Anthropomorphic dialogue response control method based on emotion time sequence joint modeling
By jointly modeling sentiment and temporal sequence, we can extract text sentiment and user behavior features, predict the optimal response interval, solve the shortcomings of response timing control in existing technologies, and achieve humanization and robustness improvement of the dialogue system.
Patent Information
- Application Number
- CN202511759087.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-02-13
AI Technical Summary
Existing dialogue response control methods struggle to simultaneously address short-term burst signals and medium- to long-term interaction patterns in complex scenarios. They lack personalization and context sensitivity, making it difficult to achieve appropriate response timing control.
We employ a sentiment-temporal joint modeling approach to extract text sentiment features and user input behavior features. By combining dialogue turns, we use causal convolution and mixed density modeling to predict the optimal response interval and generate anthropomorphic dialogue responses.
It achieves accurate prediction of response timing and adaptive adjustment, enhancing the flexibility and robustness of the dialogue and improving the naturalness and anthropomorphism of the interaction.
Smart Images
Figure CN121525733A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of dialogue response control, in particular to a humanized dialogue response control method based on emotion and time sequence joint modeling. BACKGROUND
[0002] Dialogue response control plays a core role in human-computer interaction systems, and its rationality directly affects the fluency of interaction, user satisfaction and system trust. A suitable reply time can make the system more "humanized", providing timely comfort or handling in emergency or emotional scenarios, and can also avoid frequent or abrupt intervention, which may cause user boredom or interrupt user thinking. Therefore, the reply time is one of the important dimensions of dialogue experience.
[0003] Current methods for implementing dialogue response control mainly include rule-based timing strategies, simple threshold or heuristic mapping, and supervised learning methods based on classification / regression. Rule-based or threshold-based methods are simple to implement and easy to interpret, but lack personalization and context sensitivity, making it difficult to respond to complex scenarios and diverse user behaviors. Supervised learning methods can adapt to some extent, but often use a single output form, ignoring the possible multimodal and uncertainty of the reply time itself. In addition, many methods lack sufficient time-dependent modeling of historical context, making it difficult to consider both short-term sudden signals and long-term interaction patterns, thereby limiting the capture of complex dialogue dynamics.
[0004] Currently, the introduction of representation learning, rich behavior feature engineering and probabilistic output technology can significantly improve the response control capability. Specifically, emotion and semantic representations based on pre-trained language models can provide high-quality encoding of text emotional tendencies and fine-grained semantics; behavior feature engineering structures interactive signals other than text into stable vectors, which helps to reflect the user's expression habits and urgency tendencies; and probabilistic output models can express the uncertainty and multimodal possibilities of the prediction in a distribution form, providing confidence measures for subsequent decision-making and supporting more robust deterministic decoding or strategy rollback. These technologies complement each other and have stronger adaptability and reliability in complex real-world scenarios. SUMMARY
[0005] The present application proposes a humanized dialogue response control method based on emotion and time sequence joint modeling, which extracts text emotion features and user input behavior features, and combines dialogue turns to form a time sequence input. The optimal reply interval is predicted using causal convolution and hybrid density modeling, thereby achieving intelligent control of the dialogue system in reply time and improving the naturalness and humanization of the interaction.
[0006] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:
[0007] Step 1: Obtain multi-turn dialogue dataset.
[0008] Step 2: Preprocess the obtained multi-turn dialogue data, parse the raw data into a structured format, and then clean the dialogue text to eliminate invalid information in the text.
[0009] Step 3: Extract sentiment vectors and user input behavior feature vectors from structured user input text.
[0010] Step 4: Group the sentiment vectors, behavior vectors, and dialogue turns in chronological order to form an input sequence, and input it into a time series model to predict the optimal reply interval from the output layer.
[0011] Step 5: Input the predicted interval as a control signal into the reply generation module to generate the final reply.
[0012] 1. In the above step 1, the DailyDialog multi-turn dialogue dataset is selected. DailyDialog is a high-quality multi-turn open-domain dialogue dataset widely used in dialogue system research, containing 13118 dialogues with an average of 8 speaking turns per dialogue. The dataset has a concentrated theme and standardized grammar, with annotated intent and sentiment for each sentence.
[0013] 2. In the above step 2, the DailyDialog raw data is parsed into a structured format, and the mapping relationship between "dialogue ID - session - turn - text - original sentiment label" is determined. At the same time, the text is cleaned by deleting special symbols, meaningless words, and redundant spaces and line breaks to ensure the accuracy of subsequent sentiment feature extraction and time series simulation.
[0014] 3. In the above step 3, sentiment feature extraction includes the following steps:
[0015] (1) First, perform sub-word tokenization on each round of user input text and map it to an embedding sequence. Use a pre-trained Transformer text encoder to encode the embedding sequence into a contextual token representation. For each token representation Calculate attention score:
[0016]
[0017] where, is the attention score vector, is the attention projection matrix, is the bias.
[0018] (2) Then normalize each token attention score to attention weight:
[0019]
[0020] (3) Then, the token is represented by attention weighting. Compressed into a one-dimensional emotion vector:
[0021]
[0022] 4. In step 3 above, the extraction of user input behavior features includes the following steps:
[0023] (1) First, extract quantifiable behavioral signals from each round of user input. These signals are different from the semantic content of the text, but are related to the input form and context, including the length of the input token. Does it contain a question mark? Does it contain an exclamation mark? Interval with the previous response User speaking frequency The above features are combined into the original behavior vector:
[0024]
[0025] (2) To avoid training instability caused by different units of measurement, the numerical features are standardized. The following is the calculation formula:
[0026]
[0027] in, , These are the training set mean and standard, respectively.
[0028] (3) For categorical features such as dialogue behavior categories, embedding mapping is performed first, and then the numerical features are concatenated to obtain a unified input vector. Then, it is input into a lightweight feedforward neural network for nonlinear mapping, and the output is a fixed-dimensional behavior embedding vector. The calculation process is as follows:
[0029]
[0030] (4) The final behavior vector representation is obtained through the projection layer:
[0031]
[0032] in, , All of these are trainable parameters. For the input feature dimension, For hidden layer dimensions, For output dimensions.
[0033] 5、In step 4 above, the reply interval is generated comprising the following steps:
[0034] (1) For the obtained emotion vector and behavior vector , embed them into the round , concatenate them into a fusion vector in a fixed order , to eliminate the scale difference of different source features and facilitate joint processing, apply a linear mapping to the concatenated vector to obtain a unified dimension representation:
[0035]
[0036] wherein and are the projection matrix and bias of the defined dimension.
[0037] (2) Then form a sequence from the first round to the Tth round as the input of the causal time sequence network, which adopts a causal dilated convolution network (TCN) to ensure that the output at any time only depends on the input at this time and before and can capture local and longer historical dependencies at the same time. Specifically, the TCN is composed of layers of one-dimensional causal convolution, and the th layer performs convolution calculation at time , which can be expressed as
[0038]
[0039] wherein , is the convolution kernel width, is the dilated coefficient of the th layer, which increases in an exponential sequence , and the output of each layer is connected by activation and residual to obtain a hidden sequence , and the hidden representation at the last time is taken as the prediction as a comprehensive expression of the emotion-behavior-round time sequence information up to the round .
[0040] (3) To facilitate the generation of component parameters required by the mixture density network (MDN), three independent linear transformations are used to map to the original log score vector of the component weight , the original vector of the component mean , and the original vector of the component scale .
[0041] (4) Then apply a deterministic mapping to the original parameters to obtain the effective MDN parameters, and the component weight, the component mean, and the component standard deviation can be obtained by the following mapping respectively:
[0042]
[0043]
[0044]
[0045] Thus the mixed Gaussian distribution of the reply interval is formed:
[0046]
[0047] (5) In order to obtain specific interval values directly usable for the dialogue, a deterministic decoding method is adopted, and the weighted mean of the mixed distribution is taken as the output, that is, the following formula is used to output the optimal reply interval:
[0048]
[0049] The beneficial effects of the present application are:
[0050] (1) The present application can simultaneously fuse the emotional features, behavioral features and dialogue turn information of user input, so that the prediction of the reply time is more comprehensive and accurate, and the limitations brought by single feature modeling are avoided;
[0051] (2) By introducing the time series modeling method, the dynamic evolution of the user state in the dialogue process is effectively described, so that the reply interval can be adaptively adjusted according to the context change, and is closer to the real interactive habit;
[0052] (3) The probabilistic modeling method is adopted to express the multi-modal characteristics of the reply interval, which can not only output deterministic results, but also provide uncertainty reference, thereby enhancing the flexibility and robustness of the dialogue;
[0053] (4) The output result can be directly used for reply scheduling or generation control, which is convenient for integration with existing dialogue systems. BRIEF DESCRIPTION OF DRAWINGS
[0054] Figure 1 The flow chart of the humanized dialogue response control method based on emotion time series joint modeling; DETAILED DESCRIPTION
[0055] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application, and after reading the present application, various equivalent modifications of the present application by those skilled in the art fall within the scope defined by the appended claims.
[0056] The flow chart of the humanized dialogue response control method based on emotion time series joint modeling is as shown in Figure 1 The steps include:
[0057] Step 1: Obtain multi-turn dialogue dataset.
[0058] Step 2: Preprocess the obtained multi-turn dialogue data, parse the raw data into a structured format, and then clean the dialogue text to eliminate invalid information in the text.
[0059] Step 3: Extract sentiment vectors and user input behavior feature vectors from structured user input text.
[0060] Step 4: Group the sentiment vectors, behavior vectors, and dialogue turns in chronological order to form an input sequence, and input it into a time series model to predict the optimal reply interval from the output layer.
[0061] Step 5: Input the predicted interval as a control signal into the reply generation module to generate the final reply.
[0062] 1. In step 1 above, the DailyDialog multi-turn dialogue dataset is selected. DailyDialog is a high-quality multi-turn open-domain dialogue dataset widely used in dialogue system research, containing 13118 dialogues with an average of 8 speaking turns per dialogue. The dataset is focused on a specific topic and has a standard syntax, with annotated intent and sentiment for each sentence.
[0063] 2. In step 2 above, the DailyDialog raw data is parsed into a structured format, with a clear mapping relationship between "dialogue ID - session - turn - text - original sentiment label". At the same time, the text is cleaned by deleting special symbols, meaningless words, and redundant spaces and line breaks to ensure the accuracy of subsequent sentiment feature extraction and time series simulation.
[0064] 3. In step 3 above, sentiment feature extraction includes the following steps:
[0065] (1) First, perform sub-word tokenization on each round of user input text and map it to an embedding sequence. Use a pre-trained Transformer text encoder to encode the embedding sequence into a contextual token representation. For each token representation Calculate attention score:
[0066]
[0067] where, is the attention score vector, is the attention projection matrix, is the bias.
[0068] (2) Then normalize each token attention score to attention weight:
[0069]
[0070] (3) Then, the token is represented by attention weighting. Compressed into a one-dimensional emotion vector:
[0071]
[0072] 4. In step 3 above, the extraction of user input behavior features includes the following steps:
[0073] (1) First, extract quantifiable behavioral signals from each round of user input. These signals are different from the semantic content of the text, but are related to the input form and context, including the length of the input token. Does it contain a question mark? Does it contain an exclamation mark? Interval with the previous response User speaking frequency The above features are combined into the original behavior vector:
[0074]
[0075] (2) To avoid training instability caused by different units of measurement, the numerical features are standardized. The following is the calculation formula:
[0076]
[0077] in, , These are the training set mean and standard, respectively.
[0078] (3) For categorical features such as dialogue behavior categories, embedding mapping is performed first, and then the numerical features are concatenated to obtain a unified input vector. Then, it is input into a lightweight feedforward neural network for nonlinear mapping, and the output is a fixed-dimensional behavior embedding vector. The calculation process is as follows:
[0079]
[0080] (4) The final behavior vector representation is obtained through the projection layer:
[0081]
[0082] in, , All of these are trainable parameters. For the input feature dimension, For hidden layer dimensions, For output dimensions.
[0083] 5、The reply interval in step 4 above is generated comprising the following steps:
[0084] (1) For the obtained emotion vector and behavior vector , embed them into the round , concatenate them into a fusion vector in a fixed order To eliminate the scale difference of features from different sources and facilitate joint processing, apply a linear mapping to the concatenated vector to obtain a unified dimensional representation:
[0085]
[0086] where and are the projection matrix and bias defining the dimension.
[0087] (2) Then form a sequence from the first round to the Tth round as the input of a causal temporal network, which adopts a causal dilated convolutional network (TCN) to ensure that the output at any time only depends on the input at that time and before and can capture both local and longer historical dependencies. Specifically, the TCN is composed of layers of one-dimensional causal convolution, and the convolution calculation of the th layer at time can be expressed as
[0088]
[0089] where , is the convolution kernel width, is the dilation coefficient of the th layer, which increases in an exponential sequence , and the output of each layer is activated and connected by residual to obtain a hidden sequence , and the hidden representation at the last time is taken as the comprehensive expression of the emotion-behavior-round temporal information up to round .
[0090] (3) To facilitate the direct generation of component parameters required by the mixed density network (MDN), three independent linear transformations are used to map to the original log score vector of component weights , the original vector of component means , and the original vector of component scales .
[0091] (4) Subsequently deterministic mapping is applied to the original parameters to obtain the effective MDN parameters, the component weight, component mean and component standard deviation can be obtained by the following mapping respectively:
[0092]
[0093]
[0094]
[0095] Thus the mixed Gaussian distribution of the reply interval is formed:
[0096]
[0097] (5) In order to obtain the specific interval value directly usable for the dialogue, a deterministic decoding method is adopted, and the weighted mean of the mixed distribution is used as the output, that is, the following formula is used to output the optimal reply interval:
[0098]
Claims
1. A method for controlling a human-like dialogue response based on emotion timing joint modeling, characterized in that: The method comprises the following steps: Step 1: Obtain a multi-turn dialogue data set. Step 2: Preprocess the obtained multi-turn dialogue data, parse the original data into a structured format, and then clean the dialogue text to eliminate invalid information in the text. Step 3: Extract the emotion vector and user input behavior feature vector from the structured user input text. Step 4: Group the emotion vector, behavior vector, and dialogue turn encoding in time sequence to form an input sequence, input it into the time sequence model, and predict the optimal reply interval from the output layer. Step 5: Input the predicted interval as a control signal into the reply generation module to generate the final reply.
2. The method of claim 1, wherein the method further comprises: determining a sentiment of the user input; and determining a sentiment of the response. In step 1, the DailyDialog multi-turn dialogue data set is selected. DailyDialog is a high-quality multi-turn open-domain dialogue data set widely used in dialogue system research, containing 13118 dialogues, with an average of 8 speaking turns per dialogue. The data set is concentrated in theme and has a standard syntax, with annotated intent and emotion for each sentence.
3. The method of claim 1, wherein the method further comprises: determining a sentiment of the user input; and determining a sentiment of the response. In step 2, the DailyDialog original data is parsed into a structured format, and the mapping relationship of "dialogue ID - session - turn - text - original emotion label" is determined. At the same time, the text is cleaned regularly, and special symbols, meaningless words, and redundant spaces and line breaks are deleted to ensure the accuracy of subsequent emotion feature extraction and time sequence simulation.
4. The method of claim 1, wherein the method further comprises: determining a sentiment of the user input; and determining a sentiment of the response. In step 3, emotion feature extraction includes the following steps: (1) For each round of user input text, first perform sub-word tokenization and map to embedding sequences, use a pre-trained Transformer text encoder to encode the embedding sequences into contextualized token representations, and for each token representation Compute attention scores: wherein, is an attention score vector, is an attention projection matrix, is a bias. (2) Each token attention score is then normalized to an attention weight: normalized to an attention weight: (3) token representations are then compressed into one-dimensional sentiment vectors by attention weighting compressed into one-dimensional sentiment vectors 5. The method of claim 1, wherein the method further comprises: determining a sentiment of the user input; and determining a sentiment of the response. In step 3, user input behavior feature extraction includes the following steps: (1) First, we extract quantifiable behavioral signals from each user's input, which are different from the semantic content of the text, but are related to the input form and context, including the token length of the input , whether it contains a question mark , whether it contains an exclamation mark , the interval between the previous round of replies , the user's speaking frequency . The above features are combined into the original behavior vector: (2) To avoid training instability caused by different dimensions, numerical features are standardized. The calculation formula is as follows: wherein, , are the training set mean and standard deviation, respectively. (3) For the category feature of dialogue behavior category, first, embedding mapping is performed, and then it is spliced with numerical features to obtain a unified input vector Then, it is input into a light feedforward neural network for nonlinear mapping, and a fixed-dimensional behavior embedding vector is output. The calculation process is represented as follows: (4) Then, the final behavior vector representation is obtained through the projection layer: wherein, , are trainable parameters, is the input feature dimension, is the hidden layer dimension, is the output dimension.
6. The method of claim 1, wherein the method further comprises: In step 4, reply interval generation includes the following steps: (1) For the obtained emotion vector and behavior vector , embed them with the round Concatenate them in a fixed order to obtain a fusion vector To eliminate the scale difference of features from different sources and facilitate joint processing, a linear mapping is applied to the concatenated vector to obtain a unified dimension representation: wherein with is a projection matrix defining the dimension and a bias. (2) Then a sequence formed by the first round to the Tth round As input of the causal temporal network, the causal temporal network adopts a causal dilated convolution network (TCN) to ensure that the output of any moment only depends on the input of this moment and before and can capture local and longer historical dependencies at the same time. Specifically, the TCN is composed of layers of one-dimensional causal convolution, the first layer is a one-dimensional causal convolution layer, and the last convolution calculation at the moment in, , The width of the convolution kernel. For the first The expansion coefficient of the layer, in exponential sequence Incremental, the outputs of each layer are connected through activation and residual connections to obtain the hidden sequence. And take the hidden representation at the last moment during prediction. As of the end of the round It is a comprehensive expression of emotional, behavioral, and sequential information. (3) To facilitate the direct generation of component parameters required for a mixture density network (MDN), three sets of independent linear transformations are respectively used to map the original log score vector of component weights , the original vector of component means , and the original vector of component scales , and . (4) Then, deterministic mapping is applied to the original parameters to obtain effective MDN parameters. The component weight, component mean, and component standard deviation can be obtained using the following mapping respectively: This forms a mixed Gaussian distribution of the reply interval: (5) To obtain specific interval values directly usable for dialogue, a deterministic decoding method is used, with the weighted mean of the mixed distribution as the output. The optimal reply interval is output using the following formula: