Abstract
We present open domain dialogue generation with meta-words. A meta-word is a structured record that describes 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 propose a novel goal-tracking memory network that formalizes meta-word expression as a goal in response generation and manages the generation process to achieve the goal with a state memory panel and a state controller. Experimental results from both automatic evaluation and human judgment on two large-scale data sets indicate that our model can significantly outperform state-of-the-art generation models in terms of response relevance, response diversity, and accuracy of meta-word expression.
Introduction
Open-domain dialogue systems frequently suffer from the "safe response" problem, where models default to generic, uninformative answers like "I don't know" or "Me too." This issue stems from the one-to-many relationship inherent in human conversation: a single input message can elicit a wide variety of valid responses. Standard Sequence-to-Sequence (Seq2Seq) models tend to average these possibilities, gravitating toward high-frequency, low-content patterns. While latent variable models have been proposed to address diversity, they often act as "black boxes," offering little interpretability or control for developers.
This paper introduces meta-words—structured records of response attributes—to explicitly model these variations. By defining goals such as dialogue intent, length, and specificity, we can guide the generation process toward more diverse and relevant responses.
The GTMES2S Framework
The authors propose the Goal Tracking Memory Enhanced Sequence-to-Sequence (GTMES2S) framework. The core innovation is the use of meta-words $(key, type, value)$ to define the "goal" for a response. These attributes include Act (dialogue act), Len (response length), Copy (copying words from input), Utts (multiple utterances), and Spe (specificity).
Goal Tracking Memory Network (GTMN)
Unlike models that treat attributes as static flags, GTMN manages them dynamically throughout the decoding process. It consists of two primary components:
- State Memory Panel: This acts as a "progress bar," tracking the fulfillment of each meta-word goal at every step of generation. For example, it monitors how much of the target length has been consumed or whether a specific dialogue act has been expressed.
- State Controller: This module updates the memory panel using "SUB" (subtract) and "ADD" operations. As words are generated, the controller decrements the remaining "budget" for attributes like length.
Difference Reading and Supervision
At each decoding step, the model performs Difference Reading, where it examines the residual—the gap between the current state and the final goal—to influence word selection. To ensure the memory network accurately tracks these goals, the authors introduce a State Update Loss ($L_{SU}$), which provides direct supervision during training based on the ground-truth attributes of the data.
Experimental Evaluation
The model was evaluated on two massive real-world datasets: Twitter FireHose (~4.7M pairs) and Reddit (~4.2M pairs).
Key Results
- Relevance & Diversity: GTMES2S achieved the highest BLEU scores (relevance) and Distinct-1/2 scores (diversity) compared to baselines like MMI-bidi and kg-CVAE.
- Controllability: The model demonstrated superior accuracy in following meta-word instructions, such as hitting target lengths or matching dialogue acts, significantly outperforming standard conditional training methods.
- Human Judgment: In side-by-side evaluations, human annotators preferred GTMES2S responses, citing higher quality and fewer generic "bad" responses.
Insights and Conclusion
A key takeaway is that attributes like length and copying are location-sensitive. Dynamic tracking allows the model to understand what it has already accomplished to better decide what to generate next. Furthermore, the use of explicit meta-words provides a layer of explainability, allowing developers to understand why specific words were chosen based on the active goals.
While the model shows great promise, its performance in live deployment depends on the initial prediction of meta-words. Future work involves identifying optimal sets of meta-words for specific domains and optimizing the computational overhead introduced by the state controller and memory panel.
Summary updated