Skip to content
3D illustration of AI items showcasing how Meltwater uses AI

How Meltwater Uses AI

Giorgio Orsi

May 17, 2023

This article was written by Giorgio Orsi, Senior Director for Data Science Innovation at Meltwater.

At Meltwater, we take big data, AI, and machine learning (ML) very seriously, because they have been at the core of what our company does since its very beginning.

The first ML models for sentiment analysis and entity recognition were deployed back in 2005, long before social media was mainstream. Today, we process over 1.3 billion new social media posts, news articles, and blog posts every day. We handle more than 20 billion engagement actions in real time across all social channels and make more than 15 billion inferences through our AI models.

In this blog post, we will walk you through all areas covered by our AI and Data Science teams at Meltwater, and explain how we use AI in all of our products through our unified Fairhair.AI platform:

Tip: Read about the AI ethics we're following.

AI Areas

AI areas are often classified into two main types:  

  • Symbolic (e.g. logic rules and other formal methods) 
  • Sub-symbolic (i.e. statistical) AI which includes machine learning, and famously, deep learning and neural networks.

In Symbolic AI, the knowledge and rules are explicitly programmed into the system, which means that the system has a predefined set of rules to follow. The advantage is that they manipulate data and knowledge in a transparent and easy-to-understand way, however, rules can be limited when it comes to dealing with uncertainty and learning from data.

Machine learning, on the other hand, processes data using statistical models and algorithms. It often involves training models on large datasets to learn patterns and make predictions based on the data. ML is good at dealing with highly complex relationships between data, as well as making inferences about scenarios that have not been encountered before.

The future of AI is very likely going to be a combination of the two approaches where learning and reasoning work together to learn from the data as well as "obeying" the rules, as many applications have certain safety, regulatory, and fairness requirements that must be fulfilled.

Graphic showing the symbolic approach vs. subsymbolic approach in using AI

Natural Language Processing and LLMs

Natural Language Processing (NLP) has been at the core of what Meltwater does for more than 20 years. NLP is a subfield of artificial intelligence dealing with the interaction between computers and human language

At Meltwater, NLP techniques are used to extract, structure, and analyze textual data from news, blogs,  social media platforms, and first-party data in order to gain insights into how brands are perceived by their audiences,

how and where conversations happen, and how they spread. This enables us to derive models of consumer behavior, understand emerging trends, and provide recommendations to PR and marketing professionals.

These are some of the NLP tasks supported by Meltwater's AI stack:

Sentiment Analysis and Emotion Detection

The task of understanding the tone (positive, negative, neutral) and emotions (e.g., anger, happiness, surprise) of a post, article, or entire conversations. Our models are based on Large Language Models (LLMs) to predict both sentiment and emotions.

We support more than 200 languages with >90% accuracy, with more languages on their way. Our sentiment analysis models power the sentiment analytics across all of our product suites.

Tip: Learn more about The What, Why And How of Sentiment Analysis

Sentiment and Emotion Detection at Meltwater

Entity/Keyphrase Extraction and Linking

The task of recognizing relevant entities and key phrases/terms in text.

This task uses machine learning techniques and the power of the Meltwater knowledge graph to detect and link together brands, people, products, locations, and general terms (power cable, screen protector) in more than 90 languages.

Entity and keyphrase extraction algorithms power entity analytics widgets across our products, the entity search capabilities in Explore, and Smart Alerts. It also makes sentiment analysis better, by enabling entity-level sentiment analysis (ELS).

Entity and Keyphrase Extraction at Meltwater

Topic Classification

The task of classifying text into categories, e.g. Sports, Finance, Business.

Topic classification is crucial to cut through the noise and focus the analysis on relevant and highly impactful sources.

Meltwater uses LLMs for topic classification as well and is able to classify content in more than 30 languages using a taxonomy of thousands of categories. Because one size does not fit all, we also support custom categories to help customers refine their searches to even more relevant results.

Topic Classification at Meltwater

User Profiling

Albeit not strictly an NLP task, it builds on many NLP building blocks.

At Meltwater, authors of social media posts are profiled based on features like:

  • age
  • gender
  • profession
  • relationship status
  • language spoken
  • places they mention in their posts
  • topics being talked about or engaged with

Meltwater uses the same technology used by LLMs to create models capable of predicting these profiles for post authors based on millions of other profiles.

Demographic analysis and the ability to automatically identify tribes is crucial for consumer intelligence and social media marketing and enables better targeting of marketing campaigns as well as improving customer service and customer experience.

Our user profiling capabilities power the authors tab in Explore and our Communities product. Our AI model is also able to detect when the account genuinely belongs to a human author rather than automated accounts such as product or company pages. An authority score is also calculated to identify Key Opinion Leaders (KOLs) and community authorities.

User Profiling at Meltwater

Clustering and Summarization

This task produces a meaningful gist of one or more articles and posts related to the same story. This feature enables PR and Marketing professionals to quickly identify relevant stories and the underlying sources without having to sift through thousands of irrelevant mentions.

Our best-in-class, LLM-powered clustering technology automatically identifies articles and posts belonging to the same story, as well as identifying copy-paste re-syndications or plagiarized articles.

Topic Clustering at Meltwater

Our summarization technology then uses generative AI to condense the stories into concise summaries.

Our AI can summarize single or multiple articles and posts in more than 50 languages. Every day we process more than 600k distinctive stories (clusters), each consisting of up to a few thousands articles.

Topic Summarization at Meltwater

Event Detection

Event detection identifies relevant events such as product launches, IPOs, and appointments of key officers. Event detection tracks millions of news documents every day and supports more than 40 different types of events.

Event detection uses advanced neural network technology to identify events and extract key attributes, e.g. the price paid for an acquisition, the name of the officers being appointed. Event detection is supported in both Smart Alerts and Explore, events are then linked to social media posts sharing event-related articles via our advanced LLM-powered clustering technology.

Meltewater Smart Alerts and Event detection

Spam Detection

A media monitoring and social listening product suite is not complete without the ability to separate valuable information from noise.

Our spam detection technology supports more than 200 languages and can identify all kinds of spam content, from not-safe-for-work (NSFW) content to bot-generated content and promotional ads.

This helps PR and marketing teams to cut out the noise that can affect their research and inflate impact and ROI metrics.

Tip: Learn more about Marketing ROI and PR ROI.

Speech Processing

Speech processing refers to the analysis and manipulation of speech signals, usually with the goal of extracting useful information or transforming them for specific purposes.

Speech recognition involves converting spoken words into text, and is used at Meltwater to process the broadcast data (radio, TV, and podcasts) we receive from our partners such as TVEyes. Once the text is extracted, it can be processed as any normal text. Speech processing plays a crucial role in social listening and brand monitoring as short videos and podcasts are now very popular amongst influencers.

Speech processing is supported in more than 20 languages and we process the equivalent of 37k hours of speech from more than 60k stations on a daily basis, roughly equivalent to 1.5k days of speech every single day!

Speech Processing at Meltwater

Computer Vision - Image and Video Analysis

A picture is worth a thousand words, and when it comes to brand management this is really true. Our own research shows that Twitter posts with images receive 2.6 times more engagement than text-only posts. 

Similarly, videos on social media have been found to be highly effective at capturing attention and driving engagement, with Meltwater's data showing twice as much engagement for tweets with videos compared to standard posts. As such, it's critical for a social listening platform to have AI technology capable of understanding images and videos as well as deriving insights.

At Meltwater, we have developed state-of-the-art computer vision technology based on Vision Transformers (the equivalent of LLMs for the visual world). 

Computer vision

Our unified platform currently supports more computer vision tasks than anyone in our industry, including:

  • Logo And Celebrity Detection: identifies brand logos, celebrities, and other VIPs (e.g., company officers) in images, we currently support more than 3.5k logos and 400 celebrities, with more being added every day.
  • Face Detection (Age, Gender, Density): our visual AI can detect the likely age group and gender of people appearing in images at the individual face granularity. This complements our demographic segmentation capabilities provided by our NLP stack. Our AI can even count people in images, so we can answer searches like, "retrieve results where the Nike logo appears with a crowd of people". Pretty neat!
  • Scenes And Object Detection: our AI can identify more than 700 different objects and over 120 scenes and situations. This enables searching for e.g., a brand logo in certain contexts, such as the brand logo in a sport event vs a MET Gala.
  • Emotion Detection: this capability enhances our sentiment and emotion NLP capabilities with the ability to detect the same types of emotions in images for a fuller listening experience across all media channels.
  • Memes Detection: in today's social media landscape memes can go pretty viral, no social listening suite is complete without meme tracking and detection.
  • Object Character Recognition: the task of recognizing written words inside images. No social listening suite is complete without being able to identify printed content inside visual images. This capability enhances our search and listening capabilities in Radarly and Explore, complementing our logo detection to never miss a brand mention.
  • Image Clustering: leveraging the power of visual transformers, our AI can identify similarities between images and create clusters of similar images enabling an unprecedented level of discovery and control in social listening applications.
Image Clustering at Meltwater

Visual AI technology powers image search and analytics capabilities in Radarly and Explore, as well as our Meltwater Influencer Marketing suite.

Tip: Learn more about the basics of image recognition and take a look at the best image recognition software.

Knowledge Graph

A knowledge graph is a type of database designed to represent and store knowledge in the form of entities (nodes) such as people, places, brands, products, and their relationships (edges). For example, John Box is Meltwater's CEO, iPhones are made by Apple, John Simpson writes for the BBC. 

Knowledge Graph Example
Knowledge Graph Example

Knowledge graphs are often used to power intelligent applications such as search engines, AI personal assistants, and recommendation systems. Some examples of knowledge graphs include Google's Knowledge Graph, which powers its search engine and provides information about entities such as people, places, and things, and Facebook's Social Graph, which represents the relationships between people and their interests.

Meltwater's knowledge graph contains more than 20M entities and 70M relationships, and includes companies and their officers from more than 300 industries. The knowledge graph powers our Smart Alerts, Explore's entity search, as well as our Media Relations, Influencer Marketing, and Sales Intelligence products.

We also use it to automatically recommend what competitors to follow or monitor through our AI-based industry classifiers that automatically assigns companies to their industries, linking competitors together.

Time Series Analysis And Predictive Analytics

Time series analysis is the process of understanding phenomena, patterns, trends, seasonalities, and other more exotic data relationships over time. It is used to make predictions about future values based on past observations. 

At Meltwater, our predictive analytics team crafted AI models to understand engagement and mention trends, customer behavior, identify and explain anomalies (e.g. spikes), and make predictions about future trends. These algorithms power our Smart Alerts product to notify users about spikes in mentions and engagement, shifts in sentiment, identify sponsored Facebook posts, and viral tweets, all in real time. Predictive analytics algorithms also power our spike detection inside our new Discovery offering.

Tip: Learn more about the Meltwater AI Spike Analysis feature!

Time Series Predictive Analytics at Meltwater

Language Detection and Geo Localization

Language detection and geo localization are the task of determining a user's language and location based on certain features. In some cases, Meltwater receives this information directly from our data providers, however, we always validate the location or infer it when missing using AI technology.

AI-powered language detection is supported in more than 240 languages but out platform supports search and retrieval in all known human languages (more than 7k!). Language and Geo localization are crucial for demographic analysis and customer segmentation, as well as enforcing our strict data compliance rules.

Depending on the platform, our AI is able to determine the location of a user down to the granularity of a city's district based on information available in the post, the biography of an author, and sometimes the IP address of the device used to publish the post. Currently, all recognized ISO countries and territories are supported across all Meltwater's products.

The Future of AI at Meltwater

Now that we have looked at the past and current state of Meltwater's AI, it's now time to turn our attention to its future.

As an innovation leader and AI-first company, Meltwater is doubling down on its investments in AI and Machine Learning. Our Scientific Advisory Board provides insights into shifts in the frontier of AI's technological innovation and helps to identify novel solutions or technologies for some of our most difficult AI challenges.

Top locations graph Meltwater

AI technology has a massive potential for the PR and marketing industries, and Meltwater is baking AI assistive technology in all of its products, including our enterprise and delivery services.

Stay tuned for our upcoming blog series on AI, we'll walk you through our exciting new product features including our generative AI capabilities, ground breaking video analysis, our GPT-powered PR and Engage Assistants, AI-powered custom classifiers, semantic search, spike analysis, recommendation and prescriptive analytics and a lot of tips and tricks for using our AI-powered features effectively.

Tip: Check out how Meltwater’s AI Writing Assistant for social media managers works, and learn more about Meltwater's AI-Powered PR Assistant