themefinder 0.5.4

A topic modelling Python package designed for analysing one-to-many question-answer data eg free-text survey responses.

ThemeFinder

ThemeFinder is a topic modelling Python package designed for analysing one-to-many question-answer data (i.e. survey responses, public consultations, etc.). See the docs for more info.

[!IMPORTANT] Incubation project: This project is an incubation project; as such, we don't recommend using this for critical use cases yet. We are currently in a research stage, trialling the tool for case studies across the Civil Service. Find out more about our projects at https://ai.gov.uk/.

Quickstart

Install using your package manager of choice

For example pip install themefinder or poetry add themefinder.

Usage

ThemeFinder takes as input a pandas DataFrame with two columns:

• response_id: A unique identifier for each response
• response: The free text survey response

ThemeFinder is compatible with any instantiated LangChain LLM runnable, but you will need to use JSON structured output.

The function find_themes identifies common themes in response and labels them, it also outputs results from intermediate steps in the theme finding pipeline.

For this example, import the following Python packages into your virtual environment: asyncio, pandas, lanchain. And import themefinder as described above.

If you are using environment variables (eg for API keys), you can use python-dotenv to read variables from a .env file.

If you are using an Azure OpenAI endpoint, you will need the following variables:

• AZURE_OPENAI_API_KEY
• AZURE_OPENAI_ENDPOINT
• OPENAI_API_VERSION
• DEPLOYMENT_NAME
• AZURE_OPENAI_BASE_URL

Otherwise you will need whichever variables LangChain requires for your LLM of choice.

import asyncio
from dotenv import load_dotenv
import pandas as pd
from langchain_openai import AzureChatOpenAI
from themefinder import find_themes

# If needed, load LLM API settings from .env file
load_dotenv()

# Initialise your LLM of choice using langchain
llm = AzureChatOpenAI(
    model="gpt-4o",
    temperature=0,
    model_kwargs={"response_format": {"type": "json_object"}},
)

# Set up your data
responses_df = pd.DataFrame({
   "response_id": ["1", "2", "3", "4", "5"],
   "response": ["I think it's awesome, I can use it for consultation analysis.", 
   "It's great.", "It's a good approach to topic modelling.", "I'm not sure, I need to trial it more.", "I don't like it so much."]
})

# Add your question
question = "What do you think of ThemeFinder?"

# Make the system prompt specific to your use case 
system_prompt = "You are an AI evaluation tool analyzing survey responses about a Python package."

# Run the function to find themes
# We use asyncio to query LLM endpoints asynchronously, so we need to await our function
async def main():
    result = await find_themes(responses_df, llm, question, system_prompt)
    print(result)

if __name__ == "__main__":
    asyncio.run(main())

ThemeFinder pipeline

ThemeFinder's pipeline consists of five distinct stages, each utilizing a specialized LLM prompt:

Sentiment analysis

• Analyses the emotional tone and position of each response using sentiment-focused prompts
• Provides structured sentiment categorisation based on LLM analysis

Theme generation

• Uses exploratory prompts to identify initial themes from response batches
• Groups related responses for better context through guided theme extraction

Theme condensation

• Employs comparative prompts to combine similar or overlapping themes
• Reduces redundancy in identified topics through systematic theme evaluation

Theme refinement

• Leverages standardisation prompts to normalise theme descriptions
• Creates clear, consistent theme definitions through structured refinement

Theme target alignment

• Optional step to consolidate themes down to a target number

Theme mapping

• Utilizes classification prompts to map individual responses to refined themes
• Supports multiple theme assignments per response through detailed analysis

The prompts used at each stage can be found in src/themefinder/prompts/.

The file src/themefinder.core.py contains the function find_themes which runs the pipline. It also contains functions fo each individual stage.

For more detail - see the docs: https://i-dot-ai.github.io/themefinder/.

License

This project is licensed under the MIT License - see the LICENSE file for details.

The documentation is © Crown copyright and available under the terms of the Open Government 3.0 licence.

Feedback

If you have feedback on this package, please fill in our feedback form or contact us with questions or feedback at [e-mail].