Large model dialogue system for psychotherapy healing

Abstract

The invention relates to the field of treatment healing dialogues, and particularly discloses a large model dialogue system for psychological treatment healing, which is characterized in that word segmentation processing of consultation texts is carried out by using a deep learning-based natural language processing technology and a coding mode, and then semantic embedding coding is carried out on word granularity description of each consultation text after word segmentation, so that a large model dialogue is obtained; and intelligently generating a recommended dialogue strategy tag according to the feature jump semantic propagation representation between the word granularity description semantic embedding features of each consultation text, and obtaining a dialogue text based on the tag and the consultation text. According to the method, the intention and the emotional state of the user can be better understood, and according to the real-time emotional state and the consultation content of the user, a common-situation or question-asking strategy is flexibly selected, so that a more personalized response is provided, and the problem of improper question asking or insufficient common-situation asking in a traditional system is avoided.

Description

Technical Field

[0001] The present application relates to the field of therapeutic dialogue, and more specifically, to a large model dialogue system for psychological therapy. Background Art

[0002] In today's society, fast-paced life, high-intensity work and complex interpersonal relationships have made people face unprecedented psychological pressure. Seeking psychological treatment is no longer a niche demand, but has become an urgent need for many people to relieve anxiety, resolve troubles and maintain mental health.

[0003] However, existing psychological healing systems generally have shortcomings, lack an effective dialogue guidance mechanism, and are difficult to flexibly adjust responses according to the specific circumstances of users. For example, they fail to provide sufficient support when users need empathy, or ask inappropriate questions that interrupt the user's expression. In addition, the dialogue texts generated by such systems are often stereotyped, lacking pertinence and emotional resonance, and are difficult to meet individual needs, which not only weakens the user experience, but also affects the healing effect.

[0004] Therefore, a large-scale model dialogue system for psychological healing is desired. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed.

[0006] According to one aspect of the present application, a large model dialogue system for psychological therapy is provided, which includes: a consultation audio signal acquisition module, which is used to acquire a consultation audio signal input by a user; a signal recognition module, which is used to perform speech recognition on the consultation audio signal to obtain a consultation text; a consultation text word granularity encoding module, which is used to perform word segmentation processing and semantic embedding encoding on the consultation text to obtain a sequence distribution of consultation text word granularity description semantic embedding encoding features; a consultation text semantic propagation module, which is used to perform consultation text semantic message propagation with feature jump degree on the sequence distribution of the consultation text word granularity description semantic embedding encoding features to obtain consultation text semantic propagation features, wherein the consultation text semantic propagation module includes: a message propagation weight meter A calculation unit is used to calculate the message propagation significance weight of each consulting text word granularity description semantic embedding coding feature in the sequence distribution of the consulting text word granularity description semantic embedding coding feature to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight; a message propagation aggregation unit is used to perform propagation aggregation on the sequence distribution of the consulting text word granularity description semantic embedding coding feature based on the sequence distribution of the consulting text word granularity description message propagation significance weight to obtain the consulting text semantic propagation feature; a dialogue strategy label generation module is used to obtain a recommended dialogue strategy label based on the consulting text semantic propagation feature; a dialogue text generation module is used to obtain a dialogue text based on the consulting text and the recommended dialogue strategy label.

[0007] Compared with the prior art, the present application provides a large-model dialogue system for psychological therapy, which obtains the consultation audio signal input by the user and performs speech recognition to obtain the consultation text, uses natural language processing technology and encoding methods based on deep learning to perform word segmentation processing on the consultation text, and then performs semantic embedding encoding on the word granularity descriptions of each consultation text after word segmentation, so as to intelligently generate recommended dialogue strategy labels based on the feature jump semantic propagation representation between the semantic embedding features of the word granularity descriptions of each consultation text, and obtain the dialogue text based on the label and the consultation text. The present application can better understand the user's intentions and emotional state, and flexibly select empathy or questioning strategies according to the user's real-time emotional state and consultation content, thereby providing a more personalized response, ensuring that each interaction is appropriate and productive, and avoiding the problems of inappropriate questions or insufficient empathy in traditional systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 It is a block diagram of a large model dialogue system for psychological therapy according to an embodiment of the present application.

[0010] Figure 2 This is a block diagram of a consultation text word granularity encoding module in a large model dialogue system for psychological therapy according to an embodiment of the present application.

[0011] Figure 3 This is a block diagram of a consultation text semantic propagation module in a large model dialogue system for psychological therapy according to an embodiment of the present application.

[0012] Figure 4 It is a block diagram of a message propagation weight calculation unit in a large model dialogue system for psychological therapy according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] The embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments described herein, but rather these embodiments are provided for a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0014] It should be understood that the various steps described in the method embodiments of the present disclosure may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present disclosure is not limited in this respect.

[0015] It is worth noting that in this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization of the owner of the corresponding device.

[0016] In today's society, fast-paced life, high-intensity work and complex interpersonal relationships have made people face unprecedented psychological pressure. Seeking psychological treatment is no longer a niche demand, but has become an urgent need for many people to relieve anxiety, resolve troubles and maintain mental health.

[0017] However, existing psychological healing systems generally have shortcomings, lack an effective dialogue guidance mechanism, and are difficult to flexibly adjust responses according to the specific circumstances of users. For example, they fail to provide sufficient support when users need empathy, or ask inappropriate questions that interrupt the user's expression. In addition, the dialogue texts generated by such systems are often stereotyped, lacking pertinence and emotional resonance, and are difficult to meet individual needs, which not only weakens the user experience, but also affects the healing effect.

[0018] In response to the above technical problems, the technical concept of the present application is to obtain the consultation audio signal input by the user, and perform speech recognition on it to obtain the consultation text, use natural language processing technology and encoding methods based on deep learning to perform word segmentation processing on the consultation text, and then perform semantic embedding encoding on the word granularity descriptions of each consultation text after word segmentation, so as to intelligently generate recommended dialogue strategy labels based on the feature jump semantic propagation representation between the semantic embedding features of the word granularity descriptions of each consultation text, and obtain the dialogue text based on the label and the consultation text. The present application can better understand the user's intentions and emotional state, and flexibly select empathy or questioning strategies according to the user's real-time emotional state and consultation content, thereby providing a more personalized response, ensuring that each interaction is appropriate and productive, and avoiding the problems of inappropriate questions or insufficient empathy in traditional systems.

[0019] Figure 1 is a block diagram of a large model dialogue system for psychological therapy according to an embodiment of the present application. Specifically, Figure 1 As shown, according to the large-model dialogue system 100 for psychological therapy of the embodiment of the present application, it includes: a consultation audio signal acquisition module 110, which is used to acquire the consultation audio signal input by the user; a signal recognition module 120, which is used to perform speech recognition on the consultation audio signal to obtain a consultation text; a consultation text word granularity encoding module 130, which is used to perform word segmentation processing and semantic embedding encoding on the consultation text to obtain a sequence distribution of semantic embedding encoding features of the consultation text word granularity description; a consultation text semantic propagation module 140, which is used to perform consultation text semantic message propagation with feature jump degree on the sequence distribution of the semantic embedding encoding features of the consultation text word granularity description to obtain the consultation text semantic propagation features; a dialogue strategy label generation module 150, which is used to obtain a recommended dialogue strategy label based on the consultation text semantic propagation features; a dialogue text generation module 160, which is used to obtain a dialogue text based on the consultation text and the recommended dialogue strategy label.

[0020] Specifically, the consultation audio signal acquisition module 110 is used to obtain the consultation audio signal input by the user. It should be understood that the consultation audio signal input by the user is obtained. It should be understood that the consultation audio signal input by the user refers to the audio data in the form of an electronic signal formed by the user entering the words he wants to express into the system in the form of sound through relevant audio acquisition equipment (such as a mobile phone microphone, a computer external or built-in microphone, etc.), which carries the user's various demands, ideas, emotional descriptions and other contents in the psychological healing scenario. Therefore, in order to create a cordial atmosphere similar to face-to-face chatting, in the technical solution of the present application, the consultation audio signal input by the user is obtained.

[0021] Users usually use common audio acquisition devices to input consultation audio signals. For example, mobile phone microphones are one of the most convenient and widely used methods. In modern smartphones, microphones are carefully designed and optimized to capture relatively clear sound signals. The acoustic sensor inside has a certain sensitivity and frequency response range, which can generally cover the main frequency range of human speech, ranging from tens of hertz to several thousand hertz. When the user opens the relevant psychological healing application and grants microphone permission, the mobile phone microphone starts working and converts the sound in the surrounding environment into electrical signals. Computer external or built-in microphones are also commonly used audio input devices. External microphones may have higher flexibility in performance. Users can choose different types of microphones according to their needs. For example, directional microphones can capture the user's voice more accurately and reduce the interference of surrounding noise; while built-in microphones are integrated into the computer host or laptop. Although in some cases they may be affected by the slight noise generated by the electronic components inside the computer, they can also meet the basic requirements for general consultation audio input. Regardless of the type of microphone, it is necessary to ensure its compatibility with the system at the hardware level, and transmit the collected audio signals to the system through the corresponding audio interface or wireless connection technology (such as Bluetooth).

[0022] The audio driver is responsible for communicating with the hardware device to ensure that the audio signal can be correctly recognized and acquired by the system. It will perform analog-to-digital conversion on the original analog audio signal collected by the microphone and convert it into a digital signal for subsequent processing. At the same time, the signal processing software module will perform some basic pre-processing operations on the digital audio signal, such as noise reduction. In the actual environment, there will inevitably be various background noises, such as noisy human voices in the environment, the sound of machine equipment running, etc. The noise reduction algorithm can identify the noise signal and weaken it by analyzing the frequency characteristics and amplitude changes of the audio signal, thereby improving the quality of the audio signal. Common noise reduction methods include filter-based noise reduction technology, such as high-pass filters, low-pass filters, and band-pass filters. According to the frequency distribution characteristics of the noise, a suitable filter can be selected to remove the noise; there are also noise reduction methods based on signal statistical characteristics, such as adaptive noise reduction algorithms, which can dynamically adjust the noise reduction parameters according to the real-time changes of the audio signal to achieve better noise reduction effects.

[0023] In addition, in order to ensure the integrity and accuracy of the audio signal, the system will also perform sampling and quantization operations on the audio signal. Sampling refers to the discrete value of the audio signal within a certain time interval, and the selection of the sampling frequency is crucial. According to the Nyquist sampling theorem, in order to accurately restore the audio signal, the sampling frequency should be at least twice the highest frequency of the audio signal. For human voice signals, 8kHz, 16kHz or higher sampling frequencies are generally used. Quantization converts the discrete values ​​obtained by sampling into binary digital representation, and the quantization accuracy determines the resolution of the audio signal. Higher quantization accuracy can more accurately represent the amplitude changes of the audio signal, but it will also increase the amount of data. Usually, 16-bit or 24-bit quantization accuracy is used to balance the needs of audio quality and data storage and processing.

[0024] In terms of data transmission, the audio signal that has been initially processed needs to be transmitted to the subsequent speech recognition module through the system's internal communication mechanism. If the system is based on a network architecture, such as a psychological therapy system running on a cloud server, the audio signal also needs to be securely and stably transmitted to the server through a network transmission protocol (such as HTTP, WebSocket, etc.). During the transmission process, data encryption technology is needed to protect the user's privacy and prevent the audio signal from being stolen or tampered with during transmission. Common encryption algorithms such as the SSL / TLS protocol can encrypt and decrypt data at the network transmission layer to ensure data security.

[0025] At the same time, the system also needs to perform some monitoring and management operations on the input of audio signals. For example, the system detects whether the strength of the audio signal is within a reasonable range. If the signal is too weak, it may cause inaccurate speech recognition. The system can prompt the user to adjust the microphone position or increase the speaking volume; if the signal is too strong, audio distortion may occur, and the system also needs to perform corresponding processing, such as automatically reducing the gain or prompting the user to lower the volume. In addition, the system can also monitor the duration of audio input to prevent users from wasting system resources or abnormal situations due to long-term inactivity or long audio input. In terms of user experience, in order to allow users to input consulting audio signals more naturally and comfortably, the interface design of the system also needs to consider the convenience and guidance of audio input. For example, clear audio input buttons and prompt information are provided on the interface of the application to inform users how to use the microphone correctly for recording, and how to pause, continue or stop recording during recording. At the same time, some real-time audio feedback can also be provided, such as audio waveform display or volume indicator bar, so that users can intuitively understand their audio input.

[0026] Specifically, the signal recognition module 120 is used to perform speech recognition on the consultation audio signal to obtain a consultation text. Accordingly, considering that the speech signal is a continuous acoustic waveform, it is difficult for a computer to deeply understand the semantic, logical and other information contained therein directly based on such a raw audio form. Therefore, in order to more clearly understand the semantic meaning expressed by the user and to accurately grasp what the user wants to express, the present application performs speech recognition on the consultation audio signal to convert complex semantic information into a machine-readable text form to obtain a consultation text, thereby accurately capturing the core meaning of the user's expression and ensuring that subsequent psychological healing conversations are both appropriate and fruitful.

[0027] In the signal preprocessing stage, the consultation audio signal obtained from the audio acquisition device will undergo a series of optimization operations. Since the collected audio signal may have varying degrees of noise interference, noise reduction processing is required. Common noise reduction methods include spectral subtraction-based technology, which estimates the noise spectrum and subtracts it from the original audio spectrum to reduce the impact of noise on the speech signal. At the same time, audio enhancement processing will also be performed, such as using adaptive gain control to adjust the amplitude of the audio signal to ensure that the speech part is within the appropriate volume range for subsequent processing.

[0028] Next is the feature extraction stage. Speech signals have rich acoustic features, among which the most commonly used is the Mel frequency cepstral coefficient (MFCC). When extracting MFCC, the audio signal is first framed. Usually, the duration of each frame is about 20-30 milliseconds, and the frame shift is 10-15 milliseconds. In this way, the continuous speech signal can be divided into a series of short-term signal fragments for easy analysis. Then, a fast Fourier transform (FFT) is performed on each frame signal to convert the time domain signal into a frequency domain signal to obtain the signal's spectrum information. Then, according to the human ear's auditory perception characteristics, the spectrum is filtered through a Mel filter bank to obtain a Mel spectrum. Finally, the Mel spectrum is logarithmized and discrete cosine transform (DCT) is performed to obtain the MFCC feature vector. These feature vectors can effectively characterize the acoustic characteristics of speech signals and provide key input information for subsequent recognition models.

[0029] In terms of model training, deep neural networks (DNNs) play a central role in speech recognition. Usually, architectures based on recurrent neural networks (RNNs) or long short-term memory networks (LSTMs) are used. During the training process, a large amount of speech data is required to learn the patterns and regularities of speech. These data cover speech samples of different genders, ages, accents, and language environments to ensure that the model has a wide range of adaptability. The training data will be annotated with corresponding text content, and the model will continuously adjust its parameters by minimizing the error between the predicted text and the annotated text. For example, during the training process, the cross entropy loss function may be used to measure the accuracy of the prediction, and the stochastic gradient descent (SGD) or its variants (such as Adagrad, Adadelta, Adam, etc.) optimization algorithm may be used to update the network parameters, so that the model can gradually learn the mapping relationship between speech signals and text.

[0030] For the model based on the attention mechanism, it can dynamically focus on different parts of the speech signal and has significant advantages in processing long speech sequences. The attention mechanism can automatically assign different weights to different time steps or feature dimensions according to the contextual information of the speech signal, so as to more accurately capture the key information in the speech. During the training process, the attention weights are also learned and adjusted as part of the model parameters, so that the model can better focus on the speech segments related to the current recognition task and improve the recognition accuracy. During the recognition process, the feature vector of the consultation audio signal after preprocessing and feature extraction is input into the trained speech recognition model. The model analyzes and processes the input feature vector according to the learned speech patterns and rules to generate the corresponding text sequence. This process involves complex neural network calculations and reasoning. For example, in the RNN-based model, the input feature vector of each time step is combined with the hidden state of the previous moment, and after a series of nonlinear transformations and activation function operations, the probability distribution of the text is gradually generated. Then, through decoding algorithms such as greedy search or beam search, the most likely text sequence is selected from the probability distribution as the recognition result.

[0031] In practical applications, in order to improve the accuracy and robustness of speech recognition, some post-processing technologies are also used. For example, the application of language models. Language models can further optimize and correct the recognition results based on the grammar, semantics and statistical laws of the language. Common language models include n-gram models (such as bigrams, trigrams, etc.) and neural network-based language models. They can evaluate the linguistic rationality of the generated text sequence and adjust the unreasonable parts. For example, if a word combination that does not conform to grammatical rules or semantic logic appears in the recognition result, the language model can suggest more appropriate words based on contextual information, thereby improving the quality of the consultation text.

[0032] In terms of system integration, the speech recognition module needs to work closely with the entire psychological healing dialogue system. It needs to have efficient data interaction and communication with the audio signal acquisition module, the subsequent text processing and dialogue strategy generation modules, etc. For example, when the audio signal acquisition module receives a new consultation audio signal, it needs to promptly notify the speech recognition module to process it and accurately pass the processing results to the subsequent modules. At the same time, the speech recognition module may also receive feedback information from other modules, such as user corrections or supplementary information on the recognition results, in order to continuously improve its own recognition performance.

[0033] Specifically, the consulting text word granularity encoding module 130 is used to perform word segmentation processing and semantic embedding encoding on the consulting text to obtain a sequence distribution of the consulting text word granularity description semantic embedding encoding features. Figure 2: is a block diagram of a consultation text word granularity encoding module in a large model dialogue system for psychological therapy according to an embodiment of the present application. Specifically, Figure 2 As shown, the consulting text word granularity encoding module 130 includes: a consulting text word segmentation unit 131, which is used to perform word segmentation processing on the consulting text to obtain a sequence distribution of consulting text word granularity descriptions; a consulting text word granularity description semantic encoding unit 132, which is used to use a semantic encoder to perform semantic embedding encoding on each consulting text word granularity description in the sequence distribution of the consulting text word granularity description to obtain a sequence distribution of consulting text word granularity description semantic embedding encoding vectors as the sequence distribution of the consulting text word granularity description semantic embedding encoding features.

[0034] Specifically, in the embodiment of the present application, the consulting text word segmentation unit 131 is used to perform word segmentation processing on the consulting text to obtain a sequence distribution of word granularity description of the consulting text. It should be understood that considering that the consulting text is composed of multiple meaningful units (words or phrases), each word and phrase contains semantic meaning and information between contexts. Therefore, in order to be able to understand and analyze the semantic information expressed in the user consulting text more carefully and accurately later, in the technical solution of the present application, the consulting text is segmented to divide the continuous text into words according to the grammatical rules and semantic logic of the language, and obtain a sequence distribution of word granularity description of the consulting text.

[0035] In particular, the consultation text is segmented to obtain the sequence distribution of the consultation text word granularity description, which can be performed in the following manner: First, the word boundaries in the text are identified according to pre-set rules and algorithms. Common methods include dictionary-based segmentation and statistical-based segmentation.

[0036] Dictionary-based word segmentation uses a dictionary containing a large number of common words to scan the text from the beginning of the text. When a word in the dictionary is encountered, it is segmented as an independent word. For example, for the sentence "I feel very happy today", the dictionary contains words such as "I", "today", "feeling", and "happy", and the system can accurately segment these words.

[0037] Statistical word segmentation uses a large amount of text data to train the model, counting the frequency of word occurrences and the probability of co-occurrence of adjacent words. For example, by learning from a large amount of psychology-related text, the model will know that words such as "psychological pressure" and "emotional fluctuations" often appear together, and can perform word segmentation more reasonably when encountering similar text.

[0038] In the process of word segmentation, some special cases will also be taken into consideration. For example, when processing Chinese text, attention should be paid to the semantic combination and grammatical structure of Chinese characters. For example, "rainy day" should be treated as one word, rather than being divided into two words, "rain" and "day". For English text, preliminary segmentation should be performed based on spaces and punctuation marks, and some abbreviations and hyphenated words should be handled properly.

[0039] After such word segmentation processing, the originally continuous consultation text is divided into words with independent semantics, forming a sequence distribution of word granularity description of the consultation text, providing a basis for subsequent operations such as semantic embedding coding, enabling the system to more deeply analyze and understand the meaning of the user's consultation text, thereby better promoting psychological healing dialogues.

[0040] Specifically, in an embodiment of the present application, the consultation text word granularity description semantic encoding unit 132 is used to use a semantic encoder to perform semantic embedding encoding on each consultation text word granularity description in the sequence distribution of the consultation text word granularity description to obtain a sequence distribution of the consultation text word granularity description semantic embedding encoding vector as the sequence distribution of the consultation text word granularity description semantic embedding encoding feature. In particular, the semantic encoder described here is a semantic encoder based on the Bert model. Accordingly, considering that the consultation text in the psychological healing scenario often contains rich, diverse and complex semantic information, the user may express his or her psychological state, troubles and other contents in a euphemistic, obscure or emotionally charged way. In order to accurately understand the deep semantics carried by each word in such a sentence and the relationship between them, the present application uses a semantic encoder based on the Bert model to perform semantic embedding encoding on each consultation text word granularity description in the sequence distribution of the consultation text word granularity description to map the segmented words to vectors in a high-dimensional space, and generate an embedding vector containing rich semantic information for each word, and obtain a sequence distribution of the consultation text word granularity description semantic embedding encoding vector. It is worth mentioning that the Bert model is a pre-trained language model based on the Transformer architecture. The core advantage of the Transformer architecture is that it can effectively process long sequence data and capture the semantic relationship between different positions in the input sequence through a multi-head attention mechanism. When faced with word-granular descriptions of consulting texts, it can make full use of its powerful semantic capture capabilities to accurately capture the semantic features of each word in different contexts. At the same time, the bidirectional encoding feature enables it to fully consider the context when processing word-granular descriptions, so as to deeply understand the user's consulting content, including the emotional color, intentions, and potential needs, and thus provide a more appropriate response.

[0041] The following is an example of the network structure of a semantic encoder based on the Bert model in this application scenario: The overall architecture of the Bert model is based on the Transformer structure, which is mainly composed of multiple identical layers stacked together. For the sequence distribution of the word-granularity description of the input consulting text, it will pass through an embedding layer. This embedding layer converts each word-granularity description into an initial vector representation, which combines the position information of the word and its own word vector information. For example, in a consulting text word-granularity description sequence containing "I", "recently", "pressure", and "a lot", each word will be assigned a unique initial vector in the embedding layer.

[0042] Then, these vectors processed by the embedding layer will enter multiple Transformer encoding layers in turn. In each encoding layer, there is a multi-head attention mechanism. Taking the word "pressure" as an example, in one attention head, it may focus on the word "very big" which is closely related to its semantics, while in another attention head, it may focus on "I" and "recently" at the same time to obtain more extensive contextual semantic information. Through this multi-head attention mechanism, the model can capture the semantic relationship between words from different angles.

[0043] In each encoding layer, in addition to the attention mechanism, there is also a feedforward neural network layer. This layer will further perform nonlinear transformations on the vectors processed by the attention mechanism to enhance the model's expressive power. For example, for the vector representation of the word "pressure" mentioned above, after being processed by the feedforward neural network layer, its semantic features will be further refined and enriched, which can better reflect the specific meaning in the current consultation text.

[0044] After multiple layers of encoding, each word-granular description of the consulting text will eventually get a vector that has been deeply semantically encoded. These vectors constitute the sequence distribution of the semantic embedding encoding vector of the word-granular description of the consulting text, that is, the sequence distribution of the semantic embedding encoding features of the word-granular description of the consulting text. For example, for a complete consulting text "I encountered many problems at work, I feel anxious, and I don't know what to do", after being processed by the semantic encoder based on the Bert model, the semantic embedding encoding vector of the word "work" will contain its specific semantic role in this sentence, such as it is a scene of "encountering difficulties"; the vector of "anxiety" will highlight its semantic features of negative emotions and its association with "encountering difficulties". The vectors in the entire sequence distribution can comprehensively and accurately reflect the semantic information of each word in this consulting text, providing a solid foundation for the subsequent semantic propagation of consulting texts and the generation of dialogue strategy labels.

[0045] This network structure utilizes the powerful semantic understanding ability of the Bert model. Through multi-layer encoding, complex attention mechanism and feedforward neural network processing, it can effectively transform the word-granular description in the consultation text into a sequence of encoded vectors with rich semantic information, thereby better serving the overall function of the psychological healing dialogue system.

[0046] Specifically, the consultation text semantic propagation module 140 is used to perform consultation text semantic message propagation with characteristic jumping degree on the sequence distribution of the semantic embedding coding features of the consultation text word granularity description to obtain the consultation text semantic propagation features. Furthermore, considering that in the consultation text, semantic information is often not limited to adjacent words. That is, there is still a semantic association relationship between two words that are far away in the sentence. For example, in the sentence "After experiencing a series of setbacks, I feel very confused and desperate about the future", there is a semantic association between "frustration" and "despair", but their positions in the sentence are far away. In order to capture this long-distance semantic dependency, it is very important to understand the complex sentence structure and semantics more deeply. Especially in psychological counseling, the user's expression may involve multi-level emotions and logic. The present application processes the sequence distribution of the semantic embedding coding features of the consultation text word granularity description through the consultation text semantic message propagation method with characteristic jumping degree to obtain the consultation text semantic propagation features. In particular, this method can better integrate the semantic information at different positions in the text through a special message propagation mechanism and consideration of feature jumps, thereby comprehensively reflecting how different word spans interact with each other, thereby providing a global semantic perspective for subsequent dialogue strategy recommendations.

[0047] Figure 3 : is a block diagram of a consultation text semantic propagation module in a large model dialogue system for psychological therapy according to an embodiment of the present application. Specifically, Figure 3 As shown, the consulting text semantic propagation module 140 includes: a message propagation weight calculation unit 141, which is used to calculate the message propagation significance weight of each consulting text word granularity description semantic embedded coding feature in the sequence distribution of the consulting text word granularity description semantic embedded coding feature to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight; a message propagation aggregation unit 142, which is used to perform propagation aggregation on the sequence distribution of the consulting text word granularity description semantic embedded coding feature based on the sequence distribution of the consulting text word granularity description message propagation significance weight to obtain the consulting text semantic propagation feature.

[0048] Figure 4 is a block diagram of a message propagation weight calculation unit in a large model dialogue system for psychological therapy according to an embodiment of the present application. Specifically, in the embodiment of the present application, Figure 4As shown, the message propagation weight calculation unit 141 includes: a consulting text message propagation significance factor calculation subunit 1411, which is used to calculate the message propagation significance factor of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consulting text word granularity description message propagation significance factor; a consulting text weight calculation subunit 1412, which is used to input the sequence distribution of the consulting text word granularity description message propagation significance factor into the gated propagation unit to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight.

[0049] More specifically, in the embodiment of the present application, the consultation text message propagation significant factor calculation subunit 1411 is used to: calculate the feature jump degree of each consultation text word granularity description semantic embedding coding vector in the sequence distribution of the consultation text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consultation text word granularity description feature jump degree, and the process is expressed by the following formula: ;in, is the sequence distribution of the semantic embedding encoding vector describing the word granularity of the consulting text, and They are the first, second, and third words in the sequence distribution of the semantic embedding encoding vector of the consultation text word granularity description. and The word-granular description semantic embedding encoding vector of the consulting text, yes Middle The eigenvalues ​​at the positions, yes The number of eigenvalues ​​in , and They are the first and The weighted average feature value of the semantic description of the consultation text word granularity, It is the first in the sequence distribution of the jump degree of the word granularity description feature of the consulting text. The word granularity of the consulting text describes the feature jump degree.

[0050] The message propagation space span of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector is calculated to obtain the sequence distribution of the consulting text word granularity description message propagation space span. The process is expressed by the following formula: ;in, and They are respectively the first and The word-granular description semantic embedding encoding vector of the consulting text, express arrive The number of times the message is spread, It is the first in the sequence distribution of the word granularity describing the message propagation space span of the consulting text. The word granularity of the consultation text describes the span of the message propagation space.

[0051] Based on the characteristic jump degree and message propagation space span of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector, the message propagation significance factor of each consulting text word granularity description semantic embedding coding vector is calculated to obtain the sequence distribution of the consulting text word granularity description message propagation significance factor. This process is expressed by the following formula: ;in, It is the first in the sequence distribution of the jump degree of the word granularity description feature of the consulting text. The word granularity of the consulting text describes the feature jump degree, that is, The word granularity of the consulting text describes the feature jump degree of the semantic embedding encoding vector. It is the first in the sequence distribution of the word granularity describing the message propagation space span of the consulting text. The consultation text word granularity describes the message propagation space span, that is, The message propagation space span of the semantic embedding encoding vector is described at the consultation text word granularity. and is the modulation parameter, It is the first in the sequence distribution of the significant factors of message propagation described by the word granularity of the consulting text. The word granularity of the consultation text describes the significant factor of message propagation, that is, The message propagation significance factor of the semantic embedding encoding vector described by the word granularity of the consulting text.

[0052] It should be understood that in order to determine the degree of jump of the semantic information involved in each word granularity vector in the text, that is, the degree of closeness of association with other distant semantic elements, the feature jump of each consulting text word granularity description semantic embedding encoding vector is calculated to obtain the sequence distribution of the feature jump of the consulting text word granularity description. In other words, the feature jump reflects the distance or difference of each word granularity vector in the feature space, and can effectively capture the local semantic information and mutation characteristics between nodes. With the help of the feature jump, the key semantic attributes of each word granularity vector in the feature sequence distribution can be understood, and then the nodes with large feature changes in the network can be identified, and those nodes that may have a significant impact on the communication dynamics can be highlighted.

[0053] Accordingly, the message propagation space span of the semantic embedding encoding vector described by the word granularity of each consulting text helps to identify the possible long-distance relationships and key propagation paths between words, and obtains the sequence distribution of the message propagation space span described by the word granularity of the consulting text.

[0054] It should be understood that the message propagation significance factor is calculated based on the feature jump degree of the semantic embedding encoding vector of each consulting text word granularity description and the message propagation space span to evaluate the importance of each word and its role in information propagation, and the sequence distribution of the message propagation significance factor of the consulting text word granularity description is obtained. In particular, this factor can adjust the way of information exchange according to the importance of itself and its neighbors, so as to comprehensively consider local key semantics and global factors, and accurately point out which word vectors are more critical in message propagation.

[0055] More specifically, in the embodiment of the present application, the consulting text weight calculation subunit 1412 is used to: use the gated filtering module of the softmax function and the msak function to process each consulting text word granularity description message propagation significance factor in the sequence distribution of the consulting text word granularity description message propagation significance factor to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight, and the process is expressed by the following formula: ;in, It is the first in the sequence distribution of the significant factors of message propagation described by the word granularity of the consulting text. The word granularity of the consultation text describes the significant factors of message propagation. is the normalization function, is a masking operation, is a predetermined threshold, It is the first in the sequence distribution of the significant weight of the consultation text word granularity description message propagation The word granularity of the consultation text describes the significant weight of the message propagation.

[0056] Accordingly, the sequence distribution of the message propagation significance factors of the consulting text word granularity description is input into the gated propagation unit to obtain the sequence distribution of the message propagation significance weights of the consulting text word granularity description. That is, the gated propagation unit can dynamically assign message propagation weights to each word granularity description according to the input message propagation significance factor sequence distribution, that is, flexibly adjust the influence of each element in the semantic message transmission process according to the importance of each element, thereby realizing refined control of the semantic propagation process.

[0057] Specifically, in the embodiment of the present application, the message propagation collection unit 142 is used to: describe the sequence distribution of the message propagation significance weight at the consulting text word granularity, calculate the weighted sum of the sequence distribution of the semantic embedding coding vector described at the consulting text word granularity to obtain the consulting text semantic propagation coding vector as the consulting text semantic propagation feature, and the process is expressed by the following formula: ;in, and They are respectively the first The word-granular description semantic embedding encoding vector of the consulting text, It is the first in the sequence distribution of the significant weight of the consultation text word granularity description message propagation The word granularity of the consultation text describes the significant weight of the message propagation, is the number of vectors in the sequence distribution of the semantic embedding encoding vector describing the word granularity of the consulting text, is the semantic propagation encoding vector of the consulting text.

[0058] Finally, the sequence distribution of the significant weights of message propagation is described at the word granularity of the consulting text. The weighted sum of the sequence distribution of the semantic embedding coding vector described at the word granularity of the consulting text is calculated and the semantic information with different degrees of importance is integrated to obtain a consulting text semantic propagation coding vector that can comprehensively reflect the semantics of the entire consulting text.

[0059] Specifically, the dialogue strategy label generation module 150 is used to obtain the recommended dialogue strategy label based on the semantic propagation feature of the consultation text. Specifically, in the embodiment of the present application, the dialogue strategy label generation module is used to: input the semantic propagation encoding vector of the consultation text into the dialogue strategy recommendation module based on the classifier to obtain the recommended dialogue strategy label. That is, the semantic propagation feature of the consultation text obtained by feature jump propagation using the sequence distribution of the semantic embedding encoding vector described by the consultation text word granularity is classified and processed to intelligently generate the recommended dialogue strategy label. In other words, the dialogue strategy recommendation module based on the classifier can act as a "translator" to convert this semantic level representation into a specific and operational dialogue strategy label, such as judging whether to adopt empathic response, guided questioning, providing suggestions or other specific strategies, so that the system can select the appropriate dialogue method in a targeted manner according to the semantic content input by the user, and complete the key connection from understanding semantics to formulating dialogue strategies.

[0060] It should be understood that, here, since each consulting text word granularity description semantic embedding coding vector in the sequence of the consulting text word granularity description semantic embedding coding vectors respectively represents the encoded semantic feature representation of the consulting text in the word granularity local semantic space, when performing feature sequence message propagation based on feature jump degree in the global semantic space, the semantic feature population attributes of different local semantic spaces will have sequence propagation fairness differences based on the feature jump degree level, thereby affecting the encoding feature distribution inclusiveness of the consulting text semantic propagation coding vector, and reducing the accuracy of the recommended dialogue strategy label obtained by inputting the classifier-based dialogue strategy recommendation module.

[0061] In a preferred example, inputting the semantic propagation coding vector of the consultation text into a classifier-based dialogue strategy recommendation module to obtain a recommended dialogue strategy label includes: determining the mean of the eigenvalues ​​corresponding to the semantic propagation coding vector of the consultation text and the standard deviation of the eigenvalues .

[0062] After the consultation text semantic propagation coding vector is dot-added with the eigenvalue mean, the dot-multiplied result is dot-multiplied with the eigenvalue mean to obtain the consultation text semantic propagation coding conditional probability constraint vector ,in, Indicates point addition, represents dot product, Represents the semantic propagation encoding vector of the consulting text.

[0063] The consulting text semantic propagation coding conditional probability constraint vector and the eigenvalue standard deviation dot-add vector are further subtracted from the consulting text semantic propagation coding vector to obtain the consulting text semantic propagation coding interaction fairness measurement vector ,in, Indicates point reduction.

[0064] The bit-by-bit inverse of the consultation text semantic propagation coding interaction fairness measurement vector is added to the consultation text semantic propagation coding conditional probability constraint vector, and then multiplied with the consultation text semantic propagation coding vector to obtain the consultation text semantic propagation coding group feature transfer vector ,in, Represents the bit-by-bit inverse of the consultation text semantic propagation encoding interactive fairness metric vector.

[0065] The eigenvalue mean The standard deviation of the eigenvalue The sum is multiplied by the scaling weight hyperparameter After that, it is multiplied by the semantic propagation coding group feature transfer vector of the consulting text to obtain the optimized semantic propagation coding vector of the consulting text .

[0066] The optimized consultation text semantic propagation encoding vector is input into a classifier-based dialogue strategy recommendation module to obtain a recommended dialogue strategy label.

[0067] Here, since each consulting text word granularity description semantic embedding coding vector in the sequence of the consulting text word granularity description semantic embedding coding vectors respectively represents the encoded semantic feature representation of the consulting text in the word granularity local semantic space, when performing feature sequence message transmission based on feature jump degree in the global semantic space, the semantic feature population attributes of different local semantic spaces will have sequence transmission fairness differences based on the feature jump degree level, thereby affecting the encoding feature distribution inclusiveness of the consulting text semantic propagation coding vector, and reducing the accuracy of the recommended dialogue strategy label obtained by inputting the classifier-based dialogue strategy recommendation module.

[0068] Therefore, in order to solve the problem of attribute balance deviation of the fusion features of the semantic propagation coding vector of the consulting text in cross-modal samples, it is necessary to enhance the dynamic adaptability of the semantic propagation coding vector of the consulting text under the condition of heterogeneity of the representation space. Therefore, by introducing an interactive fairness measurement model based on conditional probability constraints for the semantic propagation coding vector of the consulting text, the group feature transmission mechanism of the semantic propagation coding vector of the consulting text is iteratively optimized, and the statistical association mechanism is adopted as the adjustment factor of hierarchical fairness compensation.

[0069] In this way, a robust distribution consistency representation is achieved through the multi-dimensional feature collaborative framework of the consultation text semantic propagation coding vector, and finally a feature interaction paradigm with adaptive adjustment capability is formed, which effectively improves the fairness and collaborative efficiency of the multimodal joint encoding of the consultation text semantic propagation coding vector, and improves the accuracy of the recommended dialogue strategy labels obtained by its input classifier-based dialogue strategy recommendation module.

[0070] Specifically, the dialogue text generation module 160 is used to obtain a dialogue text based on the consultation text and the recommended dialogue strategy label. Specifically, in an embodiment of the present application, the dialogue text generation module is used to: after adding the recommended dialogue strategy label to the end of the consultation text, input it into the dialogue module based on the large model to obtain the dialogue text. That is, the consultation text carries the core content such as the user's psychological demands and troubles, and the recommended dialogue strategy label is the response direction carefully selected by the system based on the deep understanding of the text. Combining the two, the dialogue module based on the large model can obtain the most critical information, so as to generate responses around the user input and avoid the blindness of dialogue generation. For example, if the label is "empathy", the large model can focus on feeling the user's emotions and emphasize expressing understanding and resonance when generating dialogue text. It should be understood that the large model has a strong language learning and generation ability, and can learn language expressions in different contexts and semantics based on massive data. Combining the input method of consultation text and strategy labels, the big model can fully call on these knowledge reserves, accurately target users' psychological problems, and use appropriate strategies to generate conversation texts that are both professional and empathetic, maximizing the advantages of the big model in natural language processing and meeting the complex and changing needs of psychological healing conversations.

[0071] In particular, after adding the recommended dialogue strategy label to the end of the consultation text, it is input into the dialogue module based on the big model to obtain the dialogue text. The following is the specific implementation process: the system will accurately add the recommended dialogue strategy label to the end of the consultation text by text splicing. This operation seems simple, but it is necessary to ensure that the connection between the label and the text is natural and smooth, without any semantic ambiguity or confusion. For example, if the consultation text is "I feel confused about the future and don't know how to choose a career direction", and the recommended label is "provide advice", then the spliced ​​text is "I feel confused about the future and don't know how to choose a career direction to provide advice". Then, the spliced ​​text is input into the dialogue module based on the big model. After receiving the input, this dialogue module will immediately start a series of complex processing flows. The big model, with its powerful language understanding and generation capabilities trained on a large amount of text data, deeply analyzes the input text. It will identify key semantic elements in the consultation text, such as "future", "confused", "career direction", etc., and understand the user's current confusing situation.

[0072] Based on the understanding of the input text and the accompanying dialogue strategy labels, the large model begins to generate the corresponding dialogue text. If it is labeled "empathy", it will select appropriate sentences from its rich emotional expression and response knowledge base, such as "I can fully understand your confusion and helplessness now. This uncertainty about the future does make people anxious, but you have to believe that we can slowly find the direction together." These sentences are designed to resonate with the user's emotions and make the user feel understood and supported.

[0073] If it is a "guided question" label, the big model will carefully construct some inspiring questions to guide users to further think and express themselves. For example, for the above example of career choice, it may generate "Have you ever thought about your interests and hobbies and strengths? In your past experience, is there any work that makes you feel particularly fulfilled?" Through these questions, users can dig out more potential information so that the system can better help them.

[0074] In the process of generating dialogue text, the big model will also use natural language processing technology to optimize and adjust the generated sentences. It will check the correctness of the grammar to ensure that the sentences are smooth and easy to understand; at the same time, it will pay attention to the coherence of the semantics, so that the generated dialogue text is logically closely connected and conforms to the normal language communication logic.

[0075] In addition, the big model will continue to learn and evolve. Each input and generation process is an opportunity to learn. It will continuously improve its generation strategy and language expression ability based on user feedback and new data. If the user questions the generated dialogue text or provides new information, the big model will incorporate this feedback into the subsequent learning and optimization process, thereby continuously improving its performance in the psychological healing dialogue scenario, providing users with better quality and more effective dialogue texts, and helping users solve psychological problems and confusion.

[0076] In summary, a large-model dialogue system 100 for psychological therapy based on an embodiment of the present application is explained, which obtains the consultation audio signal input by the user, performs speech recognition on it to obtain the consultation text, uses natural language processing technology and encoding methods based on deep learning to perform word segmentation processing on the consultation text, and then performs semantic embedding encoding on the word granularity descriptions of each consultation text after word segmentation, so as to intelligently generate recommended dialogue strategy labels based on the feature jump semantic propagation representation between the semantic embedding features of the word granularity descriptions of each consultation text, and obtain the dialogue text based on the label and the consultation text. The present application can better understand the user's intentions and emotional state, and flexibly select empathy or questioning strategies according to the user's real-time emotional state and consultation content, thereby providing a more personalized response, ensuring that each interaction is appropriate and productive, and avoiding the problems of inappropriate questions or insufficient empathy in traditional systems.

[0077] As described above, the large model dialogue system 100 for psychological therapy according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a large model dialogue algorithm for psychological therapy. In one possible implementation, the large model dialogue system 100 for psychological therapy according to the embodiment of the present application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the large model dialogue system 100 for psychological therapy can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the large model dialogue system 100 for psychological therapy can also be one of the many hardware modules of the wireless terminal.

[0078] Alternatively, in another example, the large model dialogue system 100 for psychological therapy and the wireless terminal may also be separate devices, and the large model dialogue system 100 for psychological therapy may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0079] The above descriptions of various implementations of the present disclosure are exemplary and not exhaustive. The disclosure is not limited to the disclosed implementations, and many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The terms used herein are selected to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various practical methods disclosed herein.

Claims

1. A large-scale model dialogue system for psychological therapy, characterized in that: include: A consultation audio signal acquisition module, used to acquire a consultation audio signal input by a user; A signal recognition module, used for performing speech recognition on the consultation audio signal to obtain a consultation text; A consultation text word granularity encoding module is used to perform word segmentation processing and semantic embedding encoding on the consultation text to obtain a sequence distribution of the consultation text word granularity description semantic embedding encoding features; The consulting text semantic propagation module is used to perform consulting text semantic message propagation with feature jump degree on the sequence distribution of the semantic embedding coding features of the consulting text word granularity description to obtain the consulting text semantic propagation features, wherein the consulting text semantic propagation module includes: a message propagation weight calculation unit, used to calculate the message propagation significance weight of each consulting text word granularity description semantic embedding coding feature in the sequence distribution of the consulting text word granularity description semantic embedding coding features to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight; a message propagation aggregation unit, used to perform propagation aggregation on the sequence distribution of the semantic embedding coding features of the consulting text word granularity description based on the sequence distribution of the message propagation significance weight of the consulting text word granularity description to obtain the consulting text semantic propagation features; a dialogue strategy label generation module, used to obtain the recommended dialogue strategy label based on the consulting text semantic propagation features; a dialogue text generation module, used to obtain the dialogue text based on the consulting text and the recommended dialogue strategy label.

2. The large-scale model dialogue system for psychological therapy according to claim 1, characterized in that: The consulting text word granularity encoding module includes: a consulting text word segmentation unit, which is used to perform word segmentation processing on the consulting text to obtain a sequence distribution of consulting text word granularity descriptions; a consulting text word granularity description semantic encoding unit, which is used to use a semantic encoder to perform semantic embedding encoding on each consulting text word granularity description in the sequence distribution of the consulting text word granularity description to obtain a sequence distribution of consulting text word granularity description semantic embedding encoding vectors as the sequence distribution of the consulting text word granularity description semantic embedding encoding features.

3. The large model dialogue system for psychological therapy according to claim 2, characterized in that: The semantic encoder is a semantic encoder based on the Bert model.

4. The large model dialogue system for psychological therapy according to claim 3 is characterized in that: The message propagation weight calculation unit includes: a consulting text message propagation significance factor calculation subunit, which is used to calculate the message propagation significance factor of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consulting text word granularity description message propagation significance factor; a consulting text weight calculation subunit, which is used to input the sequence distribution of the consulting text word granularity description message propagation significance factor into a gated propagation unit to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight.

5. The large model dialogue system for psychological therapy according to claim 4, characterized in that: The consulting text message propagation significance factor calculation subunit is used to: calculate the feature jump degree of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consulting text word granularity description feature jump degree; calculate the message propagation space span of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consulting text word granularity description message propagation space span; based on the feature jump degree and message propagation space span of each consulting text word granularity description semantic embedding coding vector in the sequence distribution of the consulting text word granularity description semantic embedding coding vector, calculate the message propagation significance factor of each consulting text word granularity description semantic embedding coding vector to obtain the sequence distribution of the consulting text word granularity description message propagation significance factor.

6. The large-scale model dialogue system for psychological therapy according to claim 5, characterized in that: The consulting text weight calculation subunit is used to: use the gated filtering module of the softmax function and the msak function to process each consulting text word granularity description message propagation significance factor in the sequence distribution of the consulting text word granularity description message propagation significance factor to obtain the sequence distribution of the consulting text word granularity description message propagation significance weight.

7. The large model dialogue system for psychological therapy according to claim 6, characterized in that: The message propagation collection unit is used to: describe the sequence distribution of message propagation significance weights at the consulting text word granularity, calculate the weighted sum of the sequence distribution of the semantic embedding coding vector described at the consulting text word granularity to obtain the consulting text semantic propagation coding vector as the consulting text semantic propagation feature.

8. The large model dialogue system for psychological therapy according to claim 7, characterized in that: The dialogue strategy label generation module is used to: input the consultation text semantic propagation encoding vector into the classifier-based dialogue strategy recommendation module to obtain the recommended dialogue strategy label.

9. The large-scale model dialogue system for psychological therapy according to claim 8, characterized in that: The dialogue text generation module is used to: add the recommended dialogue strategy tag to the end of the consultation text, and then input it into the dialogue module based on the large model to obtain the dialogue text.

Images