Methods and a system for the tactical analysis of statements

Abstract

Computer-implemented method for the tactical analysis of statements made by a negotiating partner and / or for the automatic generation of a strategically aligned response proposal, comprising the following steps: S1: Capturing a negotiation data stream (104), in particular comprising audio and / or text data (106), of a negotiating partner in real time by an input unit (102); S2: Extracting a feature vector (108) from the negotiation data stream (104) by an analysis component (110), forming prosodic features, in particular a fundamental frequency and / or a volume and / or a speaking rate, and / or lexical-syntactic features; S3: Determining a psycholinguistic profile (112) of the negotiating partner using the feature vector (108) by means of at least one machine learning model (114) implemented in a processing and computing unit (100); S4: Classifying the negotiation data stream (104) into at least one tactical category (116), in particular exhibiting aggression, de-escalation, neutrality, decisiveness or obfuscation, using at least one machine learning model (114), in particular also using the psycholinguistic profile (112); S5: Determining a response strategy (118) through a decision module (120) that takes into account the tactics category (116) as well as the psycholinguistic profile (112); S6: Generating a textual and / or linguistic response proposal (122) based on the specified response strategy (118) by a generation unit (124); and S7: Outputting the proposed response (122) to an output unit (126), wherein steps S1 to S7 are preferably performed iteratively for each received negotiation data stream (104).

Description

[0001] The invention relates to a method and a system for the tactical analysis of statements made by a negotiating partner and for the automatic generation of a strategically coordinated response proposal. Furthermore, the invention relates to a corresponding computer program product and a computer-readable data carrier. State of the art

[0002] Interpersonal negotiations, particularly in business, diplomatic, legal, or automated contexts, are characterized by complex communicative dynamics. Recently, computer-assisted systems have been increasingly used to analyze conversation content, record conversation flows, or semi-automatically monitor interactions.

[0003] Known state-of-the-art systems typically analyze either the linguistic content (text) or vocal characteristics (audio) of conversational contributions. These systems employ, among other things, methods of natural language processing (NLP) and speech analysis (e.g., for recognizing emotional states).

[0004] Existing solutions are often limited to individual data modalities, such as purely textual chat histories, and only consider partial aspects of the communicative intent of the other party. The identification of deeper communicative tactics—such as targeted escalation, rhetorical obfuscation, or strategic determination—is still largely done manually or heuristically, e.g., through human moderation.

[0005] The general state of the art is known from DE 10 2022 205 123 A1 and WO 2022 / 194 686 A1.

[0006] It is therefore an object of the invention to provide a method and / or a system that enables an improved, tactically differentiated analysis of negotiating statements made by a negotiating partner and generates an automatic, strategically coordinated reply. Disclosure of the invention

[0007] The problem is solved by a method according to the features of claim 1. The problem is further solved by a system according to the features of claim 15.

[0008] According to a first aspect, a procedure for the tactical analysis of statements made by a negotiating partner and / or for the automatic generation of a strategically aligned response proposal is suggested, comprising the following steps: 1.1 Capturing a negotiation data stream, in particular including audio and / or text data, of a negotiating partner in real time; 1.2 Extracting a feature vector from the negotiation data stream, whereby prosodic features, in particular a fundamental frequency and / or a volume and / or a speaking rate, and / or lexical-syntactic features are formed; 1.3 Determining a psycholinguistic profile of the negotiating partner using the feature vector by means of at least one machine learning model; 1.4 Classifying the negotiation data stream into at least one tactical category, in particular exhibiting aggression, de-escalation, neutrality, decisiveness or obfuscation, using at least one machine learning model, in particular also using the psycholinguistic profile; 1.5 Determining a response strategy using a decision module that takes into account both the tactics category and the psycholinguistic profile; 1.6 Generating a textual and / or linguistic response proposal based on the specified response strategy; and 1.7 Outputting the suggested answer to an output unit, where steps 1.1 to 1.7 are preferably performed iteratively for each received negotiation data stream.

[0009] The claimed method serves for the tactical analysis of a negotiating partner's utterances and / or the automated generation of a content-aligned response proposal. Preferably, audio and / or text data from a negotiating partner are captured in real time. Real-time processing enables continuous interaction without perceptible delay. Features are then extracted from the received negotiation data stream, preferably analyzing both prosodic (e.g., fundamental frequency, volume, speaking rate) and lexical-syntactic information. Prosodic features relate to the non-content-related vocal characteristics of a speaker, preferably for the detection of emphasis and / or emotional expression. Lexical-syntactic features preferably relate to word choice, filler words, intonation, pauses, sentence structure, and / or grammatical structure.These features are used to calculate a feature vector, which serves as the basis for further analysis.

[0010] Based on the feature vector, a psycholinguistic profile of the negotiating partner is preferably determined. A psycholinguistic profile describes characteristic language patterns that suggest personality traits and cognitive tendencies, preferably derived using machine learning methods. This allows for an individually tailored assessment of the communication style. In the next step, the entire data stream is classified into a communicative tactic category, preferably including typical categories such as aggression, de-escalation, neutrality, assertiveness, or obfuscation. This allows for a structured interpretation of the other party's conversational strategy.

[0011] The identified tactic category is combined with the psycholinguistic profile of a decision-making unit, which then selects a suitable response strategy. Considering both factors together leads to a more situationally sensitive reaction than a purely content- or tone-based assessment.

[0012] Based on this strategy, a verbal and / or textual response proposal is generated, preferably guided by rhetorical principles and the recognized conversational situation. The response proposal is then passed to an output unit, such as an output field or a speech output module. All steps can preferably be repeated multiple times for successive utterances in an ongoing negotiation to enable dynamically adapted communication. This increases the system's responsiveness and customizability over the course of a multi-stage negotiation.

[0013] It is understood that the steps according to the invention, as well as further optional steps, do not necessarily have to be carried out in the sequence shown, but can also be carried out in a different sequence. Furthermore, additional intermediate steps may be provided. The individual steps may also comprise one or more sub-steps without thereby departing from the scope of the method according to the invention.

[0014] According to a second aspect, a system for the tactical analysis of statements made by a negotiating partner and for the automatic generation of a strategically coordinated response proposal is proposed, wherein the system has an evaluation and computing unit that is trained to perform the following steps: 15.1 Capturing a negotiation data stream, in particular including audio and / or text data, of a negotiating partner in real time; 15.2 Extracting a feature vector from the negotiation data stream, forming prosodic features, in particular a fundamental frequency and / or a volume and / or a speaking rate, and lexical-syntactic features; 15.3 Determining a psycholinguistic profile of the negotiating partner using the feature vector by means of at least one machine learning model; 15.4 Classifying the negotiation data stream into at least one tactical category, in particular exhibiting aggression, de-escalation, neutrality, decisiveness or obfuscation, using at least one machine learning model, in particular also using the psycholinguistic profile; 15.5 Determining a response strategy using a decision module that takes into account both the tactics category and the psycholinguistic profile; 15.6 Generating a textual and / or linguistic response proposal based on the specified response strategy; and 15.7 Outputting the suggested answer to an output unit.

[0015] The claimed system comprises an evaluation and computing unit configured to execute all steps of the process using a computer. This computing unit can preferably be implemented as a modular structure using software and / or hardware and is capable of processing data streams in real time. The "units" mentioned in the claim preferably refer to components that are functionally designed as modules and can be implemented either by software algorithms, dedicated hardware, or a combination thereof. The system components can be implemented locally on an end device or distributed across a server infrastructure. Preferably, execution takes place at least partially in a cloud-based environment or on a dedicated server to efficiently meet computational load, scalability, and latency requirements.

[0016] To implement the method, at least one machine learning model is used, preferably a language model and / or a multimodal classification model. This machine learning model can be trained on annotated conversation data, preferably using a corpus of real or synthetic negotiation dialogues that includes annotations for each utterance regarding tactic category and communication style. The annotations can be performed manually by experts or semi-automatically using heuristics. The training data preferably consists of combined inputs that include both prosodic features such as pitch, volume, and speech rate, as well as lexical-syntactic features. This input data is preferably normalized and segmented to enable the models to perform context-sensitive evaluations.The output data of the machine learning model preferably consists of probability values ​​for the respective tactic categories as well as additional feature values ​​that are passed as input to the decision module.

[0017] A first practical application is the use of the system in a virtual negotiation trainer, where a human user interacts with a simulated conversation partner. During the interaction, the system analyzes the user's verbal and vocal cues in real time, classifies their communication style, and provides appropriately tailored, didactically suitable responses. This can be used for training purposes in areas such as leadership development, sales, or mediation.

[0018] A second example is the integration of the system into a call center assistant that accompanies the caller during a customer conversation and analyzes their communication behavior. The system recognizes any escalation tendencies or tactically motivated obfuscation and suggests suitable phrases or rhetorical devices to the call center employee to de-escalate or structure the conversation. Both examples demonstrate that the system enables adaptive, content- and context-based communication control that goes beyond classic rule-based dialogue systems.

[0019] Another application of the method lies in the legal field, particularly in settlement negotiations within the framework of out-of-court dispute resolution. Here, the system can be used as a supporting tool for lawyers or mediators to analyze the opposing party's communication tactics—such as threats, distraction, appeasement, and / or strategic vagueness—in real time. Based on the identified tactic category and the determined psycholinguistic profile, the system preferably suggests legally appropriate yet tactically astute responses. This enables more targeted argumentation, tailored to the emotional and linguistic state of the other party. Simultaneously, the system can be used to preview one's own formulations for their potential impact and adjust them accordingly, for example, to avoid unintended escalation or to clarify contentious points.Such an application increases the quality of negotiations and can be particularly helpful in complex comparison discussions with high dynamics or psychological pressure.

[0020] The explanations given for the procedure apply accordingly to the system. It is understood that linguistic modifications of procedurally formulated features can be reformulated for the system according to common linguistic practice, without such formulations needing to be explicitly listed here.

[0021] In another aspect, it is proposed that the at least one machine learning model has a multimodal transformer network that fuses the prosodic and lexical-syntactic features extracted in step 1.2 into a common representation.

[0022] The disclosure specifies that the machine learning model is designed as a multimodal transformer network. A transformer is preferably a neural network architecture based on so-called self-attention mechanisms and is particularly well-suited for processing sequential data such as speech. The multimodal implementation means that the model simultaneously processes several different data types, in particular prosodic features from the audio signal and lexical-syntactic features from the text signal. The fusion of these features into a common representation allows the model to directly learn relationships between linguistic content and vocal expression. This increases classification accuracy, especially in complex communicative situations where purely text-based or purely acoustic signals are insufficient to reliably capture conversational tactics.

[0023] In another aspect, it is proposed that the psycholinguistic profile includes Big Five personality dimensions and / or LIWC categories and is determined in particular by fine-tuning a BERT language model.

[0024] The disclosure stipulates that the psycholinguistic profile will be based on so-called Big Five personality dimensions and / or LIWC categories. The Big Five preferably describe the five basic personality factors of openness, conscientiousness, extraversion, agreeableness, and neuroticism. LIWC (Linguistic Inquiry and Word Count) is a lexicon-based text analysis method that preferably identifies certain psychologically relevant word categories, such as emotional terms, pronoun usage, or thinking styles. The profile is determined using a BERT model, preferably a fine-tuned one, which is specialized for language processing. BERT is a bidirectional transformer language model that recognizes contextual relationships between words. The fine-tuning is based on conversational data where the speaker's personality traits are known.The resulting profile provides deeper insights into communicative preferences and makes it possible to tailor responses more precisely to the individual dynamics of the conversation.

[0025] In another aspect, it is proposed that the decision module correlates a weighted confidence rating of the tactic category determined in step 1.4 with at least one personality parameter of the psycholinguistic profile.

[0026] The disclosure stipulates that the decision-making module, when selecting a response strategy, considers both the probability of the identified tactic category and at least one personality parameter from the psycholinguistic profile. Preferably, a weighted assessment is used, allowing the influence of individual factors to be dynamically weighted depending on the situation. Such an assessment ensures that an utterance formally classified as aggressive, but made by a speaker who tends to be impulsive or extroverted, will be answered differently than the same utterance made by a speaker who tends to be controlled or introverted. This leads to a more sensitive and targeted response selection and reduces the risk of reinforcing misinterpreted statements through inappropriate reactions.The combination of both parameter flows in decision-making goes beyond simple threshold rules and opens up the possibility of systematically modeling interpersonal dynamics.

[0027] In another aspect, it is suggested that the proposed response includes rhetorical means suitable for de-escalating a statement classified as aggression and / or exposing a statement classified as obfuscation and / or responding to a statement classified as definiteness.

[0028] As part of the disclosure process, the automatically generated response suggestion is designed to address the previously identified tactic category. For example, a statement identified as aggression will preferably be met with de-escalating language to reduce tension in the conversation. A statement identified as obfuscation will preferably be addressed with a revealing or clarifying rhetorical strategy that specifically addresses contradictions or vague statements. If a statement is identified as definite, the response suggestion can react with structured counter-arguments or targeted confirmation. The selection of such rhetorical devices is preferably context-sensitive and based on previously trained response patterns. This contributes to managing the dynamics of the conversation and can help avoid misunderstandings, prevent escalation, or achieve conversational goals more effectively.

[0029] In another aspect, it is proposed that the latency between receiving an utterance and outputting the suggested response should be no more than 500 ms.

[0030] The disclosure concerns the response speed of the process. It is stipulated that the time between capturing a utterance and generating the suggested response should not exceed 500 milliseconds. This latency limit preferably ensures that the generated response can be introduced into an ongoing conversation without any perceptible delay. Such a short response time allows the system to interact with human interlocutors in real time without disrupting the natural flow of conversation. At the same time, it ensures that no crucial moments of reaction are missed, even in fast-paced negotiations. To meet this time requirement, all computational and classification processes must be implemented efficiently, preferably through parallelization or hardware-accelerated inference.

[0031] In another aspect, it is proposed that user feedback on the effectiveness of the suggested answer be collected and the parameters of the at least one machine learning model be adjusted online using reinforcement learning.

[0032] The revelation describes a dynamic learning mechanism based on user feedback. During or after the output of a suggested answer, user feedback is preferably collected, indicating whether the suggestion was perceived as helpful, accurate, or situationally appropriate. This feedback is fed into the machine learning model, which further adapts itself through so-called reinforcement learning methods. Reinforcement learning is a process in which a model learns to improve its decisions through reward signals, without needing to be explicitly trained with right / wrong labels. Through continuous feedback on real-world communication processes, the system can adapt to user preferences, industry-specific conversation styles, or culture-specific communication habits. This increases the long-term relevance and accuracy of the suggestions generated by the system.

[0033] In another aspect, it is proposed that additional video data from the user and / or the negotiating partner be recorded and features from facial expression and head posture be included in the classification of the tactic category according to step 1.4.

[0034] The disclosure expands the input modality of the procedure to include visual information. In addition to audio and text data, video data of the speaker is preferably also captured, for example, via a camera. From this video data, features of facial expression and head posture are extracted, which can be interpreted as nonverbal communication signals. Facial expressions provide clues to emotional states such as anger, embarrassment, or uncertainty, while head postures can indicate agreement, disagreement, or dominance behavior. This visual information is preferably integrated into the classification of the tactic category to enable a more complete analysis of communicative intention. This multimodal fusion means that even in cases of ambiguous or manipulative statements, additional nonverbal cues are available, increasing the reliability of the classification.

[0035] In another aspect, it is proposed that at least one machine learning model includes an adversarially trained submodel to distinguish obfuscation strategies from inconsistent or misleading statements.

[0036] The Book of Revelation describes a specialized training procedure for the machine learning model, based on so-called adversarial training methods. This involves the use of an additional, antagonistically trained sub-model designed to specifically distinguish between genuine, consistent statements and obfuscating, misleading ones. Adversarial training is a method in which two neural networks are trained against each other: a generator produces deceptively realistic examples, while a discriminator learns to recognize them. Applied to the system in question, this means that the classification model is also prepared for subtle deception strategies, such as deliberate digressions, semantic ambiguity, or targeted omission of information.The use of such a submodel improves the robustness of obfuscation detection, even in realistic, tactically demanding negotiation situations.

[0037] In another aspect, it is proposed that at least one machine learning model is trained on a corpus of at least 1,000, preferably at least 10,000, negotiation-specific annotated utterances per tactic category.

[0038] The disclosure concerns the data basis on which the machine learning model is trained. It is stipulated that the model is trained on a corpus of preferably at least 1,000, ideally at least 10,000, negotiation-specific annotated utterances per tactic category. A tactic category represents a semantically defined communication strategy, such as aggression, de-escalation, or obfuscation. The annotations in the corpus are preferably made by experts who manually assign a tactic category to each utterance. Such a structured training dataset allows the model to learn subtle differences between similar-sounding but differently motivated statements. The higher the number of examples per category, the better the model can generalize and make reliable decisions even in unknown context scenarios.A well-balanced data distribution prevents overfitting and enables consistent classification even with emotionally charged or manipulative statements.

[0039] In another aspect, it is proposed that the modules executed in steps 1.2 to 1.6 are implemented as decoupled micro-services in containers and communicate via a REST interface.

[0040] The software components used in the processing steps from feature extraction to the generation of the suggested response are preferably implemented as decoupled microservices. Microservices are functionally distinct, independently deployable units within a distributed system architecture. They preferably communicate via standardized interfaces, especially REST APIs. This structure enables a scalable, fault-tolerant, and maintainable implementation, in which individual processing steps, such as classification or psycholinguistic profiling, can be updated or replaced independently. This decoupling allows for dynamic resource allocation, which distributes the computational load and maintains stable response times.Furthermore, a REST interface allows the components to be flexibly orchestrated in both local and cloud-based environments, and external modules to be connected as needed.

[0041] In another aspect, it is proposed that the proposed answer according to step 1.6 be generated using a finely tuned language model trained on a corpus of tactical negotiation dialogues.

[0042] The generation of the suggested response is preferably carried out using a finely tuned language model. A language model is a statistical or neural model that estimates the probability of word sequences and can thus be used to generate coherent, semantically meaningful language units. In the claimed embodiment, the model is specifically trained on a corpus of tactical negotiation dialogues. This corpus preferably contains example interactions in which replies appropriate to specific tactical categories are documented. The fine-tuning is based on such dialogue structures so that the model generates not only linguistically correct but also situationally and tactically appropriate responses.The use of a specialized, finely tuned language model ensures that the generated suggestions are relevant to the conversation, rhetorically coherent, and subtly respond to the recognized behavioral patterns of the other person. This enables a context-adaptive response strategy that goes beyond simple text modules or generic replies.

[0043] In a further aspect, it is proposed that the negotiation data stream captured in step S1 is first divided into speaker-homogeneous segments (step S1a) before feature extraction takes place in step S2. For this purpose, a specialized audio model, preferably a self-attention-based audio transformer or a deep convolutional neural network, is used, which computes time-stable speaker embeddings (e.g., x-vectors) from the continuous audio signal. An agglomerative hierarchy clustering preferably iteratively assigns these embeddings to potential speaker clusters, using a structure function based on Schwarz Bayes information (BIC) as a termination criterion. BIC regularization preferably prevents unnecessary splitting into too many speakers, since an increase in the number of speakers is only acceptable if it results in a significant reduction in inhomogeneity within the clusters.

[0044] Alternatively or additionally, it is proposed that speaker-based segmentation at the text level (step S1b) be performed if only transcribed data is available or the audio signal is of low quality. Here, a set of text passages is preferably iteratively and hypothetically assigned between 1 to n speaker labels. For each assignment scenario, stylistic homogeneity measures—for example, variance in sentence length, syntactic complexity, sentiment-based mood, and word choice patterns—are calculated. An optimization procedure preferably minimizes the overall inhomogeneity, while simultaneously applying a BIC penalty to the number of speakers, so that the procedure converges without detecting individual words as separate speakers.

[0045] The speaker segments identified in this way are preferably fed separately to steps S2 to S6, thereby obtaining speaker-specific prosodic and lexical-syntactic features. This enables the machine learning model to generate individual psycholinguistic profiles and classify tactical roles more precisely, even in scenarios with changing participants, such as panel discussions or conference calls.

[0046] In a complementary aspect, it is proposed that the decision-making module mutually weights the tactical categories recognized by multiple speakers in order to identify group dynamic situations—for example, coordinated escalations—and address them with a consolidated response strategy. This clearly distinguishes the claimed method from purely sentiment-analytical systems, such as those known from Audeering: It combines multi-speaker capable signal separation, multimodal tactical recognition, and the automatic generation of dialogue-appropriate replies in a continuous pipeline approach.

[0047] Furthermore, it is emphasized that the aforementioned steps S1a and / or S1b can be implemented modularly as micro-services and integrated into the existing system architecture via a REST interface. This keeps the overall latency below the exemplary defined limit of 500 ms and ensures the system's executability even during real-time multi-speaker analysis.

[0048] In another aspect, a computer program product is proposed, comprising instructions which, when the program is executed by a computer, cause it to perform the steps of the present procedure in one of its aspects.

[0049] The computer program product preferably comprises a collection of instructions written in one or more programming languages, preferably designed to perform the tasks and / or functions described in the method when executed by a computer. The instructions in the program are preferably designed to cause the computer to execute the various steps and processes of the method according to the specified aspects.

[0050] In another aspect, a computer-readable data carrier is proposed on which such a computer program product is stored.

[0051] This computer-readable data carrier can comprise various physical media, such as CDs, DVDs, USB flash drives, hard drives, or semiconductor storage devices, e.g., SSDs, which can be read by computers or similar electronic devices. The computer program product stored on the data carrier preferably comprises a collection of instructions or code that can be executed by a computer to perform specific functions or tasks. The program can be written in various programming languages ​​and include different components such as executable files, libraries, configuration files, and documentation. The data carrier preferably enables the computer to read and execute the program stored on it in order to perform the intended functions.

[0052] Furthermore, an A / B comparison test was conducted on a proprietary negotiation corpus comprising 12,500 turns (balanced: 25% aggression, 20% de-escalation, 35% neutrality, 10% assertiveness, 10% obfuscation): Variant A corresponds to the claimed method with profile fusion (simultaneous consideration of the tactic category and the psycholinguistic profile in the decision module), variant B is identical, but without profile fusion (decision based solely on the tactic category). Both variants ran on the same GPU instance (NVIDIA A100, batch size 32, inference latency < 350 ms). The following were evaluated: (i) the success rate of the suggested response, measured as agreement with a "gold reply" set curated by communication experts, and (ii) the escalation rate after three dialogue turns. Variant A achieved a hit rate of 86.4% compared to 71.9% for variant B and significantly reduced the escalation rate from 14.2% to 5.6% (two-sample χ²). 2-Test, p < 0.01).

[0053] Further possible embodiments, developments, aspects and / or implementations of the invention also exhibit combinations of the features mentioned above or to be explained below, which are not explicitly stated. "One" or "an" is understood here to mean "at least one" or "at least one".

[0054] The described configurations and training programs can be combined in any way desired.

[0055] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described previously or subsequently with regard to the exemplary embodiments that are not explicitly mentioned. Brief description of the drawings

[0056] The accompanying drawings are intended to provide a further understanding of the embodiments of the invention. They illustrate embodiments and, in conjunction with the description, serve to explain the principles and concepts of the invention.

[0057] Other embodiments and many of the aforementioned advantages become apparent with reference to the drawings. The elements depicted in the drawings are not necessarily shown to scale. Fig. Figure 1 shows a schematic flowchart of an embodiment of the present method. Fig. Figure 2 shows a schematic block diagram of an embodiment of the present system. Detailed description of the drawings

[0058] In the figures of the drawings, identical reference symbols denote identical or functionally equivalent elements, parts or components, unless otherwise stated.

[0059] Fig. Figure 1 shows a schematic flowchart of a procedure for the tactical analysis of statements made by a negotiating partner and for the automatic generation of a strategically aligned response proposal. The procedure according to Fig. 1 is referred to Fig. 2 explained.

[0060] The method can be carried out in any embodiment, at least partially, by a system 100, which may comprise several components not shown in detail, for example, one or more provisioning units and / or at least one evaluation and computing unit 101. It is understood that the provisioning unit may be configured together with the evaluation and computing unit, or it may be different from it. Furthermore, the system 100 may comprise a storage unit and / or an output unit and / or a display unit and / or an input unit.

[0061] The computer-implemented procedure includes at least the following steps: In step S1, a negotiation data stream 104, in particular comprising audio and / or text data 106, of a negotiating partner is captured in real time by means of an input unit 102. The data is preferably continuously fed into the further analysis process.

[0062] In step S2, a feature vector 108 is extracted from the negotiation data stream 104 by an analysis component 110. In this process, prosodic features, in particular a fundamental frequency, a volume or a speaking rate, as well as lexical-syntactic features are formed and structured together.

[0063] In step S3, a psycholinguistic profile 112 of the negotiating partner is determined by a machine learning model 114, using the previously generated feature vector 108 as input. The machine learning model 114 can be based on trained language patterns and is preferably tailored to negotiation-specific personality indicators.

[0064] In step S4, the negotiation data stream 104 is classified into at least one tactic category 116, such as aggression, de-escalation, neutrality, decisiveness or obfuscation, particularly also using the psycholinguistic profile 112. This step is also carried out by the machine learning model 114, whereby a probability weighting of the categories is possible.

[0065] In step S5, a decision module 120 determines a response strategy 118. This module processes both the tactics category 116 and the psycholinguistic profile 112 and derives a communicatively appropriate strategy from them.

[0066] In step S6, a textual and / or linguistic response proposal 122 is generated based on the response strategy 118 by a generation unit 124. Rhetorical devices can be used that are tailored to the recognized communication situation.

[0067] In step S7, the suggested answer 122 is output to an output unit 126, for example in the form of a display on a screen or as an acoustic playback.

[0068] Fig.Figure 2 shows a schematic block diagram of an embodiment of the present system 100 for carrying out the procedure for the tactical analysis of negotiation statements and for generating a strategically coordinated response proposal.

[0069] System 100 has an input unit 102 for capturing audio and / or text data 106, which are assigned to the negotiation data stream 104. An analysis component 110 is connected to the input unit 102 and is configured to extract a feature vector 108 from the captured data. This is fed to a machine learning model 114, which generates a psycholinguistic profile 112 and preferably, in parallel or additionally, performs a classification into a tactical category 116.

[0070] A decision module 120 receives the tactics category 116 and the psycholinguistic profile 112 and determines a response strategy 118. This is passed to a generation unit 124, which generates a suggested response 122. The suggested response is then output via an output unit 126.

[0071] The entire System 100 can be modularly structured, with components communicating via REST interfaces and operating as decoupled microservices. Processing can be local, server-based, or cloud-based. As an example, the S1-S7 process is executed in a modular container architecture implemented on a single compute node with an NVIDIA A100 GPU (40 GB HBM2e), 32 CPU cores, and 256 GB RAM running Ubuntu 22.04 LTS. Each process stage—from real-time acquisition of the negotiation data stream (S1), feature extraction (S2), psycholinguistic profiling (S3), tactic classification (S4), and strategy determination (S5), to the generation (S6) and output (S7) of the proposed response—is preferably implemented as a decoupled microservice in a Docker container and dynamically scaled using Kubernetes.Using a multimodal transformer network trained on, for example, 50,000 negotiation-specific annotated utterances, which fuses prosodic x-vector embeddings with a finely tuned BERT language model, an end-to-end throughput of at least 120 negotiation turns per second is achieved with a total latency of no more than 350 ms. All source code, hyperparameters, and pre-trained weights are stored in a publicly referenced Git repository. Reference symbol list 100 System for process execution 101 Evaluation and computing unit 102 Input unit for recording the negotiation data stream 104 Negotiation data stream 106 audio and / or text data 108 Feature vector 110 Analysis component for feature extraction 112 Psycholinguistic Profile 114 Machine learning model 116 Tactics category (e.g. aggression, de-escalation, concealment) 118 Response strategy 120 Decision Module 122 possible answers (textual and / or linguistic) 124 Generation unit for generating the suggested answer 126 Output unit (e.g. screen, speaker) S1 Capturing the negotiation data stream S2 Extracting the feature vector S3 Determining the psycholinguistic profile S4 Classifying the negotiation data stream into a tactics category S5 Determining a response strategy S6 Generating a suggested answer S7 Output of the suggested answer

Claims

A computer-implemented method for the tactical analysis of a negotiating partner's utterances and / or for the automatic generation of a strategically coordinated response proposal, comprising the steps: S1: capturing a negotiation data stream (104), in particular comprising audio and / or text data (106), of a negotiating partner in real time by an input unit (102); S2: extracting a feature vector (108) from the negotiation data stream (104) by an analysis component (110), whereby prosodic features, in particular a fundamental frequency and / or a volume and / or a speaking rate, and / or lexical-syntactic features are formed; S3: determining a psycholinguistic profile (112) of the negotiating partner using the feature vector (108) by means of at least one machine learning model (114) implemented in a processing and computing unit (100);S4: Classifying the negotiation data stream (104) into at least one tactic category (116), in particular exhibiting aggression, de-escalation, neutrality, decisiveness, or obfuscation, using the at least one machine learning model (114), in particular also using the psycholinguistic profile (112); S5: Determining a response strategy (118) by a decision module (120) that takes into account both the tactic category (116) and the psycholinguistic profile (112); S6: Generating a textual and / or linguistic response proposal (122) based on the determined response strategy (118) by a generation unit (124); and S7: Outputting the response proposal (122) to an output unit (126), wherein steps S1 to S7 are preferably performed iteratively for the received negotiation data stream (104). Method according to claim 1, wherein the at least one machine learning model (114) comprises a multimodal transformer network that fuses the prosodic and lexical-syntactic features extracted in step S2 into a common representation. Method according to one of claims 1 or 2, wherein the psycholinguistic profile (112) comprises Big Five personality dimensions and / or LIWC categories and is determined in particular by fine-tuning a BERT language model. Method according to one of claims 1 to 3, wherein the decision module (120) correlates a weighted confidence rating of the tactic category (116) determined in step S4 with at least one personality parameter of the psycholinguistic profile (112). Method according to any one of claims 1 to 4, wherein the proposed response (122) comprises rhetorical means suitable for de-escalating a statement classified as aggression and / or exposing a statement classified as obfuscation and / or responding to a statement classified as definiteness. Method according to any one of claims 1 to 5, wherein the latency between receiving a utterance and outputting the suggested response (122) is at most 500 ms. Method according to one of claims 1 to 6, wherein user feedback on the effectiveness of the suggested answer (122) is recorded and the parameters of the at least one machine learning model (114) are adjusted online by means of reinforcement learning. Method according to one of claims 1 to 7, wherein additionally video data from the user and / or the negotiating partner are recorded and features from facial expression and head posture are included in the classification of the tactic category (116) according to step S4. Method according to any one of claims 1 to 8, wherein the at least one machine learning model (114) comprises an adversarially trained submodel to distinguish obfuscation strategies from inconsistent or misleading statements. Method according to one of claims 1 to 9, wherein the at least one machine learning model (114) is trained on a corpus of at least 1,000, preferably at least 10,000, negotiation-specific annotated utterances per tactic category (116). Method according to any one of claims 1 to 10, wherein the modules executed in steps S2 to S6 are implemented as decoupled micro-services in containers and communicate via a REST interface. Method according to any one of claims 1 to 11, wherein the proposed answer (122) according to step S6 is generated by means of a finely tuned language model trained on a corpus of tactical negotiation dialogues. Computer program product comprising instructions which, when the program is executed by a computer, cause it to perform the steps of the method according to any one of claims 1 to 12. Computer-readable data carrier on which the computer program product according to claim 13 is stored. System (100) for the tactical analysis of utterances of a negotiating partner and for the automatic generation of a strategically coordinated response proposal, wherein the system (100) comprises an evaluation and computing unit (101) trained to perform the following steps: S1: Capture a negotiation data stream (104), in particular comprising audio and / or text data (106), of a negotiating partner in real time via an input unit (102); S2: Extract a feature vector (108) from the negotiation data stream (104), comprising prosodic features, in particular a fundamental frequency and / or a volume and / or a speaking rate, and lexical-syntactic features, by means of an analysis component (110); S3: Determine a psycholinguistic profile (112) of the negotiating partner using the feature vector (108) by means of at least one machine learning model (114);S4: Classifying the negotiation data stream (104) into at least one tactical category (116), in particular exhibiting aggression, de-escalation, neutrality, decisiveness, or obfuscation, using at least one machine learning model (114); S5: Determining a response strategy (118) by a decision module (120) that takes into account both the tactical category (116) and the psycholinguistic profile (112); S6: Generating a textual and / or linguistic response proposal (122) based on the determined response strategy (118) by a generation unit (124); S7: Outputting the response proposal (122) to an output unit (126).

Images