Source-linked AI summary

Semantic Relation Classification via Convolutional Neural Networks with Simple Negative Sampling

Kun Xu, Yansong Feng, Songfang Huang, Dongyan Zhao

arXiv:1506.07650v1cs.CLcs.LG

TL;DR

Relation extraction models can be distracted by irrelevant clauses when nominals are distant, motivating representations focused on their shortest dependency path. The paper uses a CNN over these paths and negative sampling for subject–object directionality, achieving better reported performance than prior systems on SemEval-2010 Task 8.

  • Problem

    Neural relation-extraction models often suffer from irrelevant information when subjects and objects are separated by longer distances, while subject–object directionality also requires explicit handling.

  • Method

    A convolutional neural network learns relation representations from shortest dependency paths, while reversed object-to-subject paths provide negative samples for learning subject–object assignments.

  • Results

    The model outperforms state-of-the-art systems on SemEval-2010 Task 8, while negative sampling improves F1 by 2.1% without lexical features and 1.9% with them.

  • Takeaways & Limitations

    Shortest dependency paths capture syntactic features for relation extraction, and simple negative samples help the model make correct subject–object assignments.

  • Takeaways & Limitations

    The paper omits detailed objective-function formulas for space and assumes the shortest path describes the relationship between two nominals in a sentence.

Abstract

from arXiv · show

Syntactic features play an essential role in identifying relationship in a sentence. Previous neural network models often suffer from irrelevant information introduced when subjects and objects are in a long distance. In this paper, we propose to learn more robust relation representations from the shortest dependency path through a convolution neural network. We further propose a straightforward negative sampling strategy to improve the assignment of subjects and objects. Experimental results show that our method outperforms the state-of-the-art methods on the SemEval-2010 Task 8 dataset.

1 Introduction

The paper targets relation extraction with syntactic representations that avoid irrelevant clauses and explicitly model subject–object directionality. It uses shortest dependency paths and negative sampling to improve relation classification.

  • Relation extraction identifies the relationship between two nominals in a sentence, commonly using syntactic and other linguistic features.
  • Neural models can incorporate irrelevant subsequences or clauses when subjects and objects are far apart, hurting extraction performance.
  • The proposed CNN learns relation representations from the simple dependency path between subjects and objects, excluding unrelated chunks or clauses.
  • Negative sampling models relation directionality by using the shortest dependency path from the object to the subject as a negative sample.
  • The model outperforms state-of-the-art methods on the SemEval-2010 Task 8 dataset.

2 The Shortest Path Hypothesis

The shortest dependency path is assumed to describe the relationship between two nominals, while predicate order on that path indicates subject–object assignments.

  • For two nominals in one sentence, the shortest dependency path is assumed to describe their relationship.
  • If the nominals share a predicate, the path passes through it; across predicate–argument structures, it passes through predicates sharing consecutive arguments.
  • The order of predicates on the path indicates the proper assignments of subjects and objects.

3 A Convolutional Neural Network Model

The CNN converts shortest dependency paths into fixed-size syntactic representations by embedding path nodes, extracting local features, pooling globally, and classifying relations.

  • The model inputs words, dependency edge directions, and dependency labels from the shortest path between subject and object.
  • Each path node is mapped to a vector through an embedding lookup table.
  • Convolution: Fixed-size windows concatenate node vectors to form local feature matrices around each dependency-path node.
  • Convolution: A linear transformation extracts local contextual features, and max pooling selects useful features into a global vector independent of path length.
  • Objective Function and Learning: The softmax classifier predicts a K-class relation distribution using cross-entropy error and regularization over the training data.
  • Objective Function and Learning: Model parameters are learned through backpropagation using stochastic gradient descent with AdaGrad.

4 Negative Sampling

The model uses negative sampling to learn correct subject–object assignments, treating reversed assignments as informative negative examples. Development experiments show that assignment information materially improves performance.

  • Pilot Experiments: The third pilot experiment serves as an upper bound because subject–object assignments are available during both training and testing.
  • Negative Sampling: A dependency path may correspond to multiple relations between two nominals.
  • Negative Sampling: The paper omits detailed formulas because of space limitations.
  • Negative Sampling: Adding assignment information during training significantly improves the model, demonstrating that subject–object assignment is crucial to performance.
  • Negative Sampling: The proposed strategy treats the opposite subject–object assignment as a negative example for the same relationship.The reversed dependency path differs from the correct path, providing information for distinguishing subject and object roles.

5 Experimental Evaluation

The model is evaluated on SemEval-2010 Task 8 against established systems and negative-sampling alternatives. It outperforms prior approaches, benefits from dependency labels, and gains further from reversed-path negative sampling and lexical features.

  • Dataset and Setup: 10,717 annotated examples comprise SemEval-2010 Task 8, with 8,000 training instances, 2,717 test instances, and 2,182 validation samples.
  • Dataset and Setup: The model predicts both opposite subject–object dependency paths and selects other only when both predictions are other.
  • Main Results: Table 2 compares the proposed models with SVM, RNN, MV-RNN, and CNN on SemEval-2010 Task 8.
  • Negative Sampling: Table 3 compares different negative-sampling methods on the development set.
  • Main Results: The vanilla depLCNN+NS outperforms prior best systems without extra lexical features, while similar lexical features improve it by 1.6%.
  • Main Results: depCNN and depLCNN outperform MVRNN and CNN by at least 2.2%, and dependency labels make depLCNN stronger than depCNN.The results indicate that dependency-label information provides more discriminative information for relation extraction.
  • Negative Sampling: Negative sampling improves F1 by 2.1% without lexical features and by 1.9% with extra lexical features.
  • Negative Sampling: Reversed dependency paths achieve 85.4% F1, improving random negative samples by 1.9% with a similar number of examples.The reversed paths provide clues for learning correct subject–object assignments.

6 Conclusion

The paper combines shortest-dependency-path convolutional representations with simple negative sampling for relation extraction. Experiments show significant gains over state-of-the-art systems and effective capture of syntactic features.

  • Conclusion: The model learns relation representations from shortest dependency paths using a convolutional neural network.
  • Conclusion: Simple negative sampling helps the model make correct subject–object assignments within relationships.
  • Conclusion: Experimental results show significant improvements over state-of-the-art systems and effective capture of syntactic features.
Loading 1506.07650v1…