Improving Question Answering over Incomplete KBs with Knowledge-Aware Reader

• KB coverage is limited because building accurate, complex-schema knowledge bases requires substantial human effort, leaving some open-domain questions without full evidence.Source: §1 · p. 1
• Text can cover broad, evolving knowledge, but text-based QA models remain insufficient when they must handle multiple documents.The paper hypothesizes that missing background knowledge partly hinders distinguishing relevant information from irrelevant information.Source: §1 · p. 1
• The model combines a subgraph reader that accumulates knowledge about KB entities with a text reader that selectively incorporates that knowledge through conditional gating.The gate dynamically determines how much KB knowledge to use while encoding questions and passages.Source: §1 · p. 1
• The model reports consistent improvements over the previous state of the art while using a single pass over the evidence resources.Source: §1 · p. 1

• The task reads KB tuples and retrieved Wikipedia documents to answer questions, predicting from candidate entities in both sources.Source: §2 · pp. 1–2
• The model uses a graph-attention KB reader to collect entity information, passes that knowledge to a text reader, and aggregates both evidence sources for answer prediction.Source: Figure 2 · p. 2
• For each question, the KB subgraph is retrieved with Personalized PageRank from topic entities, while documents are retrieved by an existing retriever and ranked with Lucene.Entities in the documents are annotated and linked to KB entities.Source: §2 · p. 2

3 Model

• The model combines a graph-attention KB reader with a knowledge-aware text reader to use entity knowledge from both sources.The modules are described as the model’s core components.Source: §3 · p. 2

3.1 SubGraph Reader

• SGREADER uses graph attention to aggregate linked-neighbor knowledge for each subgraph entity, matching relations to question tokens and attending to topic-entity neighbors.It outputs a vector representation for each entity; neighbor attention incorporates question-relation similarity and whether the neighbor is a question topic entity.Source: §3.1 · p. 2, Extra Attention over Topic Entity… · p. 2

Extra Attention over Topic Entity Neighbors

• A gated propagation rule combines each entity’s original representation with information from linked tuples, controlling how much of the original representation is retained.The neighbor attention includes a binary indicator for whether a neighbor is a topic entity mentioned in the question.Source: Extra Attention over Topic Entity… · pp. 2–3

3.2 Knowledge-Aware Text Reader

• KAREADER updates the question with topic-entity knowledge and uses question-conditioned gates to fuse linked-entity knowledge with passage-token features.The question update is intended to help distinguish relevant information beyond text matching.Source: §3.2 · p. 3

3.3 Answer Prediction

• For each candidate entity, answer prediction matches the query representation against its learned KB representation and aggregated document representation.The document representation for an entity is formed by averaging representations of linked documents.Source: §3.3 · p. 3, §3.2 · p. 3

4.1 Setup

• Experiments downsample WebQSP's KB to different extents and keep the retrieved document set the same as prior work; evaluation uses F1 and Hits@1.The comparisons include KB-only and KB-plus-text baselines.Source: §4.1 · pp. 3–4

4.2 Results and Analysis

• Across incomplete-KB settings, SGREADER beats prior KBQA methods and KAREADER consistently improves SGREADER; the single model has the largest absolute gain over baselines and competitive results to GN-LF+EF.SGREADER also achieves competitive performance with the full KB.Source: Table 1 · p. 4, §4.2 · p. 4
• Under the 30% KB setting, ablations identify query reformulation, knowledge enhancement, and the conditional knowledge gate as important components of KAREADER.Replacing the proposed gate with a standard gate performs worse than removing knowledge enhancement.Source: §4.2 · p. 4, Table 2 · p. 4
• In 83.2% of test samples, KAREADER improves on SGREADER, including cases where missing KB links leave answer facts absent and cases where KB facts omit answer-property constraints.SGREADER can return irrelevant answers in missing-link cases or answers that violate constraints such as entity type.Source: Table 3 · p. 5, §4.2 · p. 5
• Even when the KB suffices, KAREADER sometimes adds incorrect but question-relevant answers caused by noise in machine reading.The authors identify concatenation-based knowledge aggregation as an area for future improvement.Source: §4.2 · p. 5

• The model yields consistent improvements over previous methods on WebQSP with incomplete KBs.Source: §5 · p. 5
• Graph attention efficiently and accurately accumulates question-related knowledge for each KB entity in one pass over the subgraph.Source: §5 · p. 5
• Designed gating mechanisms incorporate encoded entity knowledge while processing text documents.Source: §5 · p. 5
• Future work will extend the approach to multimodal-evidence QA, including visual QA combining symbolic approaches (Mao et al., 2019; Hu et al., 2019).Source: §5 · p. 5