This repository provides the official code for the ACL 2020 paper "You Impress Me: Dialogue Generation via Mutual Persona Perception." It addresses open-domain personalized dialogue generation by modeling mutual persona perception between interlocutors, aiming to improve conversation flow and chatbot personality. The target audience includes NLP researchers and developers working on advanced dialogue systems.

How It Works

The framework employs a transmitter-receiver architecture to explicitly model "Mutual Persona Perception." This approach is motivated by how humans build understanding by discussing persona-related topics or revealing personal information. The system aims to generate more engaging and coherent conversations by capturing this dynamic.

Quick Start & Requirements

• Install: Clone custom pytorch_pretrained_bert and ParlAI repos, then install them.
• Prerequisites: Python 3.7+, PyTorch 1.0.1, tensorboardX. Requires custom versions of pytorch_pretrained_bert (v0.6.2) and ParlAI.
• Data: Downloads PersonaChat dataset automatically during transmitter training.
• Resources: Training PSquare bot is noted to benefit from two GPUs.
• Docs: Paper details are available at ACL 2020.

Highlighted Details

• Achieves PPL of 15.12 and Hits@1 of 81.9 on PersonaChat validation set.
• Implements a transmitter-receiver framework with explicit persona modeling.
• Includes training scripts for Transmitter (GPT-based), Receiver (BERT-based), and PSquare (self-play).
• Offers pre-trained PSquare model weights for reproduction.

Maintenance & Community

The project is associated with the ACL 2020 paper. Contact information for the lead author is provided via email. Issues can be opened on the repository.

Licensing & Compatibility

The code's license is not explicitly stated in the README. However, it notes that usage of the modified ParlAI module follows its open-source license. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The code relies on specific, older versions of PyTorch (1.0.1) and Huggingface's transformers (v0.6.2), which may pose compatibility challenges with current libraries. The custom modifications to Huggingface's code for auxiliary tasks are also noted.