sending the first interactive voice information to a large language model system so that the large language model system performs semantic vector mining on the first interactive voice information to obtain the first semantic interaction vector; The method of using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector to perform digital consecutive interpretation on the first semantic interaction vector to obtain a consecutive interpretation annotated text includes: receiving a target digital annotation indication strategy corresponding to the first semantic interaction vector sent by the large language model system, wherein the target digital annotation indication strategy is obtained by the large language model system after performing digital consecutive interpretation on the first semantic interaction vector using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector; The sending the first interactive voice information to the large language model system so that the large language model system performs semantic vector mining on the first interactive voice information to obtain the first semantic interaction vector includes: performing voice enhancement and context feature annotation on the first interactive voice information to obtain target interactive voice information; performing semantic upsampling on the target interactive voice information to obtain upsampled semantic information; sending the upsampled semantic information to the large language model system so that the large language model system recognizes the upsampled semantic information to obtain the target interactive voice information, and performing semantic vector mining on the target interactive voice information to obtain the first semantic interaction vector.

7. The method according to claim 1, characterized in that After generating the first interactive voice information based on the first voice information to be translated and the second voice information to be translated, the method further includes: If the first interactive voice information does not trigger the digital consecutive interpretation mechanism, the third voice information to be translated corresponding to the third conversation scenario is collected through the voice input terminal, wherein the third voice information to be translated includes G groups of audio texts of the target conversation event, the third conversation scenario represents the interpretation task state between the target conversation event and the voice input terminal in a third consecutive interpretation cycle, the third consecutive interpretation cycle is the next consecutive interpretation cycle of the second consecutive interpretation cycle, and G is an integer not less than 1; Generate second interactive voice information based on the first interactive voice information and the third voice information to be translated, wherein the second interactive voice information includes digital interactive texts of the N group of audio texts, the M group of audio texts, and the G group of audio texts; If the second interactive voice information triggers the digital consecutive interpretation mechanism, performing semantic vector mining on the second interactive voice information to obtain a second semantic interaction vector; By using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector, digital consecutive interpretation is performed on the second semantic interaction vector to obtain a consecutive interpretation annotated text.

8. A digital consecutive interpretation device, characterized in that: The digital consecutive interpretation device comprises: A voice collection module, used for collecting first voice information to be translated corresponding to a first conversation scenario through a voice input terminal, wherein the first voice information to be translated includes N groups of audio texts of a target conversation event, the first conversation scenario represents the state of an interpretation task between the target conversation event and the voice input terminal in a first consecutive interpretation cycle, and N is an integer not less than 1; collecting second voice information to be translated corresponding to a second conversation scenario through the voice input terminal, wherein the second voice information to be translated includes M groups of audio texts of the target conversation event, the second conversation scenario represents the state of an interpretation task between the target conversation event and the voice input terminal in a second consecutive interpretation cycle, the second consecutive interpretation cycle is the next consecutive interpretation cycle of the first consecutive interpretation cycle, and M is an integer not less than 1; A speech generation module, configured to generate first interactive speech information based on the first speech information to be translated and the second speech information to be translated, wherein the first interactive speech information includes digital interactive texts of the N groups of audio texts and the M groups of audio texts; A semantic mining module, configured to perform semantic vector mining on the first interactive voice information to obtain a first semantic interaction vector if the first interactive voice information triggers a digital consecutive interpretation mechanism; wherein the semantic vector mining is implemented by a large language model system; In a case where the priority of the first conversation scenario is higher than that of the second conversation scenario, generating first interactive voice information based on the first voice information to be translated and the second voice information to be translated includes: Based on the set feature fine-grained weights, the first speech information to be translated and the second speech information to be translated are optimized in feature fine-grained manner to obtain first speech adjustment information and second speech adjustment information; Performing content interaction on the first voice adjustment information and the second voice adjustment information to obtain the first interactive voice information: obtaining a first entity relationship network corresponding to the first voice adjustment information; obtaining a second entity relationship network corresponding to the second voice adjustment information; performing mean processing on entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain a global entity feature value corresponding to each entity distribution label; generating the first interactive voice information based on the global entity feature value corresponding to each entity distribution label, or obtaining a first entity relationship network corresponding to the first voice adjustment information; obtaining a second entity relationship network corresponding to the second voice adjustment information; mapping entity feature values ​​of the same entity distribution label in the first entity relationship network and the second entity relationship network to obtain an entity mapping feature value corresponding to each entity distribution label; generating the first interactive voice information based on the entity mapping feature value corresponding to each entity distribution label; The consecutive interpretation module is used to perform digital consecutive interpretation on the first semantic interaction vector by using a plurality of pre-stored benchmark semantic interaction vectors and a digital annotation indication strategy corresponding to each benchmark semantic interaction vector to obtain a consecutive interpretation annotated text.

9. A digital consecutive interpretation device, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 7.

Images