A Transformer-based Intelligent Data Analysis Assistance System for Negotiation
By using multimodal data input and feature encoding based on the transformer architecture, dynamic negotiation strategies are generated, which solves the problems of lack of comprehensiveness in analysis results and lag in strategy adjustment in traditional systems, and achieves more efficient negotiation analysis and strategy optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- MINNAN NORMAL UNIV
- Filing Date
- 2025-11-14
- Publication Date
- 2026-07-17
AI Technical Summary
Traditional negotiation analysis systems, due to their use of a single-modal data acquisition method, lack comprehensiveness in their analysis results and struggle to capture long-distance semantic dependencies and respond to scenario changes during the negotiation process.
A multimodal data input module is used to collect text, voice, video and historical data in the negotiation scenario, and feature tokenization is performed through the transformer architecture. Combined with the adaptive policy generation module, a dynamic negotiation strategy is generated, and the loss function optimization module is used for training.
It achieves the fusion and processing of multi-dimensional information, improves the comprehensiveness of analysis results and the ability to dynamically optimize strategies in real time, and enhances the success rate of negotiations and the accuracy of risk identification.
Smart Images
Figure CN121144912B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence-assisted technology, and in particular to a negotiation data intelligent analysis assistance system based on transformer. Background Technology
[0002] Business negotiation is a communication and consultation activity conducted by enterprises or organizations to maximize economic benefits and coordinate cooperative relationships. Its core lies in reaching a mutually acceptable profit-sharing solution through strategic dialogue. Essentially, it is a process of value creation and interest balancing, requiring both adherence to one's own bottom line and an understanding of the other party's needs, seeking a win-win situation through competition and cooperation. Business negotiation has significant economic characteristics, revolving around commercial elements such as price, cost, profit, and cooperation terms, with the direct goal of optimizing economic benefits. It also possesses strategic elements, requiring negotiators to flexibly utilize techniques such as information screening, condition exchange, and psychological game theory to maintain the initiative in the offensive and defensive phases. The negotiation process often involves information asymmetry, requiring participants to conduct thorough intelligence gathering and pre-planning to cope with dynamically changing negotiation situations.
[0003] Traditional negotiation analysis systems mostly use a single-modality data acquisition method, which results in a lack of comprehensiveness in the analysis results due to the limited dimensions of information acquisition. Summary of the Invention
[0004] To overcome the above shortcomings, this invention provides a transformer-based intelligent data analysis assistance system for negotiation, aiming to improve the problem that traditional negotiation analysis systems mostly adopt a single-modal data acquisition method, resulting in a lack of comprehensive analysis results due to the single dimension of information acquisition.
[0005] This invention provides the following technical solution: a transformer-based intelligent analysis and assistance system for negotiation data, comprising: A multimodal data input module is used to collect multimodal data in negotiation scenarios, including text, voice, video, and historical data. The dynamic feature encoding module, built using the transformer architecture, is used to tokenize the multimodal data to generate corresponding feature token sequences, which include text token sequences, voice token sequences, visual token sequences, and historical feature token sequences. An adaptive strategy generation module is used to generate a dynamic negotiation strategy through the feature token sequence, the strategy including risk assessment results, concession threshold suggestions and opponent intention prediction; The real-time interactive output module is used to visualize or output the dynamic negotiation strategy by voice and to receive user feedback. The loss function optimization module is used to train the dynamic feature encoding module and the adaptive strategy generation module by combining a multi-task loss function that incorporates the success rate of negotiation results, risk coverage rate, and user satisfaction indicators.
[0006] By adopting the above technical solution, a multimodal data input module is used to collect text, voice, video, and historical data in the negotiation scenario, thereby realizing the fusion processing of multi-dimensional information. This improves the problem that traditional negotiation analysis systems mostly use a single-modal data collection method, which results in a lack of comprehensive analysis results due to the single dimension of information acquisition.
[0007] Preferably, the dynamic feature encoding module includes: a text encoder, a speech encoder, a visual encoder, a historical data encoder, and a feature fusion layer.
[0008] Preferably, the text encoder is used to tokenize text data and generate a text token sequence; the speech encoder is used to extract features from speech data and generate a speech token sequence; the visual encoder is used to extract spatiotemporal features from video data and generate a visual token sequence; the historical data encoder is used to encode features from historical data and generate a historical feature token sequence; and the feature fusion layer is used to perform cross-modal fusion of the text token sequence, speech token sequence, visual token sequence, and historical feature token sequence through an attention mechanism.
[0009] Preferably, the speech encoder uses the Mel frequency cepstral coefficient algorithm to extract acoustic features, and the Mel frequency conversion formula is as follows.
[0010] Preferably, the visual encoder includes a spatiotemporal attention submodule, which is used to extract spatiotemporal features of micro-expressions and body movements in video data and generate a dynamic attention mask.
[0011] Preferably, the text encoder includes a semantic enhancement submodule, which performs contextual enhancement on the text token sequence through positional encoding and semantic parsing, wherein the semantic parsing employs an attention mechanism.
[0012] Preferably, the adaptive strategy generation module includes: an opponent behavior prediction unit and a dynamic threshold adjustment unit; The adversary behavior prediction unit uses a generative network to construct an adversary behavior prediction model, and the generative network model generates a probability distribution of adversary bidding paths through training. The dynamic threshold adjustment unit dynamically adjusts the concession threshold using a reinforcement learning algorithm.
[0013] Preferably, the state space of the dynamic threshold adjustment unit This includes the characteristics of the current negotiation round and the historical strategic characteristics of the opponent, as well as the room for maneuver. Including threshold adjustment range, reward function Calculated based on the incremental utility of negotiation.
[0014] Preferably, the real-time interactive output module includes: a multimodal visualization unit and a feedback processing unit; The multimodal visualization unit is used to convert strategies into heatmaps and decision trees, and the heatmaps form color maps based on risk probabilities. The feedback processing unit is used to receive user ratings and adjust parameters, and to update the model weights of the dynamic feature encoding module and the adaptive strategy generation module through backpropagation.
[0015] A transformer-based intelligent data analysis assistance method for negotiation includes the following steps: S1. Collect multimodal data in the negotiation scenario through the multimodal data input module; S2. The multimodal data is tokenized using a dynamic feature encoding module to generate a feature embedding sequence; S3. A dynamic negotiation strategy is generated based on the feature embedding sequence by the adaptive strategy generation module; S4. Visualize or output the dynamic negotiation strategy through the real-time interactive output module, and receive user feedback to optimize feature representation; S5. Perform multi-task loss function training through the loss function optimization module.
[0016] The present invention has the following beneficial effects: 1. In this invention, text, voice, video and historical data in the negotiation scenario are collected by a multimodal data input module, thereby realizing the fusion processing of multi-dimensional information. This improves the problem that most traditional negotiation analysis systems adopt a single-modal data collection method, which results in a lack of comprehensiveness in the analysis results due to the single dimension of information acquisition. 2. In this invention, the dynamic feature encoding module uses the transformer architecture to tokenize multimodal data and generate feature token sequences. Then, the attention mechanism of the transformer is used to realize global correlation modeling of cross-modal features, thereby improving the problem that traditional systems mostly use CNN or RNN architectures, which are difficult to capture long-distance semantic dependencies, resulting in a lack of contextual coherence in feature representation. 3. In this invention, the adaptive strategy generation module generates a dynamic negotiation strategy based on the feature token sequence, which includes risk assessment, concession threshold suggestions and opponent intention prediction. This enables real-time dynamic optimization of the negotiation strategy, thereby improving the problem that traditional systems mostly use preset strategy templates, which cannot respond to changes in the negotiation process and thus cause a lag in strategy adjustment. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the system architecture of a transformer-based intelligent analysis and assistance system for negotiation data proposed in this invention. Figure 2 This is a schematic diagram illustrating the steps of a transformer-based intelligent analysis and assistance method for negotiation data proposed in this invention. Detailed Implementation
[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Reference Figure 1 This invention provides a transformer-based intelligent analysis and assistance system for negotiation data, comprising: The multimodal data input module is used to collect multimodal data in negotiation scenarios, including text, voice, video, and historical data. The dynamic feature encoding module, built using the transformer architecture, is used to tokenize multimodal data separately and generate corresponding feature token sequences, including text token sequences, voice token sequences, visual token sequences, and historical feature token sequences. The adaptive strategy generation module is used to generate dynamic negotiation strategies through feature token sequences. The strategies include risk assessment results, concession threshold suggestions, and opponent intention predictions. The real-time interactive output module is used to visualize or output dynamic negotiation strategies via voice and to receive user feedback. The loss function optimization module is used to train the dynamic feature encoding module and the adaptive strategy generation module by combining a multi-task loss function that incorporates the success rate of negotiation results, risk coverage rate, and user satisfaction indicators.
[0020] Specifically, the system employs a multimodal data input module to collect text, voice, video, and historical data from negotiation scenarios, providing multi-dimensional information for negotiation analysis. A dynamic feature encoding module, based on a transformer architecture, tokenizes the multimodal data, generating text, voice, visual, and historical feature token sequences. It then utilizes the transformer's attention mechanism to achieve global modeling of cross-modal features. An adaptive strategy generation module generates dynamic negotiation strategies based on the feature token sequences, including risk assessment, concession threshold suggestions, and opponent intent prediction, enabling real-time strategy generation. A real-time interactive output module visualizes or outputs the dynamic strategies via voice and receives user feedback, forming a closed-loop optimization. A loss function optimization module trains the dynamic feature encoding and adaptive strategy generation modules using a multi-task loss function that combines success rate, risk coverage, and user satisfaction metrics, improving system performance. These modules work collaboratively to achieve intelligent analysis of negotiation data and strategy-assisted generation.
[0021] The dynamic feature encoding module includes: a text encoder, a speech encoder, a visual encoder, a historical data encoder, and a feature fusion layer; A text encoder is used to tokenize text data and generate a sequence of text tokens; A speech encoder is used to extract features from speech data and generate a sequence of speech tokens; A visual encoder is used to extract spatiotemporal features from video data and generate a sequence of visual tokens; The historical data encoder is used to encode features into historical data and generate a sequence of historical feature tokens; The feature fusion layer is used to perform cross-modal fusion of text token sequences, speech token sequences, visual token sequences, and historical feature token sequences through an attention mechanism.
[0022] Specifically, the dynamic feature encoding module generates text token sequences by tokenizing text data through a text encoder, achieving a structured representation of text semantics; the speech encoder extracts features from speech data to generate speech token sequences, completing the digital conversion of acoustic features; the visual encoder extracts spatiotemporal features from video data to generate visual token sequences, capturing visual information such as micro-expressions and body movements; the historical data encoder encodes features from historical data to generate historical feature token sequences, mining the value of historical negotiation data; and the feature fusion layer fuses token sequences from various modalities across modalities through an attention mechanism, achieving correlation modeling and complementary enhancement of multi-source features, thereby constructing a comprehensive feature representation that includes text semantics, speech acoustics, visual spatiotemporal data, and historical experience, providing multi-dimensional feature support for subsequent strategy generation.
[0023] The speech encoder uses the Mel frequency cepstral coefficient algorithm to extract acoustic features. The Mel frequency conversion formula is: ; In the formula, For actual frequency, The frequency is Mel.
[0024] Specifically, the speech encoder receives speech data from the negotiation scenario, first processes it through frame segmentation and windowing, and then performs a Fast Fourier Transform on each frame to obtain the actual frequency. Finally, the actual frequency is converted using the Mel frequency conversion formula. The frequencies are mapped to Mel frequencies, and Mel frequency cepstral coefficients are then extracted as speech features. The speech encoder uses the Mel frequency cepstral coefficient algorithm to extract acoustic features, and converts the actual frequencies using the Mel frequency conversion formula. Convert to Mel frequency This transformation simulates the human ear's perception of sound frequencies, converting a linear frequency scale into a nonlinear Mel frequency scale, making the extracted speech features more consistent with the laws of human auditory perception. By processing speech data using the Mel frequency cepstral coefficient algorithm, Mel frequency cepstral coefficients reflecting the essential features of speech can be effectively extracted, generating speech token sequences. This improves the adaptability of speech features to changes in speech rate and intonation in negotiation scenarios, providing a more discriminative speech acoustic feature representation for subsequent multimodal feature fusion.
[0025] The visual encoder includes a spatiotemporal attention submodule, which is used to extract spatiotemporal features of micro-expressions and body movements in video data and generate dynamic attention masks. The text encoder includes a semantic enhancement submodule. This submodule enhances the context of the text token sequence through positional encoding and semantic parsing. The semantic parsing employs an attention mechanism, with the following formula: ,in, For the current round's text token, For historical negotiation text tokens, For the corresponding value vector, The dimension of the key vector.
[0026] Specifically, the text encoder is used to tokenize text data and generate a sequence of text tokens. The text data includes the text data from the previous round, and the text encoder obtains the text token for the current round by tokenizing the text data from the current round. The historical data encoder is used to encode features of historical data and generate a sequence of historical feature tokens. Historical data includes historical negotiation text data. The historical data encoder generates historical negotiation text tokens by tokenizing the historical negotiation text data and generating a sequence of historical feature tokens. The text encoder performs BPE segmentation on the negotiation text to generate text tokens, which are then mapped to a dimension of [dimensional value missing]. Corresponding value vector In the attention mechanism, the current round's text token serves as the query vector. Historical negotiation text tokens as key vectors Sum value vector The text encoder calculates the association weights between tokens using formulas to enhance contextual semantics. The spatiotemporal attention submodule in the visual encoder extracts spatiotemporal features from micro-expressions and body movements in the video data, generating dynamic attention masks to capture and focus on key spatiotemporal features of nonverbal behavior by negotiation participants. Positional encoding assigns temporal positional information to the text token sequence. In transformer-based text processing, by encoding the positional information of text tokens into vectors and superimposing them with text semantic feature vectors, the model can perceive the relative positional relationships of words in the text. This effectively models the temporal dependencies and contextual semantic associations of the text, improving the ability to capture long-distance semantic dependencies in negotiation texts and providing temporal dimension feature support for subsequent attention-based semantic parsing. The semantic enhancement submodule within the text encoder assigns temporal information to the text token sequence through positional encoding, combining it with the attention mechanism to enhance the current round of text tokens. Negotiating with historical text tokens Perform semantic parsing and calculate value vectors. The weighted sum of these features enhances the semantic meaning of the text context and models long-distance dependencies. Both approaches strengthen feature representation from the visual spatiotemporal dimension and the text semantic dimension, respectively, providing more accurate visual behavioral features and text semantic features for multimodal feature fusion, and improving the system's ability to understand complex information in negotiation scenarios.
[0027] The adaptive policy generation module includes: an opponent behavior prediction unit and a dynamic threshold adjustment unit; The competitor behavior prediction unit uses a generative network to build a competitor behavior prediction model. This generative network model generates the probability distribution of competitor bidding paths through training. The objective function for training is: ,in, This is a generator used to generate competitor quote path samples; The discriminator is used to distinguish between the real quote path and the generated sample. The data distribution of historical negotiation quotes, Preset input noise distribution; The dynamic threshold adjustment unit dynamically adjusts the concession threshold using a reinforcement learning algorithm. The state transition formula for the reinforcement learning algorithm is as follows: ,in, The current negotiation state is characterized by round features and the opponent's historical strategy features. For learning rate, This is the discount factor.
[0028] Specifically, This is a sample of historical negotiation price paths, showing the data distribution from historical negotiation prices. The multimodal data input module collects negotiation records from the past three years, and the historical data encoder encodes features such as the bid amount and concession range to form the data. The input noise vector follows a preset input noise distribution. A 128-dimensional noise vector is randomly generated by the opponent behavior prediction unit and used as the input to the generator; Employing a 6-layer transformer decoder architecture, input Then output the probability distribution of the competitor's bidding path; It is a bidirectional LSTM network that outputs the probability of true or false after inputting a sample. ; make real samples Discriminant probability Approaching 1, generating samples Discriminant probability The value approaches 1, thus achieving the goal of generator optimization; Log-likelihood of the true pricing path under the discriminator measures the discriminator's ability to identify true data. The log-likelihood of the generated quote path under the discriminator measures how well the generator fits the true distribution; the converged generator... It can generate data that matches the historical price distribution. The competitor's pricing path sample is output as a probability distribution. The opponent behavior prediction unit utilizes Predicting potential bidding paths from competitors provides a basis for decision-making by the dynamic threshold adjustment unit: if high-risk bids account for more than 30% of the generated samples, the dynamic threshold adjustment unit reduces the concession range; combined with the predicted bidding paths, strategy suggestions are generated. The state transition formula of the reinforcement learning algorithm is used. The input data and parameters include the current negotiation state generated by multimodal negotiation data collected by the environment perception module and encoded by the feature encoder. Based on the strategy generation unit The current action generated by the greedy strategy Instant rewards calculated and standardized in real time by the results evaluation module based on the negotiation results. and the learning rate, either preset or dynamically decaying. and a pre-set discount factor to balance immediate and long-term returns. The formula is derived by subtracting the current state from the action value. With "instant rewards" Adding the future optimal state value discount and ,in Based on the actions of the environment simulator The error of "generated by the opponent's model" multiplied by Make corrections and output the updated version. The value table stores the value estimates of actions in each state; the next step is for the policy generation unit to update the value table based on the value estimates of actions in each state. The value table adjusts the action selection, the opponent behavior prediction unit infers the opponent's strategy pattern, and enters the next state to repeat the iteration until the negotiation ends or the maximum number of rounds is reached. This allows the system to gradually learn the optimal strategy through a closed loop from state perception to action execution, then to reward feedback, and finally to strategy update, thereby achieving dynamic optimization of the concession threshold and improving the flexibility and effectiveness of the negotiation strategy.
[0029] The state space of the dynamic threshold adjustment unit This includes the characteristics of the current negotiation round and the historical strategic characteristics of the opponent, as well as the room for maneuver. Including threshold adjustment range, reward function Calculated based on the incremental utility of negotiation.
[0030] Specifically, the dynamic threshold adjustment unit dynamically optimizes the concession threshold by defining a state space, action space, and reward function. State space This encompasses the characteristics of the current negotiation round and the opponent's historical strategic characteristics, providing the system with status information reflecting the negotiation process and the opponent's behavioral patterns; action space Includes threshold adjustment range, clearly defining the range of actions the system can perform; reward function. Based on the incremental calculation of negotiation utility, quantitative feedback signals are provided to the system. With the collaboration of these three elements, the system selects the adjustment range from the action space based on the current state, evaluates the effect of the action according to the reward function, and continuously iterates and optimizes the state transition formula through reinforcement learning algorithm, thereby achieving dynamic adaptation of the concession threshold to the negotiation scenario and improving the flexibility and effectiveness of negotiation strategies.
[0031] The real-time interactive output module includes: a multimodal visualization unit and a feedback processing unit; The multimodal visualization unit is used to convert strategies into heatmaps and decision trees. The heatmap uses risk probabilities to form color maps, and the risk probability formula is: The feedback processing unit is used to receive user ratings and adjust parameters, and to update the model weights of the dynamic feature encoding module and the adaptive strategy generation module through backpropagation.
[0032] Specifically, the input data in the risk probability formula This represents the probability value for risk assessment, ranging from [0,1], output by the risk assessment model of the adaptive strategy generation module, with parameters... The function is a linear mapping function, used to transform... Map from the interval [0,1] to Color range, among which correspond For low risk, correspond The risk level is high; the output is the color value corresponding to the risk probability. The next step is for the multimodal visualization unit to apply the color value to the pixel rendering of the heatmap. The risk distribution of each stage of the negotiation strategy is displayed intuitively through color gradient. This color mapping result is used to help users quickly identify the risk points of the strategy and improve the intuitiveness and efficiency of negotiation decision-making.
[0033] The total loss function of the loss function optimization module is:
[0034] in, These represent the losses in success rate, risk coverage, and user satisfaction, respectively.
[0035] Specifically, the input data for the total loss function in the loss function optimization module includes the success rate loss calculated from the success rate statistics of historical negotiation data. The risk coverage loss is obtained by normalizing the number of uncovered risk points based on the risk assessment model. User satisfaction loss calculated from the mean squared error of user ratings of the strategy. ,parameter The weighting coefficients are determined through 10-fold cross-validation; the output is the total loss value that comprehensively reflects the multi-objective optimization error of the negotiation system. The next step is to use the backpropagation algorithm to transmit the total loss value to the dynamic feature encoding module and the adaptive policy generation module to update the model weights. This formula uses weighted fusion of multi-dimensional loss terms to guide the system to improve the negotiation success rate while taking into account risk coverage and user experience, thus achieving multi-objective balanced optimization.
[0036] Reference Figure 2 This invention provides a method for intelligent analysis and assistance of negotiation data based on transformer, comprising the following steps: S1. Collect multimodal data in the negotiation scenario through the multimodal data input module; S2. The multimodal data is tokenized using a dynamic feature encoding module to generate a feature embedding sequence; S3. A dynamic negotiation strategy is generated based on the feature embedding sequence by the adaptive strategy generation module; S4. Visualize or output the dynamic negotiation strategy through the real-time interactive output module, and receive user feedback to optimize feature representation; S5. Perform multi-task loss function training through the loss function optimization module.
[0037] Specifically, this method collects text, voice, video, and historical data from negotiation scenarios through a multimodal data input module. This data is then tokenized by a dynamic feature encoding module to generate feature embedding sequences, providing a structured representation of negotiation information. An adaptive strategy generation module generates dynamic negotiation strategies based on these feature embedding sequences, incorporating risk assessment, concession threshold suggestions, and opponent intent prediction, thus achieving intelligent strategy generation. A real-time interactive output module visualizes or outputs the strategies via voice and receives user feedback, forming a human-machine collaborative optimization loop. A loss function optimization module performs multi-task loss function training, comprehensively considering success rate, risk coverage, and user satisfaction metrics to improve system performance. These steps work together to achieve multimodal processing of negotiation data, dynamic generation and optimization of strategies, and effective utilization of user feedback, forming a complete intelligent analysis and support system.
[0038] The following examples further illustrate this point: Experimental configuration: Hardware: NVIDIA A100 GPU×4, CPU Intel Xeon Platinum 8352Y; Software: PyTorch 2.0, TensorBoard log monitoring to ensure algorithm reproducibility; Datasets: 500 business negotiation datasets: 450 rounds / sessions of text-based dialogues, 16kHz speech signals with 8 tone labels, 1080P videos with 20 micro-expressions and body language annotations, and a historical negotiation case library with 100,000 structured data entries; 200 simulated dynamic scenarios: sudden changes in competitor strategy, such as price reductions or the introduction of competition; environmental interference, such as background noise and video blurring; The embodiments of the present invention employ multimodal data input, transformer feature encoding, and dynamic strategy generation; Comparative Example 1 uses unimodal text input and rule engine analysis, but the rule engine analysis does not integrate voice, video, and historical data; Comparative Example 2 uses CNN plus RNN feature encoding plus concatenation and fusion, which is a non-transformer architecture; Comparative Example 3 uses a preset strategy template plus a fixed threshold adjustment, but the fixed threshold adjustment does not have a dynamic strategy generation module.
[0039] Table 1: Information dimension coverage (number of items) 4 1 Completeness of risk point identification (%) 92% 65% Typical scenarios Able to identify risks Text keyword matching only In Table 1, the calculation method for information dimension coverage is: the coverage of four data types: text data coverage, voice data coverage, video data coverage, and historical data coverage; the calculation method for risk point identification completeness is: the ratio of correctly identified risks to actual risks; the calculation method for typical scenarios is: accelerating the speaker's speech speed and associating it with historical failure cases, and then checking the identification status of the embodiments of the present invention and Comparative Example 1; the experimental results in Table 1 show that the present invention, through the full-dimensional collection of text, voice, video, and historical data, improves the risk identification completeness by 27%, solving the problem of the one-sidedness of traditional single-modal analysis.
[0040] Table 2: Cross-modal semantic accuracy (%) 94% 78% Long-distance dependence strength 0.82 0.35 Technical details 6 transformer layers and 8 attention heads 128 hidden layers in RNN and 16 layers in CNN In Table 2, the cross-modal semantic accuracy is calculated as follows: text to video semantic accuracy. Figure 1 The accuracy of consistency classification; the calculation method of long-distance dependency strength is: the mean of the association strength of the attention heatmap; as shown in the experimental results in Table 2, compared with the concatenation and fusion of CNN and RNN, the cross-modal semantic accuracy of this invention is improved by 16%, and the attention mechanism improves the long-distance dependency modeling ability by 134%, which verifies the significant progress of transformer in cross-modal feature association.
[0041] Table 3: Negotiation success rate (%) 85% 60% Policy response time (s) 1.2s none Utility score 88 65 In Table 3, the negotiation success rate is calculated as the percentage of rounds in which an effective agreement is reached; the strategy response time is calculated as the time taken from scenario change to the generation of a new strategy; and the utility score is calculated as (gain - risk loss) × adjustment flexibility. Based on the experimental results in Table 2, the success rate of this invention reaches 85% when facing simulated scenario changes, and the strategy response time is controlled within 1.2 seconds. Comparative Example 3, because it only uses a fixed strategy, has no scenario change leading to the generation of a new strategy. This invention solves the problem of lag in the preset template strategy by dynamically adjusting the concession threshold through reinforcement learning.
[0042] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A transformer-based intelligent data analysis and assistance system for negotiation, characterized in that, include: A multimodal data input module is used to collect multimodal data in negotiation scenarios, including text, voice, video, and historical data. The dynamic feature encoding module, built using the transformer architecture, is used to tokenize the multimodal data to generate corresponding feature token sequences, which include text token sequences, voice token sequences, visual token sequences, and historical feature token sequences. An adaptive strategy generation module is used to generate a dynamic negotiation strategy through the feature token sequence, the strategy including risk assessment results, concession threshold suggestions and opponent intention prediction; The adaptive policy generation module includes: an opponent behavior prediction unit and a dynamic threshold adjustment unit; the opponent behavior prediction unit uses a generative network to construct an opponent behavior prediction model, and the generative network model generates the probability distribution of the opponent's bidding path through training; the dynamic threshold adjustment unit dynamically adjusts the concession threshold through a reinforcement learning algorithm; the state space of the dynamic threshold adjustment unit... This includes the characteristics of the current negotiation round and the historical strategic characteristics of the opponent, as well as the room for maneuver. Including threshold adjustment range, reward function Calculated based on the incremental utility of negotiation; The real-time interactive output module is used to visualize or output the dynamic negotiation strategy by voice and to receive user feedback. The loss function optimization module is used to train the dynamic feature encoding module and the adaptive strategy generation module by combining a multi-task loss function that incorporates the success rate of negotiation results, risk coverage rate, and user satisfaction indicators. A multimodal data input module is used to collect text, voice, video, and historical data from negotiation scenarios. A transformer architecture is used to tokenize the multimodal data, generating corresponding text token sequences, voice token sequences, visual token sequences, and historical feature token sequences. These feature token sequences are then used to generate dynamic negotiation strategies that include risk assessment results, concession threshold suggestions, and opponent intent predictions. These dynamic negotiation strategies are visualized or output via voice, and user feedback is received, thus achieving multi-dimensional information fusion. A multi-task loss function, combining negotiation success rate, risk coverage rate, and user satisfaction metrics, is used to train the dynamic feature encoding module and adaptive strategy generation module, ultimately achieving real-time dynamic optimization of the negotiation strategy.
2. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 1, characterized in that: The dynamic feature encoding module includes: a text encoder, a speech encoder, a visual encoder, a historical data encoder, and a feature fusion layer.
3. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 2, characterized in that: The text encoder is used to tokenize text data and generate a text token sequence; the speech encoder is used to extract features from speech data and generate a speech token sequence; the visual encoder is used to extract spatiotemporal features from video data and generate a visual token sequence; the historical data encoder is used to encode features from historical data and generate a historical feature token sequence. The feature fusion layer is used to perform cross-modal fusion of the text token sequence, voice token sequence, visual token sequence and historical feature token sequence through an attention mechanism.
4. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 3, characterized in that: The speech encoder uses the Mel frequency cepstral coefficient algorithm to extract acoustic features.
5. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 3, characterized in that: The visual encoder includes a spatiotemporal attention submodule, which is used to extract spatiotemporal features of micro-expressions and body movements in video data and generate a dynamic attention mask.
6. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 3, characterized in that: The text encoder has a semantic enhancement submodule, which enhances the context of the text token sequence through positional encoding and semantic parsing. The semantic parsing uses an attention mechanism.
7. The intelligent analysis and assistance system for negotiation data based on transformer according to claim 1, characterized in that: The real-time interactive output module includes: a multimodal visualization unit and a feedback processing unit; The multimodal visualization unit is used to convert strategies into heatmaps and decision trees, and the heatmaps form color maps based on risk probabilities. The feedback processing unit is used to receive user ratings and adjust parameters, and to update the model weights of the dynamic feature encoding module and the adaptive strategy generation module through backpropagation.
8. A transformer-based intelligent analysis and assistance method for negotiation data, applied to the transformer-based intelligent analysis and assistance system for negotiation data as described in any one of claims 1-7, characterized in that, Includes the following steps: S1. Collect multimodal data in the negotiation scenario through the multimodal data input module; S2. The multimodal data is tokenized using a dynamic feature encoding module to generate a feature embedding sequence; S3. A dynamic negotiation strategy is generated based on the feature embedding sequence by the adaptive strategy generation module; S4. Visualize or output the dynamic negotiation strategy through the real-time interactive output module, and receive user feedback to optimize feature representation; S5. Perform multi-task loss function training through the loss function optimization module.