Using knowledge graph-based LLM for relation & event detection

Use Cases

Peio Popov, Mila Mutafchieva

·

·

Reading time: approx. 9-10 minutes

The Superpowers of Ontotext’s Relation and Event Detector

The Superpowers of Ontotext’s Relation and Event Detector

This post originally appeared on the Ontotext website and is republished and edited with the permission of our partners at Ontotext. In this blog post, we explore Ontotext’s latest knowledge graph-powered solution that works with LLMs to transform raw news content into actionable data for events impact assessment and risk and opportunity detection. 

 

Using knowledge graph-based LLM for relation & event detection

Ontotext’s Relation and Event Detector (RED) is designed to assess and analyze the impact of market-moving events. RED answers key questions such as: “What happened?”, “Who was involved?”, and “What is the financial impact?”. The answers to these foundational questions help you uncover opportunities and detect risks.

 

From a technological perspective, RED combines a sophisticated knowledge graph with large language models (LLM) for improved natural language processing (NLP), data integration, search and information discovery, built on top of the metaphactory platform. Entity linking allows events to be associated with specific companies in the graph and correlated with information from 3rd party databases, namely Crunchbase, and public information about stock prices. Further, RED’s underlying model can be visually extended and customized to complex extraction and classification tasks.

 

RED’s focus on news content serves a pivotal function: identifying, extracting, and structuring data on events, parties involved, and subsequent impacts. A challenge we continuously meet across our projects is spotting events with significant influence for organizations – either transformative or disruptive. We bundle these events under the collective term “Risk and Opportunity Events”.

 

This post is part of Ontotext’s AI-in-Action initiative aimed to empower data, scientists, architects and engineers to leverage LLMs and other AI models. Let’s dive in to see what these events are, how we define them, and how the prowess of our technology can be harnessed to decode news-based content focusing on such events.

 

Table of contents

 

 

Why do risk and opportunity events matter?

A risk and opportunity event refers to an occurrence that may positively or negatively impact the stock market performance of a company or industry sector. The ability to detect and analyze such events is crucial for several reasons:

 

  • Investment decisions: Investors rely on understanding how various events impact a company’s performance. Investors make informed decisions about buying, holding, or selling stocks by analyzing these events.

  • Risk management: Understanding the correlation between events and stock price fluctuations helps manage risk. Investors and companies can anticipate potential price movements and plan accordingly, reducing the chances of significant losses.

  • Market sentiment analysis: Events can significantly influence market sentiment. For instance, news about a particular regulatory action might impact a single company and the entire sector.

  • Strategic planning and predictive analytics: Companies can use this analysis for strategic planning. Understanding how certain types of events have historically affected their stock prices can guide future business decisions and communication strategies.

  • Economic indicators: Events affecting major companies or industries can indicate broader economic trends. For instance, significant layoffs in a leading industry could signal economic downturns, impacting individual stock prices and the overall market.

  • Competitive analysis: Companies can also use this analysis to understand competitors’ strengths and weaknesses. For example, how competitors’ stock prices react to certain events can provide insights into market positioning and consumer preferences.

 

Closely following how specific events impact company and industry stock performance can be essential for informed investment, effective risk management, strategic corporate planning, understanding market sentiment, and economic analysis.

 

RED’s technological capabilities

RED allows you to perform knowledge graph-backed information extraction, based on a predefined event extraction schema. It showcases several capabilities of the products and solutions offered by Ontotext’s technology ecosystem, such as:

 

  • Knowledge graph building, including ontology modeling, diverse data integration, and record liking across public information and non-proprietary databases.

  • Leveraging LLMs to boost information extraction tasks, such as named entities recognition, entity linking, and event and relationship extraction.

  • Quality assurance process, covering gold standard creation, extraction quality monitoring, measurement, and reporting via Ontotext Metadata Studio.

  • Semantic model management – events schema and instance management as well as taxonomical alignment.

  • Information retrieval, by the means of hybrid and faceted search and topical pages by entity or event.  

 

What makes RED special

The superpower of RED is the ability to correlate customizable company news events with stock price fluctuations. It can make these connections clearer and more accessible, enabling users to derive actionable insights and make data-driven decisions. 

 

Here’s how their tool makes it work. 

 

For every piece of processed content, RED visualizes the corresponding stock price fluctuation on a timeline, enabling users to quickly spot any immediate correlations between events and price changes. For instance, the announcement of the Microsoft-Blizzard merger might coincide with a sharp increase in the stock prices of both companies.

 

 

Relation Event Detection (RED) Demo

 

Using machine learning, RED indicates the impact of events on stock prices. It compares actual price changes to expected changes based on historical data. Then it presents customizable insights through an interactive dashboard for thorough analysis.

 

Let’s have a quick look under the hood.

 

Model-driven extraction

Model-driven extraction refers to the extraction and classification of unstructured data sources, such as text documents and web pages, based on a predefined schema, often referred to as a semantic model. This semantic model serves as a blueprint or framework against which raw data is analyzed and organized. The key capability of model-driven extraction is its ability to intelligently and dynamically extract relevant data by understanding the context, significance, and relationships embedded within the source material. 

 

The semantic model leverages ontologies and taxonomies. Taxonomies, with their hierarchical structure, enable the identification of broader categories or themes present within the content. Ontologies, on the other hand, provide a detailed representation of relationships and entities, allowing the system to discern the context in which terms and entities appear.

 

The benefits of model-driven extraction

Model-driven extraction is adaptable and customizable. It can adjust based on the specific domain or event model applied. For instance, if a company has a unique taxonomy related to its line of business, the extraction process can be tailored to prioritize information relevant to that specific task.

 

The process is automated and scalable, overcoming the challenges of traditional extraction methods. Manual extraction is time-consuming and prone to many inconsistencies due to human error. Traditional automatic approaches might be slow, hard, and, as a consequence, expensive to develop. Model-driven extraction ensures flexibility and can handle large volumes of data efficiently.

 

This method also facilitates the integration of extracted data into knowledge graphs, which allows dynamic linking and enrichment of data representation. The integration enables the interlinking of related concepts and supports advanced applications such as recommendation systems, faceted search, and data analytics.

 

The unique and critical capability, needed to combine knowledge graphs and LLMs, is the so-called entity linking – the task of associating entity mentions in the text to concrete entity identifiers and this way to nodes in the graph. LLMs are much more accessible than any other techniques in recognizing relationships and events in text. What they cannot do well on their own is entity linking. Without it is impossible both to use graph data to interpret the new events and to extend the graph with new relationships.

 

When knowledge graphs & LLM join forces

By leveraging the power of semantic models, model-driven extraction takes raw, unstructured information and transforms it into structured, meaningful, and actionable insights. The following is some of its main benefits:

 

  • Quality – by relying on well-defined models, the extraction process minimizes noise and focuses on truly relevant information
  • Efficiency – automation reduces the time and resources traditionally needed for data extraction tasks
  • Flexibility – LLMs allows end-users to define new event types to be detected, instantly, without the involvement of data scientists or engineers
  • Competitive insights – interpret freshly signals with domain knowledge combining rich 3rd party databases with proprietary data and public information.
  • Consistency – semantic models ensure a uniform extraction process, even across diverse data sources
  • Discoverability – extracted data, when integrated into systems like knowledge graphs or databases, becomes easier to search, navigate, and analyze

 

The power of storytelling in action

Storytelling has the ability to help us understand and feel more connected to large-scale business events, which we usually find dry and boring. By telling a story, we can contextualize such events and change them from distant and boring to interesting and relatable.

 

Let’s consider the merger of two giant tech companies as a series of news events. We can follow it through it as a story that unfolds in stages: the initial rumor, the announcement, the reaction, the hurdles, the process, and, finally, the outcome. Each event has a lot of angles and perspectives that yield different stories. 

 

The recent merger saga between Microsoft and Activision Blizzard provides a good illustration. A simple Google search shows over 3000 hits.

 

 

As you see in the diagram above, these stories present the merger not just as a business but as a human event. It has implications that resonate with individuals, creating moments of tension, excitement, fear, and hope. The power of storytelling lies in this ability to turn dry business events into tangible human experiences. It invites the audience to see themselves in the story and creates a more profound understanding of the event.

 

Transforming content into data

News content is a rich source of information that, when effectively harnessed, can be converted into actionable data. This data can fuel a range of activities from strategic decision-making to predictive analysis. In this context, RED emerges as a valuable tool designed to extract events from news and convert them into actionable data points. 

 

RED’s event extraction pipeline leverages NLP and the latest generation of LLMs. We illustrate how this works with the example of the Microsoft-Blizzard merger:

 

 

Through these steps, RED transforms raw news content into a structured dataset, providing a more manageable and actionable format for analysis and decision-making.

 

The Knowledge Graph as RED’s secret sauce

The secret power of RED is using a knowledge graph as an effective way to contextualize and enrich the data we collect and analyze. In the specific case of the Microsoft-Blizzard merger, the knowledge graph enhances RED’s analysis. It provides information about Microsoft and Blizzard beyond just the merger, including company information, key people, previous mergers or acquisitions, and major milestones. 

 

The knowledge graph also maps out the key people associated with both companies such as CEOs, board chair and members, and other significant personnel. Furthermore, RED links the current merger to previous similar events – these may include other tech industry mergers or past mergers involving either Microsoft or Blizzard. 

 

All of this contextual information sheds more light on the histories of the two companies and their strategic decisions, which enables users to gain market insights and make informed decisions. By leveraging a knowledge graph, RED provides a comprehensive and multi-dimensional view of the Microsoft-Activision Blizzard merger event. This makes the system not just a news event extractor, but a powerful contextual analysis platform.

 

To wrap it up

Ontotext’s Relation and Event Detector transforms the intricate landscape of business information into actionable insights. Powered by advanced technologies, RED not only addresses the challenges of contextualizing unstructured text. It also enables users to understand the nuanced relationships and historical contexts within complex content. 

 

RED can make a big difference when applied to real life events, showcasing how it can correlate news articles with stock price fluctuations. By seamlessly blending storytelling with technology, RED humanizes large-scale business events, making them relatable and providing deeper understanding. 

Do you want to see what Ontotext’s RED is capable of?

 

Contact Ontotext for a free consultation

Peio Popov

Peio Popov is a legal counsel with strong IT background, specializing in business development, intellectual property, and technology-related contract negotiations. Besides being highly skillful in business operations, Peio’s professional focus spans software licensing and development, free software and Open Data licensing issues, text analysis and Information extraction, PKI and Identity services and applicable cryptography, among others.

Mila Mutafchieva

Mila's background is rooted in account management within Advertising and international Event Management. Prior to joining Ontotext, she contributed to a fledgling London-based startup focused dedicated to building audio ML for Ford's spin-off company, Canopy. Currently, Mila is a project manager, leading the RED work stream at Ontotext.