Abstract
Current state-of-the-art semantic role labeling (SRL) uses a deep neural network with no explicit linguistic features. However, prior work has shown that gold syntax trees can dramatically improve SRL decoding, suggesting the possibility of increased accuracy from explicit modeling of syntax. In this work, we present linguistically-informed self-attention (LISA): a neural network model that combines multi-head self-attention with multi-task learning across dependency parsing, part-of-speech tagging, predicate detection and SRL. Unlike previous models which require significant pre-processing to prepare linguistic features, LISA can incorporate syntax using merely raw tokens as input, encoding the sequence only once to simultaneously perform parsing, predicate detection and role labeling for all predicates. Syntax is incorporated by training one attention head to attend to syntactic parents for each token. Moreover, if a high-quality syntactic parse is already available, it can be beneficially injected at test time without re-training our SRL model. In experiments on CoNLL-2005 SRL, LISA achieves new state-of-the-art performance for a model using predicted predicates and standard word embeddings, attaining 2.5 F1 absolute higher than the previous state-of-the-art on newswire and more than 3.5 F1 on out-of-domain data, nearly 10% reduction in error. On ConLL-2012 English SRL we also show an improvement of more than 2.5 F1. LISA also out-performs the state-of-the-art with contextually-encoded (ELMo) word representations, by nearly 1.0 F1 on news and more than 2.0 F1 on out-of-domain text.
Introduction
Semantic Role Labeling (SRL) is the task of determining "who did what to whom" in a sentence. Historically, SRL systems relied heavily on complex, hand-crafted syntactic features. However, with the rise of deep learning, "syntax-free" models (like LSTMs) began to outperform these traditional systems, leading many to wonder if explicit linguistic structure was even necessary.
The authors argue that syntax remains valuable, especially for resolving tricky ambiguities like prepositional phrase attachments. The challenge is integrating this syntactic knowledge into modern neural architectures effectively and efficiently, without a brittle, multi-stage pipeline.
Methods
LISA is built on the Transformer architecture, specifically utilizing multi-head self-attention. The core innovation is Syntactically-Informed Self-Attention.
Core Components
- Supervised Attention: Instead of letting all attention heads learn freely, one specific head is forced to learn the sentence's dependency structure. This head is trained to attend to the syntactic "parent" of each token using a bi-affine scoring mechanism.
- Multi-Task Learning (MTL): The model is trained to simultaneously perform four related tasks: Part-of-Speech (POS) tagging, predicate detection, dependency parsing, and SRL.
- Efficiency: Unlike many previous models that re-encode a sentence for every predicate, LISA encodes the sequence once. It then uses a bilinear operation to predict roles for all predicates in parallel.
- Parse Injection: Because syntax is localized to one attention head, the model is modular. At test time, this head can be "hard-coded" to use a high-quality parse from an external parser (e.g., Dozat & Manning) to boost performance without retraining.
Results
The authors evaluated LISA on CoNLL-2005 (WSJ and Brown corpus) and CoNLL-2012 (OntoNotes 5.0) using both GloVe and ELMo embeddings.
Key Findings
- New State-of-the-Art: On CoNLL-2005, LISA outperformed previous best models by 2.5 F1 points on newswire and over 3.5 F1 points on out-of-domain data.
- Robustness to Domain Shift: One of the biggest wins was on the Brown out-of-domain corpus, where the model achieved a nearly 10% reduction in error. Syntax is relatively consistent across domains, allowing the model to generalize better than syntax-free counterparts.
- Synergy with ELMo: Even with powerful ELMo embeddings, adding explicit syntactic supervision provided an additional boost of ~1.0 to 2.0 F1 points.
- Syntax as an Oracle: By replacing predicted attention weights with gold-standard syntactic parents at test time, the authors proved the model learned to use that specific head as a syntactic oracle.
Conclusion
LISA demonstrates that supervising internal mechanisms like attention weights is an effective way to bake domain knowledge into neural models. While a performance gap remains between predicted and gold parses, the model's ability to efficiently incorporate syntax leads to significant gains, particularly in out-of-domain scenarios. Future work may explore improving the internal parsing head or applying supervised attention to other linguistic structures.
Summary updated