Open-Domain Dialogue Systems: Current State, Challenges, and Future Directions

Introduction The talk focuses on open‑domain dialogue systems, presenting cutting‑edge technologies including knowledge‑driven generation, reinforcement‑learning‑based controllable dialogue, and large‑scale pre‑trained models, followed by a discussion of future research directions.

Dialogue System Classification Dialogue systems are divided into task‑oriented systems (e.g., Baidu UNIT, customer service) that follow modular pipelines, and chat‑oriented systems that lack specific goals. Open‑domain dialogue aims for meaningful conversations without strict domain constraints, and end‑to‑end modeling is becoming mainstream for both types.

End‑to‑End Dialogue Generation

1. Opportunities – Encoder‑decoder models encode context and decode responses, trained by minimizing negative log‑likelihood on human dialogue corpora.

2. Challenges – Current models suffer from bad cases such as logical contradictions, uncontrolled background information (e.g., age), and over‑use of safe but dull replies.

3. Data Limitations – Human dialogue data contain hidden attributes (personal traits, commonsense, knowledge, intent, emotion) that are rarely annotated, making it hard for models to capture them.

Baidu NLP: Knowledge‑Controlled Dialogue Generation

Four modules are introduced: diversified generation, knowledge‑driven generation, automated evaluation & flow control, and large‑scale latent‑space models.

1. Diversity Generation – Proposes a multi‑mapping mechanism where each response is generated via a specific mapping M₁…Mₖ, eliminating randomness in the decoding process.

2. Related Work – Highlights limitations of CVAE (poor diversity), MHAM and MARM (ineffective prior‑posterior modeling).

3. Problem – Training uses a single mapping selected by the response (posterior), while inference must pick a mapping without the response, causing a training‑inference gap.

4. Solution – Introduces discrete mapping mechanisms and separates prior/posterior inference, enabling accurate mapping selection during training.

5. Model Structure – During training, context (Post) and response (Response) are encoded to vectors x and y, a Gumbel‑Softmax layer selects a mapping M, and losses include NLLLoss and a matching loss for the posterior selector.

During inference, the model samples a mapping based only on the prior selector, as the response is unavailable.

Automated Evaluation and Dialogue Flow Control

1. Self‑evolving Dialogue System (SEEDS) – Uses reinforcement learning to improve knowledge or latent‑space selection, addressing the lack of long‑term feedback in supervised learning.

2. Automated Evaluation – Builds multiple models to assess coherence, informativeness, and logic, producing a compound reward that outperforms previous metrics.

3. SEEDS Results – Demonstrates significant improvements in multi‑turn dialogue quality and reduced logical conflicts.

Large‑Scale Latent‑Space Dialogue Models (PLATO & PLATO‑2)

Recent NLP trends show that scaling pre‑trained models (e.g., BERT, GPT‑2) improves generation quality. PLATO introduces a latent variable to enhance diversity, consisting of Generation, Recognition, and Prior modules. PLATO‑2 removes the Prior module and adds a Retrieval component, training in two stages: coarse‑grained generation (no latent variable) and fine‑grained generation (with latent variable).

PLATO‑2 is released in 300M‑parameter (24‑layer) and 1.6B‑parameter (32‑layer) versions, trained on 1.2B Chinese and 0.7B English tokens. Its architecture blends bidirectional attention for context and unidirectional attention for response, differing from pure GPT‑2 or encoder‑decoder designs.

Static and dynamic evaluations show PLATO consistently surpasses other methods while using fewer parameters, and its generated responses are more fluent and informative.

Future of Open‑Domain Dialogue Systems

Despite recent progress, current systems still fail the rigorous Turing‑style test where experts probe for deep understanding. Future work should focus on richer corpora and knowledge bases, memory and few‑shot learning capabilities, and virtual environments with self‑play to provide missing background knowledge.

Corpus & Knowledge – foundation for any model Memory & Few‑Shot Learning – enable continual learning in dialogue Virtual Environments & Self‑Play – supply missing context

Thank you for attending the talk.