- Unveiling the Realm of AI Technologies: A Glimpse into the Augmented Future
- Generative AI and Language Models
- LLMs and Their Applications
- LLMs and Customer Support
- Development, Optimization, Localization, and Personalization Based on LLMs
- Unveiling the Power of Clustering and Topic Modeling
- Enhancing Customer Support Through Hybrid AI: LLMs Meet Clustering and Topic Modeling
- Endnotes
This chapter is from the book
This chapter is from the book
This chapter is from the book
Enhancing Customer Support Through Hybrid AI: LLMs Meet Clustering and Topic Modeling
Customer support is evolving, and businesses seek more sophisticated and powerful solutions to handle the vast amount of textual data generated in interactions. A hybrid approach, blending the capabilities of LLMs and traditional machine learning techniques, emerges as a robust strategy. We’ll explore a few of these machine learning techniques often utilized in support organizations to make sense of the large amounts of data to help optimize the business.
Clustering and Customer Support
Clustering is an unsupervised learning approach of grouping a set of samples based on their similarity without using any predefined labels or categories. Clustering aims to discover the natural structure or patterns of the data, as well as to reduce its complexity and dimensionality. Clustering can be used for various purposes, such as data exploration, summarization, organization, retrieval, and visualization. There are several different clustering methods:
Hierarchical clustering: This method builds a hierarchy of clusters, where each cluster is either a subcluster or a supercluster of another cluster. Hierarchical clustering can be either agglomerative or divisive. Agglomerative clustering starts with each sample as a singleton cluster and then merges the most similar clusters until a single cluster remains. Divisive clustering starts with all documents in one cluster and then splits the most dissimilar clusters until each cluster contains only one sample.
Partitioning clustering: This method divides the data points into a predefined number of non-overlapping clusters, where each point belongs to exactly one cluster. K-mean clustering is one of the most popular algorithms for partitioning clustering. Partitioning clustering can be either distance-based or centroid-based. Distance-based clustering assigns each data point to the cluster with the closest or most similar representative, such as the nearest neighbor. Centroid-based clustering assigns each data point to the cluster with the smallest or least average distance to the center or the cluster’s mean, such as K-mean clustering. K-mean clustering classifies samples based on attributes or features into k clusters. It starts with a first group of randomly selected centroids, which are used as the beginning points for every cluster, and then assigns each point to the cluster whose mean has the least squared Euclidean distance and optimizes the centroid based on the distances from the points to it. The hard assignment stops creating and optimizing clustering when either the centroids have stabilized or the defined number of iterations has been reached.
Density-based clustering: This method identifies clusters based on the density or the concentration of the data points in the feature space, where regions of separate low-density clusters can be uncovered and assist in identifying unforeseen patterns. Density-based clustering can handle outliers, noise, and arbitrary shapes of clusters. One of the popular algorithms for density-based clustering is DBSCAN (density-based spatial clustering of applications with noise). DBSCAN defines a cluster as a set of densely connected core points; a point is a core point if it has at least a minimum number of points within a given radius or neighborhood.
Clustering is a powerful technique for identifying patterns and insights from large and complex data sets. It can be used to segment customers, optimize services, categorize issues based on their similarities or differences, and provide personalized and efficient solutions. In the field of customer service and support, clustering has been a popular approach for solving some problems, such as:
Customer segmentation: Clustering can help discover different groups of customers based on their demographics, preferences, needs, behaviors, or characteristics, such as age, gender, location, income, spending habits, loyalty, satisfaction, or feedback. This can help tailor the marketing strategies, product recommendations, pricing policies, or communication channels for each segment and to improve customer retention and acquisition.
Service optimization: Clustering can help optimize the service delivery and support processes based on the complexity, urgency, or frequency of customer requests, issues, or inquiries, such as order status, product information, technical support, billing, or feedback. This can help allocate the appropriate resources, staff, or channels for each service type and improve service efficiency and quality.
Support case categorization: Clustering can help resolve customer issues faster and more effectively by grouping similar or related issues based on their causes, symptoms, or solutions, such as product defects, software bugs, network failures, or user errors. When AI technology is used to cluster similar cases together, these groupings can help by offering new insights that are not obvious when looking at cases individually or by product. An example might be multiple unrelated services experiencing login or profile creation issues. Viewed on their own, these could be hard to relate or determine the root cause of the issue, but after clustering them together, it might be more obvious that this is a problem with shared code providing identity services to multiple workloads. This clustering can help diagnose the root causes, find the best solutions or prevent future occurrences of the issues, increase customer satisfaction, and enhance retention.
Topic Modeling and Customer Support
Topic modeling is a technique for extracting hidden topics or concepts from a collection of text documents, such as customer reviews, feedback, complaints, or inquiries. Topic modeling can help discover the main themes or patterns of customer needs, preferences, opinions, or issues and provide valuable insights for customer support improvement, product development, marketing strategy, or sentiment analysis.
There are several different topic modeling methods. These algorithms differ in their assumptions, mathematical models, and implementations, but they all share the same basic idea: finding a low-dimensional representation of the documents and the words in terms of topics and probabilities. The output of a topic modeling algorithm is usually a matrix that shows the relationship between documents and topics, and another matrix that shows the relationship between topics and words. These matrices can be used to infer the topics of new documents, find similar documents, visualize the topics, and extract insights from the text data. These methods include:
Latent Dirichlet Allocation (LDA): This is one of the most popular topic modeling methods. LDA is an unsupervised learning algorithm that describes a set of observations as a mixture of distinct categories. These categories are themselves a probability distribution over the features. LDA is most commonly used to discover a user-specific number of topics shared by a collection of documents within a text corpus. Each observation is a document, the features are the presence or occurrence count of each word, the categories are the topics. LDA uses a generative process to assign topic probabilities to each document and word probabilities to each topic, based on the observed word frequencies in the documents. LDA can be applied to large, diverse text corpora and produce interpretable and coherent topics.
Non-negative Matrix Factorization (NMF): NMF is a linear algebra method that decomposes a matrix of word-document frequencies into two lower-dimensional non-negative matrices, one representing the word-topic associations and the other representing the topic-document associations. NMF imposes a non-negativity constraint on the matrices, which ensures that the topics and the documents have additive and meaningful components. NMF can be faster and more robust than LDA and can handle sparse and noisy data.
Hierarchical Dirichlet Process (HDP): HDP is a Bayesian nonparametric model that extends LDA by allowing the number of topics to be automatically inferred from the data rather than fixed in advance. HDP uses a hierarchical structure of Dirichlet processes to generate a potentially infinite number of topics and assigns them to the documents based on their relevance and specificity. HDP can adapt to the complexity and diversity of the text data and can avoid overfitting or underfitting the topics.
Topic modeling is a valuable technique in the customer service and support field for extracting insights from large volumes of textual data, such as customer reviews, feedback, and support cases. Here’s how topic modeling is leveraged in this domain:
Automated support case categorization: Customer support teams often deal with a variety of issues and requests. Topic modeling can be leveraged to automatically categorize support tickets into different topics or categories based on their content. This helps in routing tickets to appropriate product support teams and improves response time and efficiency. Moreover, topic modeling can help automate some processes in the customer support workflow. For example, it can point customers to the self-help knowledge base, diagnostics, or websites with the accurate topic category prediction. This can enhance the customer experience, reduce customer effort, and increase operational efficiency.
Identifying emerging issues: Topic modeling can help uncover emerging trends or issues in customer feedback and support cases. It provides actionable insights for companies to address top issues before they escalate proactively.
Improving search and retrieval: Topic modeling helps organize and index articles based on the topics for a large knowledge base of support or self-help articles. This improves the search and retrieval process for support agents or engineers and the customers looking for solutions.
Customer feedback analysis: Topic modeling can help analyze and summarize customer feedback from multiple channels and platforms. This can help identify the most common and important topics, issues, compliments, complaints, and suggestions that customers express. This can also help products and companies measure and track key performance indicators related to customer support, customer satisfaction, and loyalty. For instance, it can help measure the volume of support cases in different categories, identify resolution time, and assess customer satisfaction for each topic. Furthermore, product teams can prioritize and address customer complaints and grievances more effectively.
Content creation and knowledge management. Topic modeling aids in content creation for FAQs, manuals, and support articles. It helps identify the most discussed topics, allowing companies to create relevant and helpful content that addresses common customer queries.
In essence, topic modeling enhances the efficiency and effectiveness of customer service and support operations by providing automated tools for organizing, analyzing, and extracting insights from large volumes of textual customer data.
Hybrid AI Opportunity
Traditional machine learning methods like topic modeling and clustering have their own limitations and challenges. One of the main drawbacks is that they rely on statistical methods that do not account for the semantic and contextual nuances of natural language. For example, topic modeling may fail to distinguish between different meanings or senses of the same word, such as apple as a company but not as a fruit, or group together words that are syntactically similar but semantically different, such as bass as a type of fish but not low-frequency sound in music. Moreover, topic modeling may produce topics that are too broad, too narrow, or not coherent, depending on the choice of parameters and algorithms. In contrast, large language models, such as GPT and Gemini, have demonstrated remarkable proficiency in understanding context, generating human-like responses, and extracting intricate patterns from textual data. In customer support, LLMs can be employed for tasks like sentiment analysis, intent recognition, and even generating responses to common queries.
While LLMs excel in understanding context and generating text, traditional machine learning methods like clustering and topic modeling offer strengths in structuring and organizing information. Clustering can group similar customer queries or issues, facilitating efficient handling by support agents. Topic modeling, on the other hand, extracts underlying themes from a vast dataset, aiding in understanding prevalent customer concerns. Moreover, when computational resources and budget are limited, it is easier and cheaper to leverage traditional machine learning methods like topic modeling and clustering.
In the dynamic landscape of customer support, a hybrid approach, integrating the capabilities of LLMs with the structuring prowess of traditional methods, proves to be a holistic solution. By combining LLMs with topic modeling, more accurate, robust, and interpretable models can be utilized for customer feedback analysis. For instance, language models can help generate more natural and fluent texts from topics and can also help capture the semantic and contextual information that topic modeling may miss. Furthermore, LLMs can help generate new and novel topics that may not be present in the existing data or suggest relevant and personalized content based on the topics of interest of each customer, while topic modeling and clustering can bring more interpretability and flexibility. This hybrid solution addresses the complexities of customer interactions, providing businesses with a powerful tool for improving customer satisfaction and support efficiency.