Intelligent Question Answering in WeChat Search: Knowledge‑Graph QA and Document‑Based QA Techniques

The talk begins with a background on why traditional search is insufficient for answering user queries and motivates the need for intelligent question answering (QA) in the WeChat Search scenario.

Four types of user QA needs are described: factual (entity) queries, opinion (yes/no) queries, summary (long‑answer) queries, and list/step queries.

The technical roadmap is split into two main tracks: Knowledge‑Graph QA (KBQA) and Document‑Based QA (DocQA). KBQA relies on a structured knowledge base and semantic parsing, while DocQA uses retrieval plus machine reading comprehension over unstructured text.

For KBQA, the pipeline includes entity linking, relation recognition, topic entity detection, constraint identification, and query‑graph reasoning. Entity linking uses NER, candidate recall (both dictionary‑based and vector‑based), and a BERT‑based disambiguation model that also incorporates attribute information. Relation recognition builds a template library from seed triples and expands it via synonym patterns, then encodes relations with one‑hot and token embeddings for matching.

Complex queries are handled by generating candidate query graphs, pruning them with binary classification, one‑hop pruning, and heuristic rules, and ranking candidates with either BERT semantic matching or graph‑based encoders.

DocQA follows a retrieval‑plus‑MRC workflow: filter QA intent, retrieve relevant paragraphs, apply a multi‑paragraph reader, re‑rank answers, and output the final result. Challenges such as limited labeled data, training‑inference mismatch, and answer extraction for long or yes/no answers are addressed with large‑scale pre‑training, distant supervision, soft label smoothing, and cross‑paragraph normalization.

The future outlook highlights three research directions: more robust parsing of complex multi‑intent queries, improving MRC stability through adversarial training, and extracting conditioned short‑answer entities across multiple passages.

A short Q&A section clarifies practical details about KBQA seed‑triple labeling, entity role identification, and negative‑sample selection.