One problem business leaders face when communicating with their technical counterparts is trying to describe their AI problem types. To simplify some of the communication, here are 22 common AI problem types.
1. Classification
A classification problem is about assigning one or more categories to a document, product, person, or image—essentially anything. Examples include:
- Categorizing incoming support tickets by relevant topics
- Classifying images of silicon wafers as containing defects or no defects
2. Regression
A regression problem is about estimating numerical values given some input. For example, trying to predict the number of months before a machine needs service given the conditions of the current machine, or predicting how specific drug dosage affects blood pressure.
3. Recommendation
A recommendation problem is about suggesting personalized content or products to a group of people. Examples include:
- Product recommendation (e.g., Amazon)
- Recommendations on who to follow (e.g., LinkedIn)
- Recommendations on jobs to apply for
- Recommendations on articles to read (e.g., Medium)
4. Search Relevance
A search relevance problem is about improving the rankings of search results shown to users. Often search relevance improvement starts with the analysis of search logs to diagnose problems using hard data. Search improvement may or may not require heavy use of machine learning.
5. Information Extraction (IE)
An information extraction problem is about extracting specific information from large volumes of text data. One of the goals of information extraction is to fill templates using data extracted from raw text. Examples include:
- Extracting patient symptoms from large volumes of clinical notes.
- Extracting pertinent information from large volumes of legal case files
- Pre-populating a candidate application form/database by extracting pertinent information from resumes
6. Text Summarization
Text summarization is about creating an accurate synopsis of a longer document or a set of documents.
7. Clustering
Clustering is about grouping people, content, documents, topics and etc based on some logical structure—for example, grouping customers by their purchase behavior.
More generally, clustering divides data points into a number of fixed (or dynamic) groups such that the data points in one group are more similar to each other than data points in other groups.
Example: Grouping Customers by Purchase Behavior.
A retail store collects data on how customers spend money across different product categories:
A clustering algorithm groups customers by shopping habits:
Grocery Shoppers → Alice, Charlie (High grocery spending, low on others)
Tech Enthusiasts → Bob, Dave (High electronics spending)
Fashion Shoppers → Eve (High clothing spending)
This type of analysis can be used by businesses to personalize promotions and improve customer experience with targeted recommendations.
8. Time Series Forecasting
Time series forecasting is the process of using historical data points, typically collected over time at regular intervals, to predict future values. It is widely used in business and industry for applications such as demand forecasting, stock price prediction, sales trends, and resource planning.
By analyzing patterns like seasonality, trends, and cyclic behaviors, time series forecasting models help businesses reduce costs and increase revenue by assisting them in optimizing tasks such as inventory planning and keeping up with changes in seasons and other external factors that might be out of their control.
9. Virtual AI Assistant
Virtual AI Assistant is used for having short conversations with humans to complete simple tasks. Examples include:
- Providing answers to common customer questions without human involvement
- Using text messages to check bank balances or to get a refund
Alexa and Siri are examples of virtual AI assistants.
10. Sentiment Analysis
Sentiment Analysis is about discovering emotions and feedback patterns in text data such as user reviews, social media comments, and surveys. Sentiment analysis can even be applied to images for understanding emotions from facial expressions.
Example: Here’s an example of extracting sentiments related to specific aspects such as customer service, product quality, pricing, and overall experience, determining if each aspect is viewed positively, negatively, or neutrally.
Customer Review:
“I can’t say enough good things about Pretty Petals and their incredible team. From the moment I stepped into their charming store, I was greeted with warmth and enthusiasm. The staff was attentive, offering expert advice on flower selection and care, and they crafted a gorgeous arrangement that exceeded my expectations. However, I did find the prices to be a bit on the higher side compared to other florists in the area. Their work’s quality and presentation are outstanding, and I will continue to trust Pretty Petals with all my floral needs.”
Sentiment Analysis Output:
The review shows a mostly positive experience with Pretty Petals, highlighting great customer service, atmosphere, and flower arrangements. The only downside mentioned was the higher pricing compared to other florists, but the reviewer still expressed trust in the company and plans to return.
11. Object Detection
An object detection problem is about discovering specific objects such as humans, buildings, or cars in digital images and videos.
12. Document Segmentation Problem
Document segmentation is about trying to subdivide documents into meaningful parts. For example, segmenting unstructured clinical texts to extract their past medical history and family history.
13. Keyword Extraction
Keyword extraction is about identifying terms that best describe the subject of a document—for example, extracting keywords from large volumes of legal documents to understand the themes of discussion.
While there are many keyword extraction tools readily available (including open-source tools), you’d need to ensure that these work on your data. Often, keyword extraction tools are best customized or custom-developed. To show how keyword extraction works, we provide a few pages from a research paper and leverage a keyword extraction algorithm to extract some relevant keywords.
Extracted Keywords:
1. Vertical AI Agents
2. Context-Aware Systems
3. SaaS Platforms
4. Real-Time Adaptability
14. Speech Recognition
Speech recognition, also known as speech-to-text (STT) or automatic speech recognition (ASR), is about having a computer program understand and transform spoken language into a written format (or text).
Speech recognition is often used to complete downstream tasks. For example, speech recognition is used behind the scenes to surface relevant search results when you use Google voice search. Specifically, your speech is translated into a human-readable format, and that generated text is used to surface relevant search results.
Many vendors offer speech recognition solutions, and therefore, speech recognition systems rarely need to be developed from scratch. Of course, these systems will benefit from customization for the target data.
15. Machine Translation
Machine translation is the automatic software translation of text from one language to another. For example, translating English sentences into German with reasonable accuracy. Machine translation programs rarely need to be developed from scratch but may benefit from customization.
Machine translation is used for many purposes, including:
- Localizing website text for a particular country
- Customer support conversations across countries
- Understanding documents written in a different language
16. Paraphrasing
Paraphrasing is the process of rewording text while keeping its original meaning. It can help make documents easier to read or explain complex ideas in simpler terms.
Let’s consider an example: We provided a text to the Large Language Model (LLM) Mistral for paraphrasing.
Here’s the response we got.
17. Named Entity Recognition (NER)
NER is a common NLP task that identifies entities like people, organizations, or locations in text. These entities can be used for various applications, such as gaining insights from company documents, improving search engines, or building a structured database from a corpus.
Let’s identify the named entity in this text:
“Apple is looking at buying U.K. startup for $1 billion.”
18. Question-Answering
Question answering is a natural language processing task where a model is given a question and is expected to extract an accurate answer from the provided knowledge source. There are many approaches for extracting answers from questions from extractive approaches to more abstractive approaches. In some cases, large language models are used to perform question-answering tasks due to their ability to understand and generate accurate answers. Here’s an example of how it works:
Example:
Question: “What are the benefits of exercise?”
Text: “Exercise is good for heart health, builds muscle strength, and can help with weight loss.”
Answer: “Exercise helps improve heart health, build muscles, and manage weight.”
19. Retrieval-Augmented Generation (RAG)
For more advanced use cases where you need to find answers from your data sources, RAG is a more robust approach. RAG dynamically retrieves relevant information from external or proprietary databases before generating an answer, making it highly effective for domain-specific applications. To learn more about RAG, you can look at this article
20. Automated Feature Engineering
This is the process of automatically extracting useful features from raw data, which can then be applied to various machine learning tasks. For instance, state-of-the-art LLMs can be prompted to extract abbreviations from a given text, making it easier to handle domain-specific language or simplify complex data. You can also perform tasks like extracting key information from text. For example, the following shows how to extract price and trends from text.:
Input Text: "The price of gold decreased from $3,032 to $3,026 per ounce in the past week."
Extracted Feature: {"price_change": 6, "price_trend": "decrease"}21. Optical Character Recognition (OCR)
OCR is a technology that converts different types of documents, such as scanned documents, PDFs, or images, into editable and searchable data. It enables computers to recognize and extract text from images, which can then be processed and analyzed using NLP techniques. This allows for automated data entry, document classification, and information extraction from visual content.
22. Textual Entailment
Textual entailment is the task of determining whether one statement (hypothesis) logically follows from another (premise). It helps in sentiment analysis, and customer feedback analysis, by ensuring AI understands and verifies meaning accurately. For example, it can detect if a customer review implies satisfaction without explicitly stating it.
Example 1: Positive Entailment
Premise: “The company delivered the product two days earlier than expected.”
Hypothesis: “The customer received the product ahead of schedule.”
Entailment: True (Yes, the hypothesis logically follows from the premise.)
Example 2: Contradiction
Premise: “The smartphone has a battery life of 12 hours on a full charge.”
Hypothesis: “The battery lasts only 5 hours after a full charge.”
Entailment: False (The hypothesis contradicts the premise.)
Example 3: Neutral
Premise: “Mohammad Ali is the most renowned boxer of all time”
Hypothesis: “Mohammed Ali likes playing golf.”
Entailment: Neutral (The premise provides no information about Mohammad Ali liking golf.)
Summary of AI Problem Types
In this short guide, we discussed 22 common AI problem types that often overlap. For example, you can apply a classification approach for sentiment analysis.
However, the key is to identify the problem type that best fits the task at hand. It doesn’t have to be 100% accurate—it’s just semantics. You can continually refine these definitions with the help of your AI experts.