Source-linked AI summary
A Dependency-Based Neural Network for Relation Classification
Yang Liu, Furu Wei, Sujian Li, Heng Ji, Ming Zhou, Houfeng Wang
TL;DR
Relation classification needs to exploit complementary information from dependency shortest paths and subtrees. The paper introduces augmented dependency paths and DepNN, which combines recursive and convolutional neural networks to model them. Experiments on SemEval-2010 report state-of-the-art performance.
Problem
Prior work established the usefulness of dependency shortest paths or subtrees, but the paper asks how to fully combine these two kinds of dependency information.
Method
The paper combines each entity pair’s shortest dependency path with its attached subtrees as an augmented dependency path, modeled by DepNN using RNNs and CNNs.
Results
DepNN achieves state-of-the-art performance on the SemEval-2010 relation classification task.
Takeaways & Limitations
Combining shortest-path and subtree information in DepNN provides a unified semantic representation for relation classification.
Abstract
from arXiv · showhide
Previous research on relation classification has verified the effectiveness of using dependency shortest paths or subtrees. In this paper, we further explore how to make full use of the combination of these dependency information. We first propose a new structure, termed augmented dependency path (ADP), which is composed of the shortest dependency path between two entities and the subtrees attached to the shortest path. To exploit the semantic representation behind the ADP structure, we develop dependency-based neural networks (DepNN): a recursive neural network designed to model the subtrees, and a convolutional neural network to capture the most important features on the shortest path. Experiments on the SemEval-2010 dataset show that our proposed method achieves state-of-art results.
1 Introduction
Relation classification benefits from dependency information, but shortest paths and attached subtrees capture complementary signals. The paper combines them in an augmented dependency path and models it with DepNN, achieving state-of-the-art performance on SemEval-2010.
- Motivation: Dependency shortest paths and subtrees provide complementary information for relation classification, and their collaboration can improve performance.Prior work identifies shortest paths as information-rich and subtrees as useful supplemental structure.
- Architecture: Each shortest-path word is combined with a representation generated from its attached subtree to strengthen the path representation.The resulting augmented dependency path is represented as a continuous semantic vector for relation classification.
- Contribution: The augmented dependency path combines the shortest dependency path between two entities with the subtrees attached to that path.This structure extends the shortest path to incorporate additional dependency-tree information.
- Contribution: DepNN uses a CNN for shortest dependency paths and an RNN for dependency subtrees.The architecture combines convolutional modeling of flat paths with recursive modeling of hierarchical subtrees.
- Results: DepNN outperforms baseline methods and achieves state-of-the-art F1 measure on the SemEval-2010 relation classification task.The result is reported for experiments on the SemEval-2010 dataset.
2 Problem Definition and Motivation
The task is to identify a predefined semantic relation between two annotated entities. The paper motivates augmenting shortest dependency paths with attached subtrees because similar paths can correspond to different relation types.
- Problem Definition: Relation classification identifies the semantic relation between two annotated entities according to predefined relation types.SemEval-2010 task 8 defines nine ordered relation types, with other instances labeled Other.
- Motivation: The shortest dependency path captures predicate-argument sequences and provides useful information for judging relations between entities.For example, a path containing “broke prep-with screwdriver” helps identify Instrument-Agency.
- Motivation: Similar shortest dependency paths can belong to different relation types, making path-only classification difficult.The paper uses examples whose attached subtrees differ despite similar shortest paths.
- Motivation: Attached subtrees such as “dobj→commandment” and “dobj→ignition” provide supplemental information for distinguishing relations.The paper proposes combining these subtrees with the shortest path in an augmented dependency path.
- Approach: DepNN represents each shortest-path word using both its own meaning and the semantic information from its attached subtree.An RNN builds subtree embeddings, after which a CNN models the resulting flat path structure.
3 Dependency-Based Neural Networks
DepNN represents augmented dependency paths by recursively encoding attached subtrees and applying a CNN to the shortest dependency path, then combines this representation with lexical features for classification.
- Modeling Dependency Subtree: DepNN uses an RNN to recursively build subtree representations for words on the shortest dependency path.Each word representation concatenates its word embedding with an embedding derived bottom-up from its subtree.
- Modeling Shortest Dependency Path: The augmented dependency path is modeled as a sequence of word and dependency-relation representations.The sequence intersperses words with dependency relations, preserving both path and subtree information.
- Modeling Shortest Dependency Path: A CNN applies filters over windows of the shortest dependency path and uses max-over-time pooling to retain salient features.Each window concatenates neighboring word or relation representations before convolution and pooling.
- Dependency-Based Neural Networks: The resulting ADP representation is a continuous semantic vector that combines the advantages of recursive and convolutional networks.The RNN models dependency subtrees, while the CNN captures important features in the path sequence.
- Learning: Lexical features are concatenated with the ADP representation and passed through a softmax layer to predict relation-label probabilities.Training uses cross-entropy error with stochastic gradient descent and backpropagation.
4 Experiments
Experiments on SemEval-2010 evaluate DepNN’s components, semantic path representations, subtree effects, and baseline performance. The results support CNN modeling of shortest paths, added subtree information, and the overall DepNN design.
- Experimental Setup: The experiments use SemEval-2010, whose official metric is macro-averaged F1 excluding the Other relation.The dataset contains 8,000 training instances and 2,717 test instances, with collapsed dependency trees generated by the Stanford Parser.
- Contributions of Different Components: The evaluation builds from a CNN-only shortest-path system, adds WordNet or NER features, and then adds RNN-modeled attached subtrees for the complete ADP.These configurations isolate contributions from shortest paths, lexical features, and subtree information.
- Contributions of Different Components: Attached subtrees provide an obvious performance boost and outperform NER tags and WordNet hypernyms as supplemental information for shortest-path word representations.The authors note that WordNet may be less useful because many hypernyms can cause overfitting, while larger Gigaword embeddings also improve classification performance.
- Intuitive Analysis of Shortest Path: The learned CNN path representation captures relation-relevant dependency and lexical patterns across similar instances.Examples include “nsubj with” for Instrument-Agency, “nsubj ... verb1 xcomp verb2 dobj” with words such as “pencil” and “create” for Product-Producer, and semantically related words for Message-Topic.
- Influence of Attached Subtree: Subtree information generally improves relation types and is especially salient for Instrument-Agency and Product-Producer, which shortest paths alone can confuse.Attached structures such as “conj-and→valves,” “amod→manufacturing,” and the subtree attached to “scaled” provide supplemental information for correct classifications.
- Comparison with Baselines: 83.6 F1 is achieved by DepNN with NER features, the best result in the baseline comparison.SVM is comparable, while MV-RNN is about 1.7 percent lower without additional features and 4.6 percent lower with them; DT-RNN performs worst among the listed baselines.
5 Related Work
Relation classification research has used linguistic features, kernels, and neural networks to model relations between entities. These approaches differ in how they represent syntactic structure and learn features.
- Task formulation: Relation classification is commonly treated as supervised multiclass classification over predefined entity or nominal relation types.The field includes feature-based, tree-kernel-based, and composite-kernel approaches.
- Feature-based methods: Feature-based methods combine syntactic and semantic cues into vectors for Max Entropy or SVM classifiers.They usually require handcrafted features and poorly represent structures such as parse trees and word order.
- Kernel methods: Kernel methods capture structural similarity through inner products in high-dimensional latent feature spaces.Dependency-tree kernels extend tree-based structural modeling with POS and word-chunk information.
- Neural methods: Deep neural networks learn distributed word representations and automatically derive features, helping address sparsity in traditional methods.Prior neural approaches include recursive models over constituent trees and convolutional models over sentences.
6 Conclusion
The paper introduces augmented dependency paths and DepNN to combine dependency-path and subtree information in neural relation classification. Experiments show that DepNN achieves state-of-the-art performance.
- Contribution: The augmented dependency path combines the shortest path between two entities with its attached subtrees.This extends shortest-path modeling to include additional dependency information.
- Contribution: DepNN models the augmented dependency path by combining convolutional and recursive neural networks.The approach uses both neural architectures to represent the combined dependency structure.
- Results: DepNN achieves state-of-the-art performance in the reported experiments.The conclusion attributes this result to experiments evaluating the proposed approach.