Text Mining with Small Language Models: Unlocking Data with Composable Language Models

In the latest episode of the Data Product Workshop Podcast, we dive deep into the advanced capabilities of Small Language Models (SLMs) and their application in real-world text mining use cases. From extracting information to building composable workflows, SLMs are transforming how organizations interact with unstructured text data.

The Power of SLMs in Text Mining

Small Language Models (SLMs) offer a lightweight yet powerful alternative to massive LLMs, enabling targeted and cost-effective solutions for text extraction and classification. As part of a Composable Enterprise, SLMs are particularly suited for creating Composable Language Models (CLMs) that retrieve, extract, and interpret data from documents through structured prompts and other advanced techniques.

How It Works: Composable SLM Pipelines

SLMs form key components of a text mining process that involves multiple stages:

  1. Ingestion Flow: Extract and classify text data from various formats like PDFs, CSVs, HTML files, and more.
  2. Data Product Execution Graph: Raw classified data, enhanced by SLMs, flows into a virtual data fabric for processing and integration.
  3. Querying the Graph: SLMs are used to interact with the data graph and generate insights through natural language prompts.

Breaking Down the Text Mining Process

Hereโ€™s a closer look at the end-to-end ingestion flow for text mining with SLMs:

  1. Doc to Text:

    • Use tools like PDFMiner and OCR for text extraction.
    • Extract text from tables and structured documents.

  2. Sentence Extraction:

    • Tools like Spacy and OpenNLP help extract meaningful sentences for further analysis.

  3. Table Extraction:

    • Models like table-transformer are used to analyze and extract data from complex tables within documents.

  4. Data Point Extraction:

    • Techniques like Named Entity Recognition (NER) (e.g., BERTTopic) classify and tag relevant data points.
    • Prompts convert data into structured JSON formats for use in downstream processes.

Connecting SLMs to the Data Product Graph

One of the most exciting parts of this workflow is connecting SLMs with a data product graphโ€”a composable execution graph that integrates vector data and enterprise data. By leveraging functional calling, SLMs can seamlessly retrieve, organize, and extract insights from the data fabric.

For example:

  • SLMs query the data product graph to return actionable insights for business users.
  • Results are delivered via a clean and intuitive Query UX/UI, making it easy for analysts and business users to interact with the underlying data.

Real-World Use Cases of SLMs in Text Mining

  • Information Extraction: Extract specific entities, topics, or keywords from unstructured text using NER and topic extraction techniques.
  • OCR-Based Table Extraction: Automate the extraction of tables and structured data from scanned PDFs and documents.
  • Automatic Taxonomy Creation: Use SLMs to identify relationships and organize extracted data into meaningful taxonomies.
  • Document Understanding: Generate insights from long-form text, creating structured outputs for downstream analytics.

Composable Language Models: A Game-Changer

The ability to combine multiple SLMs as Composable Language Models (CLMs) creates a flexible and scalable architecture. Instead of relying on a single monolithic model, CLMs allow you to:

  1. Deploy smaller, task-specific models for different components of a problem.
  2. Improve cost-efficiency by running models on lower-end GPUs or edge infrastructure.
  3. Integrate seamlessly into a composable enterprise framework, enabling modular and adaptable workflows.

Join Us for the Podcast

If you're curious about leveraging SLMs for advanced text mining and document understanding, tune into our latest Data Product Workshop Podcast. Weโ€™ll cover the full range of NLP approaches, including:

  • Named Entity Recognition (NER)
  • Topic Extraction
  • OCR-Based Table Extraction
  • Functional Calling for Integrated Workflows

Discover how composable, lightweight models can unlock tremendous business value without the heavy infrastructure of massive LLMs.

At Dataception Ltd, weโ€™re helping businesses build cutting-edge solutions with Small Language Models. Let us show you how to turn unstructured data into actionable insights.

Reach out today and explore how SLMs can transform your text mining workflows! ๐Ÿš€