Tencent Xiaowei Conversational AI Platform: Architecture, Models, and Applications

On December 15, the first "Tencent Cloud+ Community Developer Conference" was held in Beijing, organized by Tencent Cloud. The conference theme was "New Trends • New Technologies • New Applications" and gathered more than 40 technical experts to discuss the latest developments in artificial intelligence, big data, IoT, mini‑programs, and operation‑development. The following is a summary of the AI & Big Data track.

The talk began with an overview of human‑machine dialogue systems, distinguishing task‑oriented and non‑task‑oriented conversations, and describing their overall frameworks and key technical points.

Tencent Xiaowei is presented as an open, easy‑to‑integrate conversational platform rather than a single model. It enables partners to connect smart devices (speakers, cars, TVs) and issue commands such as weather queries or music playback. User utterances are sent to Tencent Cloud for speech recognition and semantic analysis, after which the inferred intent and relevant content are returned to the device.

Key capabilities of Xiaowei include:

Rich content ecosystem (QQ Music, Tencent Video, WeChat, Mini‑Programs) that can be leveraged in dialogues.

A three‑stage evolution of dialogue systems: rule‑based (1960s), data‑driven (early 2000s), and modern context‑aware systems applied to smart hardware and customer service.

Core components: Natural Language Understanding, Dialogue Management, and Dialogue Generation, built on user profiles, multi‑turn conversation technology, and extensive content resources.

Precise intent detection (single and multi‑intent) and keyword extraction, supporting hundreds of intent categories such as weather, reminders, stock queries, music playback, and smart‑device control.

The intent classification models discussed include:

Bidirectional RNNs that encode the whole sentence and use the final hidden state for intent prediction.

CNNs with kernels of different sizes to capture local n‑gram features.

Hybrid approaches derived from the above architectures.

Large‑scale annotated data and careful data cleaning are emphasized as essential for achieving high model performance. Parameter tuning further improves results.

Beyond intent detection, the talk covered multi‑turn context handling. A dialogue history is treated as memory slots and processed by Memory Networks (including a gating mechanism to balance raw input and contextual information). This approach was evaluated on the DSTC7 dataset, which contains multi‑turn teacher‑student conversations with rich user profiles, external knowledge, and up to 100 candidate answers per turn.

Experimental results showed that Memory Networks and Hierarchical LSTMs effectively leverage context and knowledge, outperforming baseline models. An enhanced Memory Network adds a learnable gate to weight the contribution of the original input versus the contextual memory dynamically.

Case studies demonstrated Xiaowei’s deployment in smart speakers, TVs, and customer‑service bots, illustrating how the platform can be used to build customized robots and reduce data‑labeling costs through active‑learning tools.

The future outlook highlighted continued integration of hardware interrupts and software terminals, more accurate dialogue understanding, and deeper exploitation of Tencent’s content resources.