XiaoAi Intelligent QA: Information Extraction, Event Extraction, and Knowledge Graph Question Answering

1. XiaoAi Application Scenarios XiaoAi provides quick access to Xiaomi’s smart hardware (robots, TV, air‑conditioner) and offers six service categories: content, information query, interaction, control, life services, and basic tools. The focus is on intelligent QA for factual queries.

2. Information Extraction Information extraction transforms unstructured text into structured knowledge (entities, relations, events). Relation extraction uses a predefined schema to extract SPO triples; challenges include complex objects and overlapping SPOs. The presented solution (runner‑up in the 2021 Language & Intelligence Competition) includes two models:

2.1 SPO Model A pointer‑network first extracts the subject, then uses the schema to find predicates and corresponding object spans, handling complex objects by separate slot processing and position‑distance matching.

2.2 PSO Model Similar to SPO but first classifies the relation type, then extracts subject and object via pointer networks, also handling complex objects with slot‑wise processing.

2.3 Fine‑grained Scoring Model Candidate SPO triples are concatenated with the original sentence and modeled as a semantic similarity task to assign confidence scores and filter high‑confidence triples.

3. Event Extraction Event extraction identifies event types, triggers, and arguments. Three methodological streams are discussed:

• Pipeline approach: predict event type first, then extract triggers and arguments based on the predicted type.

• Joint model: simultaneously learn event type and argument extraction via multi‑task learning.

• End‑to‑end approach: combine event type and role labels into a single tag set for BIO tagging, extracting both in one step.

Given the low overlap rate, a sequence‑labeling model is chosen for its ease of training and performance. A joint model predicts event type and trigger together, using a shared encoder, a linear layer for type prediction, and a CRF layer for trigger detection.

Two parallel extraction streams are built: one conditioned on the predicted event type, the other on the detected trigger. Their results are combined by voting to produce the final event arguments.

4. Knowledge Graph Question Answering (KGQA) Three KGQA techniques are described:

4.1 Grammar‑based parsing: match user queries to templates, convert them to structured queries, and retrieve answers from the graph. Template acquisition can be driven by high‑frequency queries or automated mining.

4.2 Cross‑domain coarse‑grained slot and intent extraction: unify slots across domains using a concept graph (e.g., person, work, title) and train a joint intent‑slot model to improve generalization to low‑resource domains.

4.3 Path‑retrieval method (champion solution in CCKS2021): identify entities and attribute values, expand candidate paths by adding triples, attach constraints (filtering and ordering) as additional triples, and rank paths via semantic matching. The final path is used to query the graph for the answer.

The path retrieval system uses integer ID mapping for entities/relations and a node index table to enable fast binary‑search lookups, reducing latency and memory usage.

5. Q&A Session Participants asked about neural‑network‑based path generation versus manual templates, the role of position‑distance features in SPO models, and whether all possible event roles are predicted before filtering by the event type. The answers highlighted the current reliance on manually defined templates for performance, the use of position‑distance as a selection feature, and the extraction of all roles followed by type‑based filtering.

Thank you for attending.