Abstract
Existing open-domain dialogue generation models are usually trained to mimic the gold response in the training set using cross-entropy loss on the vocabulary. However, a good response does not need to resemble the gold response, since there are multiple possible responses to a given prompt. In this work, we hypothesize that the current models are unable to integrate information from multiple semantically similar valid responses of a prompt, resulting in the generation of generic and uninformative responses. To address this issue, we propose an alternative to the end-to-end classification on vocabulary. We learn the pair relationship between the prompts and responses as a regression task on a latent space instead. In our novel dialog generation model, the representations of semantically related sentences are close to each other on the latent space. Human evaluation showed that learning the task on a continuous space can generate responses that are both relevant and informative.
Introduction: The Problem of Generic Responses
Open-domain dialogue generation is a "high entropy" task. For any given prompt, thousands of valid, specific responses exist. Standard models typically rely on Maximum Likelihood Estimation (MLE) with cross-entropy loss, which forces the model to predict the exact tokens of a single "gold" response.
The fundamental issue with MLE is that it treats every word in the vocabulary independently. Specific, interesting words (e.g., "guitar" or "pizza") appear far less frequently than generic "stop words" (e.g., "I," "don't," "know"). Consequently, the model spreads its probability mass too thin across specific options, making the safest statistical bet a bland, high-frequency generic response. This paper argues that current models fail because they cannot effectively integrate the shared meaning of multiple valid, specific responses.
Methodology: Semantic Latent Space Regression
The authors propose a novel architecture that shifts from word-by-word classification to a regression task within a shared semantic latent space.
Mapping to a Shared Latent Space
The model utilizes a modified version of Deep Canonical Correlation Analysis (DCCA) to create a shared space where the representation of a prompt ($X$) and its corresponding response ($Y$) are highly correlated. In this space, semantically similar sentences are clustered together.
Two-Part Response Encoding
The response representation is split into two distinct components:
- Correlated ($Y$): Information directly related to the prompt.
- Uncorrelated ($Y_u$): Information specific to the response but not necessarily triggered by the prompt (e.g., a sudden topic change). This is modeled using a Variational Autoencoder (VAE) approach to ensure a continuous and smooth latent space.
Generation and Denoising
During inference, the model encodes a prompt into the latent space and uses a decoder ($G_y$) to "regress" that vector back into a full sentence. To ensure grammatical output from a continuous vector, a "denoising" technique is used during training. Additionally, an "attention bottleneck" prevents the model from bypassing the latent space and reverting to standard MLE-style behavior.
Experimental Results
The model was evaluated on two benchmarks: PersonaChat (persona-based chit-chat) and DailyDialog (high-quality human conversations).
Quantitative and Human Evaluation
The researchers compared their model against baselines including standard Seq2seq (MLE), MMI (Maximum Mutual Information), and SpaceFusion.
Human evaluation revealed that the proposed model significantly outperformed baselines in both Relevance and Informativeness. On the "Useful Information" (UI) score—a product of relevance and informativeness—the model scored 1.97 on PersonaChat, compared to 1.43 for the standard Seq2seq baseline. Interestingly, despite not being trained for word-for-word mimicry, the model achieved higher BLEU scores, suggesting that capturing semantic meaning improves lexical prediction.
Discussion and Limitations
Key Insights
- Probability Dilution: Traditional beam search fails because it calculates candidate probabilities independently. While many specific responses might collectively have high probability, that mass is "diluted," leaving generic responses as the statistical winners.
- Regression vs. Classification: Treating dialogue as finding a point in semantic space allows the model to "average" information from many similar specific responses.
- Separating Correlation: Isolating the "uncorrelated" part of a response allows for a cleaner mapping between prompts and relevant answers.
Limitations
Generating text from a continuous latent space is inherently more difficult than standard word selection. The authors noted a slightly higher rate of minor grammatical errors. Furthermore, the multi-encoder setup with DCCA and VAE components is more complex to tune than standard architectures. The attention bottleneck, while necessary, may also limit performance on very long or complex prompts.
Future work may involve integrating this latent space approach with larger pre-trained models to combine semantic clustering with superior linguistic fluency.
Summary updated