TL;DR
This paper introduces a suite of "probing tasks" to understand what linguistic information neural relation extraction (RE) models actually learn from sentences. They find that different model architectures and input features (like entity masking or contextualized word embeddings) lead to distinct learned representations, with trade-offs between RE performance and the ability to capture specific linguistic properties.
What Problem Does This Paper Address?
Neural networks have achieved impressive results in relation extraction (RE), a task focused on identifying semantic relationships between entities in text (e.g., "org:subsidiaries" between "Google" and "Alphabet"). However, despite their success, it's often unclear what linguistic knowledge these models are actually leveraging. Are they truly understanding syntax and semantics, or are they relying on superficial cues? This lack of interpretability makes it hard to diagnose errors, improve models, and build trust in their predictions. This paper aims to bridge this gap by systematically probing the learned sentence representations to understand which linguistic features are encoded and how different architectural choices and input enhancements influence this encoding.
Methods
The core of this work is the development and application of 14 "probing tasks." These tasks are essentially diagnostic classifiers trained on the internal representations of RE models. The idea is that if a model's representation encodes a specific linguistic feature (like sentence length, argument order, or entity type), a simple classifier trained on these representations should be able to predict that feature with high accuracy.
The authors then evaluate over 40 combinations of:
- Four Encoder Architectures: Convolutional Neural Networks (CNN), Bidirectional Long Short-Term Memory networks (Bi-LSTM), Graph Convolutional Networks (GCN), and Multi-Headed Self-Attention (S-Att.).
- Supporting Linguistic Knowledge:
- Entity Masking: Replacing entity mentions with their types and grammatical roles to force models to focus on context.
- Contextualized Word Representations: Incorporating embeddings from ELMo and BERT.
These models are trained on two standard RE datasets (TACRED and SemEval 2010 Task 8) and then evaluated on the probing tasks.
Datasets
- TACRED: A large dataset for relation extraction, notable for its high proportion of negative (no-relation) examples.
- SemEval 2010 Task 8: A standard benchmark for binary relation classification.
The authors also introduce REval2, a framework extending the SentEval toolkit with their new probing tasks, to facilitate future research.
Key Insights and Ideas
- Architecture Matters: Different neural architectures (CNN, Bi-LSTM, GCN, S-Att.) imprint distinct biases on the learned representations. For instance, CNNs and Bi-LSTMs seem to excel at capturing local context, while GCNs leverage dependency information more effectively.
- Linguistic Features are Explicitly Encoded (Sometimes): Models trained for RE do encode many linguistic properties relevant to the task, such as entity types and grammatical roles. Performance on probing tasks directly reflects this.
- Trade-offs Exist: There's a fascinating tension between performance on the downstream RE task and performance on probing tasks. Models that achieve state-of-the-art RE performance (like S-Att. with BERT) sometimes perform poorly on probing tasks, suggesting they might be learning "deeper" or more abstract linguistic features not captured by the current probes.
- Contextualized Embeddings are Powerful: Adding ELMo and BERT embeddings significantly boosts performance on probing tasks related to named entities and part-of-speech tags, indicating these embeddings capture rich syntactic and semantic information crucial for RE.
- Entity Masking: A Double-Edged Sword: While entity masking can improve RE performance by preventing overfitting to specific entity mentions, it can also hurt performance on entity-type-related probing tasks, suggesting models rely more on explicit entity information when it's available.
Key Results
- Entity Type is Key: All proper encoders (CNN, Bi-LSTM, GCN, S-Att.) show consistently high performance on entity type probing tasks (TypeHead, TypeTail), underscoring its importance for RE.
- Sentence Length is Ignored: Models trained on complex downstream tasks like RE tend to perform worse on simple tasks like predicting sentence length (SentLen), suggesting they "forget" superficial features as they learn deeper linguistic properties.
- BERT Outperforms: BERT embeddings generally lead to better and more balanced performance across probing tasks compared to ELMo.
- S-Att. vs. Probing: The self-attention (S-Att.) encoder, while achieving top RE performance, consistently ranked last on many probing tasks, hinting at a different learning strategy. For example, on TACRED, S-Att.
- BERT cased with masking achieved the highest RE F1 score (66.9%), but performed poorly on probing tasks.
Limitations and Future Work
- Probing Task Coverage: The authors acknowledge that their 14 probing tasks might not cover all relevant linguistic patterns. Future work could include tasks for appositions and generalization to specific entity types (e.g., company vs. person names).
- "Deeper" Features: The observation that high-performing RE models sometimes do poorly on probing tasks suggests they might be learning more abstract or "deeper" linguistic features that current probes don't capture. Investigating these further is a promising direction.
- Generalizability: While the study uses two datasets, further validation across more diverse RE datasets and tasks would be beneficial.
Summary updated