Source-linked AI summary

Simple Question Answering by Attentive Convolutional Neural Network

Wenpeng Yin, Mo Yu, Bing Xiang, Bowen Zhou, Hinrich Schütze

arXiv:1606.03391v2cs.CL

TL;DR

SimpleQA asks systems to answer single-relation Freebase questions by identifying the relevant entity and fact. The paper combines simple entity linking with character- and word-level CNN matching, including attentive maxpooling for predicate patterns. Its system outperforms prior systems and achieves optimal performance in the reported FB2M and FB5M experiments.

  • Problem

    SimpleQA remains difficult because many paraphrases express the same single-relation question, while entity linking can have candidate-collection limitations when entities share names.

  • Method

    The system ranks entity candidates with simple coverage and position factors, then uses char-CNN for entity–mention matching and attentive word-CNN for predicate–pattern matching.

  • Results

    The system outperforms the state-of-the-art system, while active-entity-linker AMPCNN has optimal performance on FB2M and FB5M.

  • Takeaways & Limitations

    Simple entity linking and attentive convolution provide an effective fact-selection approach for SimpleQA and also show outstanding performance on relation classification.

  • Takeaways & Limitations

    Surface-form entity linking is limited when distinct entities have the same names, constraining candidate collection.

Abstract

from arXiv · show

This work focuses on answering single-relation factoid questions over Freebase. Each question can acquire the answer from a single fact of form (subject, predicate, object) in Freebase. This task, simple question answering (SimpleQA), can be addressed via a two-step pipeline: entity linking and fact selection. In fact selection, we match the subject entity in a fact candidate with the entity mention in the question by a character-level convolutional neural network (char-CNN), and match the predicate in that fact with the question by a word-level CNN (word-CNN). This work makes two main contributions. (i) A simple and effective entity linker over Freebase is proposed. Our entity linker outperforms the state-of-the-art entity linker over SimpleQA task. (ii) A novel attentive maxpooling is stacked over word-CNN, so that the predicate representation can be matched with the predicate-focused question representation more effectively. Experiments show that our system sets new state-of-the-art in this task.

1 Introduction

The paper frames SimpleQA as selecting a single Freebase fact whose subject and predicate match the question. It proposes lightweight entity linking and CNN-based matching for entity mentions and predicate-focused question patterns.

  • Task motivation: SimpleQA answers single-relation questions by selecting a Freebase fact matching the question’s entity and topic.The target fact has the form (subject, predicate, object).
  • Entity linking: Entity linking reduces Freebase’s overwhelming fact inventory to a small candidate subset.The proposed linker searches entity names and ranks candidates using coverage and mention-position factors.
  • Fact selection: Fact selection matches each candidate fact’s subject with the question mention and its predicate with the question pattern.The system treats surface-form matching and predicate-pattern matching as distinct signals.
  • Neural architecture: A character-level CNN handles entity–mention surface forms, while a word-level CNN with attentive maxpooling handles predicate–pattern matching.Character-level representations are intended to tolerate typos, spaces, and other character violations; attentive pooling emphasizes predicate-like phrases.
  • Overall result: The simple architecture outperforms a more complicated state-of-the-art system on the task.The full pipeline uses top-N entity candidates, then applies char-CNN and word-CNN matching to select a fact.

2 Related Work

Prior SimpleQA research used retrieval, relation extraction, distributed representations, semantic parsing, and CNN matching. This work positions its approach as CNN-based fact selection with distinct surface-form and predicate-pattern modeling.

  • Research landscape: Factoid QA research distinguishes single-relation questions from multi-relation questions commonly addressed with semantic parsing.Multi-relation systems often map questions to formal meaning representations.
  • Limitations: Surface-form entity linking can fail during candidate collection when distinct entities share the same name.Matching entity descriptions with another word-CNN did not improve results in the reported experiment.
  • Candidate retrieval: Earlier approaches retrieve candidate answers from knowledge bases using relation extraction or distributed representations.These methods form alternative routes to candidate generation and answer selection.
  • CNN approaches: Prior CNN work matched entity mentions and predicate patterns, but used the same word-hashing-based CNN architecture for both.The present work instead separates character-level entity matching from word-level predicate-pattern matching.
  • Benchmarks: The SimpleQuestions benchmark framed SimpleQA as a single lookup in memory and was also studied with embedding-based Memory Networks.Training could combine SimpleQuestions with WebQuestions and Reverb datasets.

3 Task Definition and Data Introduction

SimpleQuestions contains single-relation questions paired with ground-truth Freebase facts, where the object is the default answer. Evaluation requires recovering both the correct subject and predicate.

  • Freebase: Freebase represents entities connected by directional predefined predicates, with each triple (h, p, t) describing a fact.For example, a route can be connected to a city through a major-cities predicate.
  • Benchmark: The SimpleQuestions benchmark provides single-relation questions, each paired with a ground-truth fact whose object is the default answer.The benchmark is split into train, dev, and test sets.
  • Data scale: SimpleQuestions includes FB2M and FB5M Freebase subsets with millions of entities, thousands of predicates, and millions of atomic facts.FB2M contains 2,150,604 entities and FB5M contains 4,904,397 entities.
  • Task formulation: The task assumes that a question can be answered by querying Freebase with a single subject and predicate.Only the tuple (h, p) is used for matching, and accuracy requires both subject and predicate to match ground truth.

4 Entity Linking

The entity-linking stage identifies candidate Freebase entities and the question span referring to the entity. It supports passive and active procedures, with LCCS-based scoring and mention construction.

  • Entity linking goals: Entity linking must identify candidate Freebase entities and the question span that refers to the entity.The order of these operations yields passive and active entity linkers.
  • Passive Entity Linker: The passive linker retrieves entities containing question words, computes word-level LCCS overlap, and retains top-N ranked candidates.Its score combines question coverage, entity-name coverage, and mention position through factors a, b, and c.
  • Passive Entity Linker: The passive linker prefers candidates covering more question or entity words and candidates whose matching span occurs near the question’s end.The authors report that it outperforms other state-of-the-art SimpleQuestions entity linkers and runs unsupervised and fast.
  • Mention detection: Mention detection can expand word-level LCCS matches, fall back to character-level LCCS, and create a pattern by replacing the mention with <e>.Passive mention detection can produce multiple mention–pattern pairs, whereas active detection produces one.
  • Active Entity Linker: The active linker detects the entity mention first, then searches for entity candidates using the detected span.It produces one mention–pattern pair per question and requires entity-labeled training questions.

5 Fact Selection

Fact selection matches each question with candidate Freebase facts using character-level entity matching and predicate-focused word-level matching. Attentive maxpooling reweights question subsequences according to their similarity to the predicate before extracting the pattern representation.

  • Candidate generation: The entity linker supplies top-N entity candidates, whose subject-linked facts form the fact-search pool.Fact selection then ranks these candidate facts against the question.
  • Two-part matching: The system uses char-CNN for entity–mention surface-form matching and word-AMPCNN for predicate–pattern matching.The two CNNs model complementary aspects of the question–fact match.
  • CNN framework: The CNN framework represents each input as embedded entries, applies sliding n-gram convolutions, and uses shared weights for the two text inputs.Entries may be characters or words, depending on the model.
  • Traditional maxpooling: Traditional maxpooling takes dimension-wise maxima across n-gram representations without using predicate context.This produces the same pooled representation for different predicate representations.
  • Attentive maxpooling: Attentive maxpooling computes predicate–pattern similarities, reweights pattern features, and retrieves maxima from the original feature map.Higher-similarity subsequences therefore receive greater influence in the final pattern representation.
  • Attentive maxpooling: AMP is designed to extract n-gram features related to the predicate, unlike TMP’s context-free feature selection.The predicate guides which question subsequences are emphasized during pooling.

6 Experiments

Experiments evaluate entity linking, SimpleQuestions fact selection, and attentive maxpooling through comparisons with ablations and established baselines. The system achieves strong results while using a comparatively simple CNN-based architecture.

  • 6.1 Training Setup: Training uses fact pools from top-N linked entities, with 99 negative facts sampled per ground-truth fact during training and all candidates retained for development and testing.The model uses Adagrad, L2 regularization, diversity regularization, and development-set hyperparameter tuning.
  • 6.2 Entity Linking: The entity linker outperforms the baseline by large margins and exceeds its reranked scores by 2–3 percent.All three ranking factors matter, with factor a dominant at small N and factor c contributing less as N increases.
  • 6.2 Entity Linking: The active entity linker performs significantly better than both the baseline linker and the passive linker.The experiments compare passive and active linkers and release the resulting entity-linking outputs.
  • 6.3 SimpleQuestions: AMPCNN with the active entity linker achieves optimal performance on both FB2M and FB5M.FB5M performance is slightly lower, mainly because entity-linking coverage is about 2% below FB2M; removing attention still yields competitive performance.
  • 6.4 Effect of Attentive Maxpooling: AMPCNN outperforms the best relation-classification baseline, APCNN, by 0.8%.AMPCNN also uses fewer parameters and runs faster than APCNN.

7 Conclusion

The paper concludes that CNNs support SimpleQA through improved entity linking and attentive predicate–question matching. Its model performs strongly on both SimpleQA and relation classification.

  • 7 Conclusion: The work contributes a simple entity linker that provides higher ground-truth entity coverage.It also contributes attentive maxpooling above the convolution layer to model predicate–question-pattern relationships more effectively.
  • 7 Conclusion: The model shows outstanding performance on both SimpleQA and relation classification.The conclusion summarizes the system-level outcome of the proposed CNN approach.
Loading 1606.03391v2…