A Diversity-Promoting Objective Function for Neural Conversation Models
TL
TL;DR
This work proposes using Maximum Mutual Information (MMI) as the objective function in neural models, and demonstrates that the proposed MMI models produce more diverse, interesting, and appropriate responses, yielding substantive gains in BLEU scores on two conversational datasets and in human evaluations.
Abstract
Sequence-to-sequence neural network models for generation of conversational responses tend to generate safe, commonplace responses (e.g., "I don't know") regardless of the input. We suggest that the traditional objective function, i.e., the likelihood of output (response) given input (message) is unsuited to response generation tasks. Instead we propose using Maximum Mutual Information (MMI) as the objective function in neural models. Experimental results demonstrate that the proposed MMI models produce more diverse, interesting, and appropriate responses, yielding substantive gains in BLEU scores on two conversational datasets and in human evaluations.
References50 items
1
Attention with Intention for a Neural Network Conversation Model
2
Semantically Conditioned LSTM-based Natural Language Generation for Spoken Dialogue Systems
3
Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models
4
Hierarchical Neural Network Generative Models for Movie Dialogues
5