Abstract
We present open domain response generation with meta-words. A meta-word is a structured record that describes various attributes of a response, and thus allows us to explicitly model the one-to-many relationship within open domain dialogues and perform response generation in an explainable and controllable manner. To incorporate meta-words into generation, we enhance the sequence-to-sequence architecture with a goal tracking memory network that formalizes meta-word expression as a goal and manages the generation process to achieve the goal with a state memory panel and a state controller. Experimental results on two large-scale datasets indicate that our model can significantly outperform several state-of-the-art generation models in terms of response relevance, response diversity, accuracy of one-to-many modeling, accuracy of meta-word expression, and human evaluation.
Introduction
Open-domain chatbots built on standard Sequence-to-Sequence (Seq2Seq) architectures often suffer from the "safe response" problem, generating repetitive and uninformative replies like "I don't know." This stems from the one-to-many nature of dialogue: a single input can lead to numerous valid outputs. Standard models tend to favor high-frequency, generic patterns. While latent variable models have been proposed to increase diversity, they often act as "black boxes," lacking interpretability and direct control.
To address this, researchers from Microsoft and Peking University introduced meta-words—structured records of response attributes such as dialogue acts, length, and specificity. By explicitly defining these attributes, response generation becomes both explainable and controllable.
Methodology: Goal Tracking Memory Network
The core of this approach is the Goal Tracking Memory Network (GTMN), which treats meta-words as goals to be achieved during the decoding process. The GTMN consists of three primary components:
- State Memory Panel: This panel tracks the fulfillment of each attribute in the meta-word. It functions like a set of progress bars, updated at each decoding step.
- State Controller: This component manages the interaction between the memory and the decoder. It uses subtraction and addition operations to update the state based on the words generated so far.
- Difference Reading: The model calculates the "residual"—the gap between the current state and the target goal. This information is fed back into the decoder to guide the selection of the next word.
A key innovation is the State Update Loss, a supervised loss function that ensures the memory network accurately tracks the progress of attribute expression. This allows for dynamic control; for instance, if a specific sentence length is desired, the model knows exactly how many words remain to be generated at any point.
Experimental Setup and Results
The model, GTMES2S, was evaluated on two massive, real-world datasets: Twitter FireHose (4.7 million pairs) and Reddit (4.2 million pairs). It was compared against state-of-the-art baselines, including MMI-bidi and kg-CVAE.
Key Findings
- Diversity and Relevance: GTMES2S achieved significantly higher "Distinct-1/2" scores, indicating a more diverse vocabulary, while maintaining superior BLEU scores for relevance.
- Controllability: The model demonstrated high accuracy in expressing the target meta-words (e.g., correctly generating a question when the "Act" attribute was set to "yes-no question").
- Human Evaluation: In side-by-side comparisons, human judges preferred GTMES2S responses for being more informative and natural. On Twitter data, it earned an average score of 1.04 (on a 0-2 scale), outperforming the MMI-bidi baseline (0.85).
Conclusion
By incorporating meta-words and the GTMN architecture, this research provides a unified framework for handling both categorical and real-valued constraints in dialogue generation. This approach not only improves response diversity but also offers developers a level of explainability and control previously missing in neural chatbots. Future work will explore more sophisticated methods for predicting meta-words from context and extending the framework to multi-turn conversations.
Summary updated