# Inducing Character-level Structure in Subword-based Language Models with Type-level Interchange Intervention Training

arXiv: https://arxiv.org/abs/2212.09897 · HTML: https://mlsanity.com/paper/2212.09897/

## Key facts

- Cited by 16 papers in the map (data through 14 Oct 2026).
- Builds on 40 of its 61 references that are in the map.
- First posted on arXiv 19 Dec 2022.

## Builds on (most cited first, 30 of 40)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://mlsanity.com/paper/1910.10683/) (2019)
- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [OPT: Open Pre-trained Transformer Language Models](https://mlsanity.com/paper/2205.01068/) (2022)
- [Neural Machine Translation of Rare Words with Subword Units](https://mlsanity.com/paper/1508.07909/) (2015)
- [Deep contextualized word representations](https://mlsanity.com/paper/1802.05365/) (2018)
- [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://mlsanity.com/paper/2006.03654/) (2020)
- [BLOOM: A 176B-Parameter Open-Access Multilingual Language Model](https://mlsanity.com/paper/2211.05100/) (2022)
- [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://mlsanity.com/paper/2204.06745/) (2022)
- [Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates](https://mlsanity.com/paper/1804.10959/) (2018)
- [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://mlsanity.com/paper/2105.13626/) (2021)
- [Causal Abstractions of Neural Networks](https://mlsanity.com/paper/2106.02997/) (2021)
- [End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF](https://mlsanity.com/paper/1603.01354/) (2016)
- [Synthetic and Natural Noise Both Break Neural Machine Translation](https://mlsanity.com/paper/1711.02173/) (2017)
- [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://mlsanity.com/paper/2103.06874/) (2021)
- [Byte Pair Encoding is Suboptimal for Language Model Pretraining](https://mlsanity.com/paper/2004.03720/) (2020)
- [Charformer: Fast Character Transformers via Gradient-based Subword Tokenization](https://mlsanity.com/paper/2106.12672/) (2021)
- [Injecting Numerical Reasoning Skills into Language Models](https://mlsanity.com/paper/2004.04487/) (2020)
- [Between words and characters: A Brief History of Open-Vocabulary Modeling and Tokenization in NLP](https://mlsanity.com/paper/2112.10508/) (2021)
- [Inducing Causal Structure for Interpretable Neural Networks](https://mlsanity.com/paper/2112.00826/) (2021)
- [Abstracting Causal Models](https://mlsanity.com/paper/1812.03789/) (2018)
- [CharacterBERT: Reconciling ELMo and BERT for Word-Level Open-Vocabulary Representations From Characters](https://mlsanity.com/paper/2010.10392/) (2020)
- [Approximate Causal Abstraction](https://mlsanity.com/paper/1906.11583/) (2019)
- [What do tokens know about their characters and how do they know it?](https://mlsanity.com/paper/2206.02608/) (2022)
- [Mimicking Word Embeddings using Subword RNNs](https://mlsanity.com/paper/1707.06961/) (2017)
- [Causal Proxy Models for Concept-Based Model Explanations](https://mlsanity.com/paper/2209.14279/) (2022)
- [Models In a Spelling Bee: Language Models Implicitly Learn the Character Composition of Tokens](https://mlsanity.com/paper/2108.11193/) (2021)
- [Why don't people use character-level machine translation?](https://mlsanity.com/paper/2110.08191/) (2021)
- [AmbiPun: Generating Humorous Puns with Ambiguous Context](https://mlsanity.com/paper/2205.01825/) (2022)

## Built on by (most cited first, 16 of 16)

- [Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations](https://mlsanity.com/paper/2303.02536/) (2023)
- [ReFT: Representation Finetuning for Language Models](https://mlsanity.com/paper/2404.03592/) (2024)
- [Interpretability at Scale: Identifying Causal Mechanisms in Alpaca](https://mlsanity.com/paper/2305.08809/) (2023)
- [Rigorously Assessing Natural Language Explanations of Neurons](https://mlsanity.com/paper/2309.10312/) (2023)
- [CUTE: Measuring LLMs' Understanding of Their Tokens](https://mlsanity.com/paper/2409.15452/) (2024)
- [MrT5: Dynamic Token Merging for Efficient Byte-level Language Models](https://mlsanity.com/paper/2410.20771/) (2024)
- [ScoNe: Benchmarking Negation Reasoning in Language Models With Fine-Tuning and In-Context Learning](https://mlsanity.com/paper/2305.19426/) (2023)
- [Causal interventions expose implicit situation models for commonsense language understanding](https://mlsanity.com/paper/2306.03882/) (2023)
- [Knowledge of Pretrained Language Models on Surface Information of Tokens](https://mlsanity.com/paper/2402.09808/) (2024)
- [EXECUTE: A Multilingual Benchmark for LLM Token Understanding](https://mlsanity.com/paper/2505.17784/) (2025)
- [Fast Byte Latent Transformer](https://mlsanity.com/paper/2605.08044/) (2026)
- [Learning Mutually Informed Representations for Characters and Subwords](https://mlsanity.com/paper/2311.07853/) (2023)
- [Multi-teacher Distillation for Multilingual Spelling Correction](https://mlsanity.com/paper/2311.11518/) (2023)
- [Retrieval Augmented Spelling Correction for E-Commerce Applications](https://mlsanity.com/paper/2410.11655/) (2024)
- [Inducing Causal Structure for Interpretable Neural Networks Applied to Glucose Prediction for T1DM Patients](https://mlsanity.com/paper/2503.14442/) (2025)
- [PACUTE: Phonology-, Affix-, and Character-level Understanding of Tokens for Filipino](https://mlsanity.com/paper/2606.15144/) (2026)

## Abstract (arXiv)

Language tasks involving character-level manipulations (e.g., spelling corrections, arithmetic operations, word games) are challenging for models operating on subword units. To address this, we develop a causal intervention framework to learn robust and interpretable character representations inside subword-based language models. Our method treats each character as a typed variable in a causal model and learns such causal structures by adapting the interchange intervention training method of Geiger et al. (2021). We additionally introduce a suite of character-level tasks that systematically vary in their dependence on meaning and sequence-level context. While character-level models still perform best on purely form-based tasks like string reversal, our method outperforms character-level models on more complex tasks that blend form, meaning, and context, such as spelling correction in context and word search games. Compared with standard subword-based models, our approach also significantly improves robustness on unseen token sequences and leads to human-interpretable internal representations of characters.