Cloud Hosting

Review of seismic imaging

The goal of seismic imaging is to infer subsurface structures based on observed seismic data. This can be achieved by solving inverse problems. Reverse Time Migration (RTM) is an imaging technique based on the wave equation25, which utilizes the cross-correlation of the underground forward and backward wavefields, demonstrating excellent adaptability, especially in areas with complex structures and high velocity variations. The formula for the cross-correlation imaging condition is expressed as:

$$I(x,z)={int }_{0}^{T}{u}_{text{f }}(x,z,t)*{u}_{text{b }}(x,z,t)dt$$

(1)

Here, (I(x,z)) represents the RTM result, ({u}_{text{f }}(x,z,t)) denotes the forward wavefield, and ({u}_{text{b }}(x,z,t)) is the backward wavefield.

However, RTM suffers from low-frequency noise and inaccurate amplitudes, limiting its application in seismic imaging. To address the shortcomings of RTM, Least Squares Reverse Time Migration (LSRTM) associates the migration imaging result with seismic data26, constructing the least squares objective function:

$$E({varvec{m}})=frac{1}{2}{parallel {varvec{L}}{varvec{m}}-{{varvec{d}}}_{{text{obs}}}parallel }^{2}$$

(2)

Here, ({{varvec{d}}}_{{text{obs}}}) represents the observed data, ({varvec{L}}) is the forward operator, and ({varvec{m}}) is the subsurface structural parameter.

LSRTM involves key steps such as forward simulation, backpropagation, gradient computation, and optimization algorithms. Through iterative optimization to minimize the error between observed and simulated data, LSRTM enhances the quality of seismic imaging.

In this study, we introduce a hybrid architecture (Fig.1) that integrates Transformer and CNN to address seismic imaging tasks. Within the Transformer framework, the need for a one-dimensional sequence as input necessitates an initial transformation of the input image. The Image Patching phase involves partitioning the input image into a series of equally sized image patches, each with a size of ({P}^{2}). This transforms the original (Htimes W) image into an (Ntimes Ptimes P) sequence, where (N) represents the sequence length, encompassing (frac{Htimes W}{{P}^{2}}) image patches. Consequently, the input image is reshaped into a one-dimensional sequence, with each image patch corresponding to a vector. The adoption of a smaller patch size enables enhanced capture of intricate details within the image, thus elevating the model's accuracy, albeit at the expense of increased computational overhead27. In view of balancing between model efficacy and computational efficiency, we establish (P=16). In the Input Embedding stage, a linear transformation is applied to each segmented image patch, mapping it to a continuous vector representation. As the Transformer model abstains from utilizing recurrent or convolutional layers for sequence processing, positional encoding is incorporated into the input embedding vector to discern the positional information of each image patch.

Network architecture diagram.

$${mathbf{Z}}_{0}=[{{varvec{X}}}_{p}^{1}mathbf{E};{mathbf{X}}_{p}^{2}mathbf{E};dots ;{mathbf{X}}_{p}^{N}mathbf{E}]+{mathbf{E}}_{text{pos}}$$

(3)

The proposed model employs a Transformer Encoder comprising ({text{L}}=12) layers to process the image sequence, with each encoder layer composed of Multi-Head Self-Attention (MSA) and Multi-Layer Perceptron (MLP).

$${mathbf{Z}}_{l}^{mathrm{^{prime}}}={text{MSA}}({text{LN}}({mathbf{Z}}_{l-1}))+{mathbf{Z}}_{l-1},l=1dots L$$

(4)

$${mathbf{Z}}_{l}={text{MLP}}({text{LN}}({mathbf{Z}}_{l}^{mathrm{^{prime}}}))+{mathbf{Z}}_{l}^{mathrm{^{prime}}},l=1dots L$$

(5)

Here, ({text{LN}}(cdot )) denotes layer normalization, (l) is the identifier for intermediate blocks, and L is the number of Transformer layers.

These stacked Transformer layers facilitate capturing the complexity of the data from a multiscale perspective. To prevent the loss of primary features by solely relying on the last layer output, we employ a multi-level feature extraction strategy. In addition to the final layer (12th layer), features are extracted from the 6th and 9th layers, representing deep, intermediate, and shallow features, providing a rich and multiscale feature space. These three layers of features are adjusted to different resolutions of feature maps and fused through ASFF, resulting in adaptive aggregation at each scale.

ASFF constitutes an attention-based spatial feature integration strategy devised to amalgamate feature maps originating from diverse spatial resolutions within deep neural networks28. Its principal objective is to augment the model's perceptual acuity concerning targets across varying scales. ASFF dynamically weights and fuses features from distinct spatial resolutions by learning task-specific attention weights.

We represent features at resolution level ({ell}) (where ({ell}in left{mathrm{1,2},3right})) as ({x}^{l}). For level ({ell}), we resize features from other levels (n) ((nne {ell})) to the same shape as ({x}^{l}). Let ({x}_{ij}^{nto {ell}}) denote the feature vector at position ((i,j)) on the feature map, adjusted from level (n) to level ({ell}). We perform the following fusion of corresponding level ({ell}) features:

$${y}_{ij}^{{ell}}={alpha }_{ij}^{{ell}}cdot {x}_{ij}^{1to {ell}}+{beta }_{ij}^{{ell}}cdot {x}_{ij}^{2to {ell}}+{gamma }_{ij}^{{ell}}cdot {x}_{ij}^{3to {ell}}$$

(6)

Here, ({y}_{ij}^{{ell}}) signifies the vector at position ((i,j)) in the output feature map ({y}^{{ell}}) across channels. The spatial importance weights ({alpha }_{ij}^{{ell}}), ({beta }_{ij}^{{ell}}), and ({gamma }_{ij}^{{ell}}) for features from three different levels to level ({ell}) are adaptively learned by the network. To ensure the effectiveness of weights, constraints ({alpha }_{ij}^{{ell}}+{beta }_{ij}^{{ell}}+{gamma }_{ij}^{{ell}}=1) and ({alpha }_{ij}^{{ell}}, {beta }_{ij}^{{ell}}, {gamma }_{ij}^{{ell}}in [mathrm{0,1}]) are enforced. These constraints ensure the validity and range of the weights. The weights are computed using softmax functions with control parameters as follows:

$${alpha }_{ij}^{{ell}}=frac{{e}^{{lambda }_{{alpha }_{ij}}^{{ell}}}}{{e}^{{lambda }_{{alpha }_{ij}}^{{ell}}}+{e}^{{lambda }_{{beta }_{ij}}^{{ell}}}+{e}^{{lambda }_{{gamma }_{ij}}^{{ell}}}}$$

(7)

$${beta }_{ij}^{{ell}}=frac{{e}^{{lambda }_{{beta }_{ij}}^{{ell}}}}{{e}^{{lambda }_{{alpha }_{ij}}^{{ell}}}+{e}^{{lambda }_{{beta }_{ij}}^{{ell}}}+{e}^{{lambda }_{{gamma }_{ij}}^{{ell}}}}$$

(8)

$${gamma }_{ij}^{{ell}}=frac{{e}^{{lambda }_{{gamma }_{ij}}^{{ell}}}}{{e}^{{lambda }_{{alpha }_{ij}}^{{ell}}}+{e}^{{lambda }_{{beta }_{ij}}^{{ell}}}+{e}^{{lambda }_{{gamma }_{ij}}^{{ell}}}}$$

(9)

The calculation of control parameters ({lambda }_{{alpha }_{ij}}^{{ell}}),({lambda }_{{beta }_{ij}}^{{ell}}), and ({lambda }_{{gamma }_{ij}}^{{ell}}) is performed through (1{text{x}}1) convolution layers from ({x}_{ij}^{1to {ell}}),({x}_{ij}^{2to {ell}}), and ({x}_{ij}^{3to {ell}}), respectively. These parameters are learned through standard backpropagation during network training.

Overall, this approach furnishes the model with a rich and multiscale feature space, thereby contributing to its performance in complex seismic imaging tasks.

Visit link:
Deep learning for high-resolution seismic imaging | Scientific Reports - Nature.com

Data collection Data collection

The research group collected 17-dimensional basic trait data (Supplementary Information1. andSupplementary Information 2.) of 25480 samples of community correction prisoners in Zhejiang Province, China, and the corresponding Symptom Checklist-90 (SCL-90) and Health Survey Short Form (SF-12) data. These data were collected through the standardized community correction digital management platform of the Zhejiang Provincial Department of Justice, covering the period from January 2020 to December 2020. The 17-dimensional characteristics mainly include age, sex, treatment level (general control, strict control), whether adult, education level, dmicile (urban or rural), whether there are infectious diseases, whether belongs to the following three categories (unemployed individuals, those without relatives to rely on, individuals without a place to live), whether there is a criminal record, crime type, supervision time, whether there is recidivism, whether there is anti-government tendency, whether there are five kinds of involvement (terrorism, cults, drugs, gangs, and gun trafficking), whether there are four histories (drug use history, escape history, suicide history, police assault history), correction status (in correction, released from the status of correction), occupation before arrest. The SCL-90 traditional scale obtained 9 kinds of psychological measurement indicators: somatization, obsessive-compulsive symptoms, interpersonal sensitivity, depression, anxiety, hostility, terror, paranoia, and psychosis. Due to the incomplete basic information registered in some judicial offices, the samples with missing values in the basic information were removed and matched, resulting in a total of 25,214 sample data.

Due to the privacy and compliance issue of patients, it is difficult to collect a large number of medical data, especially the data of specific groups. The research group has invested a lot of manpower, material and financial resources in the construction of this data set (Supplementary Information 3.).

The research design has been approved by the Ethics Research Committee of the Zhejiang Community Correction Management Bureau. This study was carried out in accordance with the Declaration of Helsinki, and all procedures were carried out in accordance with relevant guidelines and regulations. The Committee waived the requirement of informed consent for this study because the researchers only access the database for analysis purposes, and all personnel, including patient data, are desensitized, and there is no conflict of interest among personnel of each unit.

The pretreatment of tabulated data described in the paper includes missing value imputation, outlier detection and removal and data standardization, as follows:

Missing values refer to situations where the values of certain features or variables in a table are missing or not recorded. In machine learning modeling, handling missing values is crucial36. Choosing appropriate filling methods can improve the predictive performance of the model, making the data more complete and reliable37. In this study, there were some missing values in the raw data we used, and most of the missing values were filled in by manually tracing the raw materials. For a small amount of other missing values such as age and other quantitative data, we use mean interpolation to fill in, as the mean can represent the central trend of the data and help maintain its distribution.For qualitative data such as crime types, we use the median to fill in, which is a better choice because it can reduce the impact of extreme values while maintaining the order and level of the data38.

Outliers refer to data points that are significantly different from other data points or deviate from the normal range. Outliers may have adverse effects on data analysis and modeling, so they need to be eliminated or handled. To ensure the accuracy and reliability of the data, we carried out outlier detection and elimination. We use the Rajda criterion to deal with outliers. The process takes the given confidence probability of 99.7% as the standard, and is based on the standard deviation of 3 times of the data column. The abnormal data row greater than the value is deleted, and when the residual error vb of the measured value xb is greater than 3 times , outliers should be eliminated.

$$left| {vb} right| = left| {xb - x} right| > 3sigma .$$

Data standardization is to transform the data of different scales and ranges into a unified standard scale to eliminate the influence of dimensions and make different features comparable. In the stage of data preprocessing, we normalize the numerical features from minimum to maximum. By linearly mapping the values of each feature to the range of 0 to 1, we eliminate the differences of different feature scales and make them comparable.

Based on symptom checklist-90(SCL-90), this study constructed an adaptive scale (between question groups) simplification screening evaluation model based on multi-label classification algorithm, and used Health Survey Short Form(SF-12), a primary screening tool commonly used by community correction management institutions, as a simplified baseline method for comparative analysis.

We used the multi-label classification model for scale (between question groups) simplification to analyze the risk degree of individuals in nine categories of psychological measurement indicators, and simplified the scale structure based on the risk distribution. The goal of scale simplification is to simplify the questions, make the scale more readable and easy to understand, and help readers get core information and insight more quickly. During the process of scale simplification, it is necessary to make trade-offs and decisions according to the data and the needs of the audience to ensure that enough information is retained while maintaining simplicity and clarity.

The basic principle of the multi-label classification algorithm (as shown in Fig. 1 and Table 1) is to recognize the association between features and labels by learning historical data, so as to predict new labels. It can integrate the results of multiple tags, find the association between multiple tags, and solve the multiple conflicts that may exist in the multi-tag classification problem, so as to effectively improve the accuracy of classification. It can also help us quickly identify features, thus reducing the time of classification.

Binary relevance (first-order, y tags are independent of each other). It is a problem transformation method. The core idea is to decompose the multi-label classification problem. BR is simple and easy to understand. When there is no dependent relationship between Y values, the effect of the model is good.

Classifier chains (high-order, y tags are interdependent). Its principle is similar to the BR conversion method. In this case, the first classifier is trained only on the input data, and then each classifier is trained on all previous classifiers in the input space and chain. A certain number of binary classifiers can be combined into a single multi-label model to explore the correlation between multiple targets.

Rakle (random k-labelsets, high-order, y tags are interdependent). It can divide the original large tag set into a certain number of small tag sets, then use RF to train the corresponding classifier, and finally integrate the prediction results. RakeID is a high-order strategy algorithm, which can mine the correlation of multiple tags according to the size of the tag subset.

Multi label classification algorithm.

For the latter two algorithms, if there is a clear dependency between tags, the generalization ability of the final model is better than that of the model constructed by binary relevance. The problem is that it is difficult to find a more suitable tag dependency.

The core principle of oversampling method is to increase some samples in the category with fewer samples to achieve category balance. SMOTE is the representative algorithm of the oversampling method. In the process of modeling, SMOTE (Synthetic Minority Over-sampling Technique) is used to solve the problem of category imbalance. SMOTE increases the number of minority samples by synthesizing new minority samples, to balance the unbalanced data set.

Because the total number of samples collected is sufficient, the training data adopts 5-fold cross-validation to prevent the model from overfitting and increase the robustness of the model. The extracted feature data is randomly divided into five parts, four of which are used for training, and one part is retained as test data. The above process is repeated five times, using different test data each time. Then the results of these five times are summarized, and the average value is taken as the estimation of the algorithm performance index. Five cross-validation is a popular algorithm choice at present.

In this paper, SF-12 was used as a comparison tool. SF-12 is a commonly used health questionnaire survey tool, which is used to assess the health status and quality of life of individuals. SF-12 is a simplified version derived from the SF-36 questionnaire, which retains the core concepts and dimensions of SF-36. However, it reduces the number of questions and improves the efficiency of questionnaire implementation. The simplicity and efficiency of the SF-12 questionnaire make it a common tool in large-scale epidemiological research and clinical practice. It can be used to evaluate the health status of different groups and the effect of health intervention, and compare the health differences between different groups.

If all SCL-90 subscales of the actual sample are diagnosed as risk-free, the sample is defined as a negative sample. If any subscale test is risky, the sample is defined as a positive sample. Similarly, if all the sub-tags predicted by the multi-label model are 0, the sample is negative. If there is any positive sub-tag, the sample is positive:

If the actual 9 labels are all negative, the mental state is healthy and marked as a negative sample.

If one of the actual 9 labels is positive, the mental state is unhealthy and marked as a positive sample.

Similarly, if all of the predicted 9 tags are negative, the mental state is healthy and the tag is negative.

If one of the predicted 9 tags is positive, the mental state is unhealthy and marked as a positive sample.

According to the actual mental state and the predicted value, the confusion matrix (as shown in Table 2) is drawn, which is composed of the following four important definitions: true positive (TP), false positive (FP), false negative (FN) and true negative (TN).

The overall effect of the model is evaluated by the following indicators, including accuracy, sensitivity, specificity and F1. The relevant measurement standards are as follows:

$${text{Accuracy }} = left( {{text{TP }} + {text{ TN}}} right)/left( {{text{TP }} + {text{ TN }} + {text{ FP }} + {text{ FN}}} right),$$

$${text{Sensitivity }} = {text{ TP}}/left( {{text{TP }} + {text{ FN}}} right),$$

$${text{Precision }} = {text{ TP}}/left( {{text{TP }} + {text{ FP}}} right),$$

$${text{F1}} = {2} times {text{Sensitivity}} times {text{Precision}}/left( {{text{Precision}} + {text{Sensitivity}}} right).$$

In the multi label classification problem, accuracy_Score, Hamming loss and 0-1 loss related evaluation indicators can be based on the prediction results of a single tag or the overall prediction results.

Accuracy_Score is the correctly predicted score (default) or count. In multi-label classification, the function returns the subset precision. If the whole set of predicted tags of the sample matches the real tag combination, the subset accuracy is 1. Otherwise, it is 0.

Hamming loss: Hamming loss measures the prediction accuracy of the model for each label, that is, the ratio of the number of labels with average prediction errors to the total number of labels. It calculates the prediction result of each tag and returns a value between 0 and 1. The smaller the value, the more accurate the prediction is.

0-1 loss is a common classification loss function, which is used to measure the prediction error of the classification model. It takes 1 when the prediction is wrong and 0 when the prediction is correct, so it is named 0-1 loss.

Simplification rate refers to the proportion of the simplified scale to the original scale, which can be used to evaluate the degree of simplification of the scale. Scale simplification refers to simplifying the structure of the original scale by reducing the number of items, deleting redundant or unnecessary items, or merging multiple items. The simplification rate of the scale can be calculated in the following way: simplification rate (number of simplified items/original number of items) 100%. In other words, the simplification rate based on the multi-label model is calculated as follows: simplification rate (the number of sub-labels predicted to be negative)/(the total number of samples).

The Ethics Committee of the Zhejiang Community Correction Management Bureau has waived the informed consent requirement for this study, as researchers accessing the database is only for analytical purposes, including patient data, which is desensitized, and there are no conflicts of interest between personnel in each unit. The research design has been approved by the Ethics Research Committee of the Zhejiang Community Correction Management Bureau. This study was conducted in accordance with the Helsinki Declaration, and all procedures were conducted in accordance with relevant guidelines and regulations.

Continued here:
Research on a machine learning-based adaptive and efficient screening model for psychological symptoms of ... - Nature.com

Conversational AI vs. Generative AI: Which solution will turbocharge your contact centers performance and help you achieve your CX goals? Worldwide, the evolution of artificial intelligence has unlocked new waves of productivity for business leaders and teams.

While the impact of advanced AI algorithms can be felt everywhere, its particularly prominent in the contact center. In the last year alone, weve lost count of the number of contact center, CRM, and CX software vendors introducing new AI capabilities for customer service teams.

Though ChatGPT, Microsoft Copilot, and even solutions like NICEs Enlighten AI suite are driving focus to the rise of generative AI, its not the only intelligent tech making waves. Conversational AI is also emerging as a critical part of contact center success.

The question is, which of these two solutions do you need, and do you need to choose between one or the other? Heres your guide to conversational AI and generative AI in the contact center.

Conversational AI is a type of artificial intelligence that allows computer programs (bots) to simulate human conversations. It combines various AI techniques to ensure people can interact with computer systems just like talking to another human being.

Examples of conversational AI are everywhere. Smart assistants like Alexa and Siri use conversational AI to interact with users. Many of the chatbots installed on company websites leverage the same technology.

So, how does it all work?

While the nature of each conversational AI solution can vary depending on your chosen vendor, most tools feature the same central components:

After processing input, conversational AI tools can generate responses based on their data. Some more advanced solutions can even enhance their responses by using additional forms of analysis, such as sentiment analysis.

Conversational AI has become the backbone of many advances in the customer experience and contact center landscapes. It forms part of the tech behind conversational intelligence tools, such as those offered by CallMiner, Calabrio, and Talkdesk.

Its also a common component in the chatbots and virtual assistants customers interact with through text and speech, for self-service interactions.

The most common examples of conversational AI in customer service include:

Older chatbots were primarily rule-based solutions that used scripts to answer customer questions. Advanced chatbots, powered by conversational AI, use natural language processing to recognize speech, imitate human interaction, and respond to more complex inputs.

They can also operate across multiple channels, accompanying your contact center IVR system, chat apps, social media service strategies, and more. Plus, they can learn from interactions over time, becoming more effective and advanced.

Modern IVR systems also leverage conversational AI. Instead of giving customers a list of limited options to choose from, they can listen to what customers say, recognize their intent, and route them to the best agent or department.

With NLP, IVR systems can provide more accurate responses and even draw insights from company databases and CRMs to personalize interactions. They can also be configured to route conversations based on various factors, such as customer sentiment or agent skill level.

As mentioned above, conversational AI tools are a common component of conversational intelligence. Because they can process language and analyze interactions, they can offer companies insight into customer sentiment, track customer service trends, and highlight growth opportunities.

Some solutions can also automatically transcribe and translate calls, which can be ideal for enhancing compliance, as well as training initiatives.

When analyzing conversational AI vs. generative AI, its worth noting that both solutions have strengths and limitations. Conversational AI, for instance, can empower teams to deliver fantastic service across multiple channels 24/7. It can also help personalize interactions.

By analyzing previous discussions and real-time sentiment or intent, conversational AI can help ensure every customer gets a bespoke experience with your contact center.

Beyond that, conversational AI can:

However, conversational AI also has limitations. Although conversational AI tools are more advanced than traditional chatbots, they can still struggle with complex linguistic nuances and requests. They dont always understand customer accents or things like humor or sarcasm.

Plus, since theyre reliant on collecting and processing customer data, theres always a risk to the privacy and security of your contact center. Business leaders need to ensure they have the right security strategies in place to protect sensitive data.

Generative AI is a form of artificial intelligence that can generate new, original content, such as text and images, based on basic prompts. It uses deep learning and neural networks to produce highly creative answers to queries and requests.

Like conversational AI, generative AI is becoming a more common component of the contact center. CCaaS vendors offer companies access to generative AI-powered bots that can provide real-time coaching and assistance to agents or enhance the customer service experience.

Most of these solutions build on the foundations of conversational AI, enhancing bot performance with access to large language models (LLMs).

Alongside leveraging NLP technologies, most generative AI solutions rely on:

Since generative AI tools share many of the same features as conversational AI solutions, they can also address many of the same use cases. Were already seeing an increase in companies using generative AI to create intuitive chatbots and virtual assistants.

However, there are also additional opportunities for generative AI in the contact center, such as:

Generative AI excels at producing original content. It can help contact centers create knowledge bases, drawing on existing data in their ecosystem to design comprehensive guides. Generative AI bots can then surface this information to contact center agents in real-time and offer recommendations to guide them through a conversation.

They can even help organizations create more comprehensive training resources and onboarding tools for new contact center agents, boosting team performance.

Like conversational AI, generative AI tools can have a huge impact on customer service. They can understand the input shared by customers in real time and use their knowledge and data to help agents deliver more personalized, intuitive experiences.

Generative AI solutions can automatically create responses to questions on behalf of an agent and recognize keywords spoken in a conversation to surface relevant information. It can even draw insights from multiple different environments to help answer more complex queries.

One major use case for generative AI in the contact center is the ability to automate repetitive tasks, improving workplace efficiency. Generative AI bots can transcribe and translate conversations like their conversational alternatives and even summarize discussions.

They can pinpoint key action items and discussion trends, automatically classify and triage customer service tickets, and improve the routing process.

Like conversational AI, generative AI has both its pros and cons to consider. It can significantly enhance team productivity and creativity and guide agents through the process of delivering exceptional customer service. It can also help improve team efficiency by automating repetitive tasks like call summarization.

Plus, generative AI solutions can:

However, there are risks to generative AI, too. Like most forms of AI, generative AI relies on access to large volumes of data, which needs to be protected for compliance purposes. It can cause issues with data governance, particularly when teams have limited transparency into how an LLM works.

Plus, since generative AI creates unique original content, its subject to AI hallucinations, which means not all of the answers it gives will be correct.

Conversational AI and generative AI have a lot of overlapping capabilities and features. They both make it easier for human beings to interact intuitively with machines, and they can both understand natural input. However, there are some major differences:

So, conversational AI vs generative AI: which do you actually need?

Though conversational AI and generative AI have different strengths, they can both work in tandem to improve customer experience. Tools like Microsoft Copilot for Sales are considered generative AI models, but they actually use conversational AI, too.

There are various ways contact centers can connect generative AI and conversational AI. For instance, conversational AI bots can generate better answers to customer questions by calling on the insights of back-end generative models.

Smart conversational assistants can analyze inbound ticket information and assign issues to specialized generative models to help with customer service. Conversational bots can even draw insights from FAQs and knowledge bases created by generative AI during discussions.

Ultimately, weaving conversational and generative AI together amplifies the strengths of both solutions. While conversational AI bots can handle high-volume routine interactions in contact centers, solutions powered with generative algorithms can address more complex queries and offer additional support to agents.

The chances are, as both of these technologies continue to mature, well see CCaaS and contact center leaders introducing more tools that allow users to design their own systems that use the best of both models, such as Five9s generative AI studio.

Link:
Conversational AI vs Generative AI: Which is Best for CX? - CX Today