Learning Syntax Without Planting Trees: Understanding Hierarchical Generalization in Transformers
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866912275560923136 |
|---|---|
| author | Ahuja, Kabir Balachandran, Vidhisha Panwar, Madhur He, Tianxing Smith, Noah A. Goyal, Navin Tsvetkov, Yulia |
| author_facet | Ahuja, Kabir Balachandran, Vidhisha Panwar, Madhur He, Tianxing Smith, Noah A. Goyal, Navin Tsvetkov, Yulia |
| contents | Transformers trained on natural language data have been shown to learn its hierarchical structure and generalize to sentences with unseen syntactic structures without explicitly encoding any structural bias. In this work, we investigate sources of inductive bias in transformer models and their training that could cause such generalization behavior to emerge. We extensively experiment with transformer models trained on multiple synthetic datasets and with different training objectives and show that while other objectives e.g. sequence-to-sequence modeling, prefix language modeling, often failed to lead to hierarchical generalization, models trained with the language modeling objective consistently learned to generalize hierarchically. We then conduct pruning experiments to study how transformers trained with the language modeling objective encode hierarchical structure. When pruned, we find joint existence of subnetworks within the model with different generalization behaviors (subnetworks corresponding to hierarchical structure and linear order). Finally, we take a Bayesian perspective to further uncover transformers' preference for hierarchical generalization: We establish a correlation between whether transformers generalize hierarchically on a dataset and whether the simplest explanation of that dataset is provided by a hierarchical grammar compared to regular grammars exhibiting linear generalization. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2404_16367 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Learning Syntax Without Planting Trees: Understanding Hierarchical Generalization in Transformers Ahuja, Kabir Balachandran, Vidhisha Panwar, Madhur He, Tianxing Smith, Noah A. Goyal, Navin Tsvetkov, Yulia Computation and Language Machine Learning Transformers trained on natural language data have been shown to learn its hierarchical structure and generalize to sentences with unseen syntactic structures without explicitly encoding any structural bias. In this work, we investigate sources of inductive bias in transformer models and their training that could cause such generalization behavior to emerge. We extensively experiment with transformer models trained on multiple synthetic datasets and with different training objectives and show that while other objectives e.g. sequence-to-sequence modeling, prefix language modeling, often failed to lead to hierarchical generalization, models trained with the language modeling objective consistently learned to generalize hierarchically. We then conduct pruning experiments to study how transformers trained with the language modeling objective encode hierarchical structure. When pruned, we find joint existence of subnetworks within the model with different generalization behaviors (subnetworks corresponding to hierarchical structure and linear order). Finally, we take a Bayesian perspective to further uncover transformers' preference for hierarchical generalization: We establish a correlation between whether transformers generalize hierarchically on a dataset and whether the simplest explanation of that dataset is provided by a hierarchical grammar compared to regular grammars exhibiting linear generalization. |
| title | Learning Syntax Without Planting Trees: Understanding Hierarchical Generalization in Transformers |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2404.16367 |