LLM Forensic Interview Assistant

An AI-powered web application that assists police officers in conducting forensic interviews with children in child sexual abuse (CSA) investigations — suggesting safe, non-leading questions in real time using GPT-3.5.

---

Overview

Forensic interviews with child witnesses are highly sensitive. Research shows that leading, suggestive, or directive questions can contaminate a child's testimony and compromise an investigation. This tool uses large language model (LLM) prompt engineering to suggest high-quality, open-ended interview questions to officers during live interviews — and automatically filters out any potentially harmful or biased questions before surfacing them.

The system was built as a capstone project exploring prompt engineering strategies for a real-world, high-stakes NLP application.

---

How It Works

1. Question Generation

Given an interview history (e.g. Q1: Tell me what happened. A1: Something bad happened in the park.), the system generates candidate follow-up questions using one of three prompting strategies:

Strategy	Description
Zero-Shot	Generates questions without domain-specific examples, generalizing from broad training
In-Context Learning	Uses real CSA interview examples to guide the model toward appropriate questioning patterns
Chain-of-Thought (CoT)	Reasons step-by-step through the child's response before generating a question — best for complex or ambiguous situations

2. Automatic Question Filtering

Each generated question is automatically classified into one of 5 categories based on forensic interview research:

Category	Safe?	Description
Invitation focus	✅	Open-ended, free-recall questions — most desirable
Facilitator	✅	Non-suggestive encouragements to continue
Directive	⚠️	Focuses on details already mentioned by the child
Option posing	❌	Introduces new details the child hasn't mentioned
Suggestive utterance	❌	Leads the child or implies an expected response

The system generates 5 candidate questions, filters out Suggestive and Option posing responses, and returns the best recommendation plus alternatives. If no safe question is found, it flags the result as potentially biased.

3. Evaluation

A separate evaluation pipeline (prompt_engineering_evaluation.py) benchmarks each prompting strategy against a labeled dataset of 500 real forensic interview questions (100 per category), measuring how often generated questions fall into desirable vs. undesirable categories.

---

Project Structure

LLM-Forensic-Interview-Assistant/
│
├── init1.py                              # Flask app entry point
├── openai_logic.py                       # GPT-3.5 prompt logic + filtering
├── templates/
│   └── index.html                        # Main UI
├── static/
│   ├── css/style.css                     # Styling
│   └── js/index.js                       # Frontend logic + consulting history
├── evaluation/
│   └── prompt_engineering_evaluation.py  # Offline evaluation pipeline
├── .env                                  # API key — never committed
├── .gitignore
└── README.md

---

Getting Started

Prerequisites

• Python 3.8+
• An OpenAI API key (get one here)

Installation

# Clone the repo
git clone https://github.com/qiyahuang72/LLM-Forensic-Interview-Assistant.git
cd LLM-Forensic-Interview-Assistant

# Install dependencies
pip install flask openai python-dotenv pandas openpyxl

Set up your API key

Create a .env file in the root directory:

OPENAI_API_KEY=your_openai_api_key_here

Run the app

python init1.py

Open your browser and go to http://127.0.0.1:5000

---

Key Findings

• Chain-of-Thought produced the most contextually appropriate questions for complex interview situations by reasoning through the child's response before generating
• In-Context Learning performed well on straightforward scenarios matching the training examples
• Zero-Shot was the most generalizable but least domain-specific
• The automatic classifier achieved strong performance on Directive, Invitation focus, and Facilitator categories
• ~58% of Suggestive utterance questions were misclassified as Option posing — but since both are filtered out, this does not affect the quality of recommendations surfaced to the user

---

Tech Stack

• Backend: Python, Flask
• LLM: OpenAI GPT-3.5-turbo
• Frontend: HTML, CSS, JavaScript
• Prompt strategies: Zero-shot, In-context learning, Chain-of-Thought
• Evaluation: pandas, OpenAI API

---

Background & Motivation

This project is grounded in forensic interview research, particularly the NICHD (National Institute of Child Health and Human Development) protocol, which emphasizes open-ended questioning to preserve the integrity of children's statements. Studies show that suggestive or leading questions can significantly alter a child's recollection, making question quality critical in abuse investigations.

This tool is intended as a decision-support aid — not a replacement for trained forensic interviewers.

---

Limitations & Future Work

• Currently uses GPT-3.5-turbo; upgrading to GPT-4 may improve question quality
• The classifier could be fine-tuned on a larger labeled dataset for better accuracy
• Session history is stored in-browser only — a database backend would enable persistent case tracking
• Multilingual support is partially implemented (prompts include translation instructions) but not fully tested

---

Author

Qiya Huang — NYU
GitHub