This repository contains my research work on "Dialogue Schema Induction: An Unsupervised Approach." This research focuses on the automation of dialog schemas from domain-specific conversations using an unsupervised method. This advancement paves the way for the development of dialogue systems that are more comprehensive, versatile, and controllable.
- Developed an end-to-end pipeline for unsupervised dialog schema induction from domain-specific conversations.
- Leveraged BERT-based sentence transformer method for generating embeddings of utterances.
- Utilized Agglomerative Clustering from sklearn for clustering similar utterances.
- Employed GPT-3 language model for high-level dialogue action recognition.
- Optimized cluster merging using spacy lemmatizer and cosine similarity to improve coherence of the resulting dialogue schema.
- Derived dialog schemas using heuristics like transition probability.
- Successfully experimented with the MetaWoZ dataset, demonstrating the efficacy of the approach.
An example of a schema for the ordering a pizza domain:
.
- Natural Language Processing
- Unsupervised Machine Learning
- Dialogue Systems
- BERT-based sentence transformer
- GPT-3 Language Model
- Python
- Spacy
- PyTorch
- Sklearn
This research work is organized into the following sections:
- Finding embeddings for utterances
- Clustering the utterances
- Labeling the clusters
- Merging the clusters
- Using heuristics to find the schema for clusters
Please refer to the src directory for the source code, and the data directory for the dataset used in this research.
To replicate the research findings, follow these steps:
- Clone the repository
- Install the necessary dependencies (mentioned in
requirements.txt) - Run the code
python main.py
- Introduced a new approach for automatically constructing dialog schemas in an unsupervised manner from domain-specific conversations.
- Successfully categorized and labeled conversational utterances using a multi-phase framework, thereby identifying the distinct dialog actions that serve as nodes in a graph-based schema.
- Achieved promising results on the MetaWoZ dataset with the approach.
- Extending the schema induction approach to open-domain conversations.
- Exploring new strategies for schema induction.
I am a motivated individual with a knack for problem-solving and the ability to independently manage project pipelines. I possess strong skills in programming, which I have applied to my research in the field of Dialogue Systems. My goal is to continually innovate and contribute to the development of emerging technologies. I am currently open to full-time Software Engineer roles where I can apply my technical expertise to challenging and impactful real-world problems.
For any further queries or discussions related to my research work, feel free to reach out to me at stutia3@illinois.edu.