Another Diversity-Promoting Objective Function for Neural Dialogue Generation
This addresses the issue of repetitive responses in dialogue systems for users, though it is incremental as it modifies the loss function without changing the model architecture.
The paper tackled the low diversity problem in neural dialogue generation by proposing a new objective function, Inverse Token Frequency (ITF) loss, which achieved a state-of-the-art DIST-1 score of 7.56 on the OpenSubtitles dataset while maintaining good BLEU-1 scores.
Although generation-based dialogue systems have been widely researched, the response generations by most existing systems have very low diversities. The most likely reason for this problem is Maximum Likelihood Estimation (MLE) with Softmax Cross-Entropy (SCE) loss. MLE trains models to generate the most frequent responses from enormous generation candidates, although in actual dialogues there are various responses based on the context. In this paper, we propose a new objective function called Inverse Token Frequency (ITF) loss, which individually scales smaller loss for frequent token classes and larger loss for rare token classes. This function encourages the model to generate rare tokens rather than frequent tokens. It does not complicate the model and its training is stable because we only replace the objective function. On the OpenSubtitles dialogue dataset, our loss model establishes a state-of-the-art DIST-1 of 7.56, which is the unigram diversity score, while maintaining a good BLEU-1 score. On a Japanese Twitter replies dataset, our loss model achieves a DIST-1 score comparable to the ground truth.