TL;DR
Character-level language models (LMs) don't just memorize character sequences; they internally develop mechanisms to identify higher-order linguistic units like morphemes and words. By probing a "wordless" LSTM, researchers discovered that the model learns to recognize word boundaries and understands the syntactic rules governing how suffixes attach to specific parts of speech.
What Problem Does This Paper Address?
Character-level models are widely used in NLP because they handle "open vocabularies" (they don't break when they see a new word) and share parameters between related words (like run and running). While researchers often claim these models learn "morphosyntactic" properties, these claims are usually based on intuition rather than direct evidence.
The authors set out to answer two fundamental questions:
- What do these models actually learn about morphology (word structure) and syntax (word relationships)?
- How do they learn it without being explicitly told where words or morphemes begin and end?
Methods
The researchers trained a "wordless" character-level Recurrent Neural Network (RNN) using Long Short-Term Memory (LSTM) units. "Wordless" means the model treats spaces just like any other character, with no prior knowledge of word boundaries.
To peek inside the "black box," they used two main techniques:
- Hidden Unit Analysis: They tracked the activation of individual neurons in the LSTM to see if any specific units responded to linguistic patterns.
- Probing Models: They built two auxiliary models—C2M (Character-to-Morpheme) for segmentation and C2T (Character-to-Tag) for Part-of-Speech (POS) tagging. These probes used the frozen hidden states of the LM as input, testing how much linguistic information was already encoded in the LM's representations.
- Nonce Word Analysis: They sampled the model to generate new, non-existent words (like indicatement) to see if the model applied English rules productively.
Datasets
- Training: The first 7 million character tokens from the English Wikipedia corpus.
- Morphological Segmentation (C2M): Gold standard segmentations from MorphoChallenge 2010 and Hutmegs 1.0.
- POS Tagging (C2T): The English Universal Dependencies (UD) corpus.
- Linguistic Reference: Selectional restrictions were validated against Fabb’s (1988) study of 43 English suffixes.
Key Insights and Ideas
- The "Word Unit": The most striking discovery was a single hidden unit that consistently peaks at word and morpheme boundaries. This unit isn't just predicting a space; it acts as a signal that a familiar linguistic pattern has been completed.
- Learning by Extrapolation: The model learns to identify morpheme boundaries (like the ing in drinking) by extrapolating from word boundaries. Because many morphemes can also stand alone as words, the model uses its knowledge of word edges to "guess" where internal breaks occur.
- Productive "Smart" Errors: When generating text, the model produced nonce words like breaked and indicatement. While these are technically incorrect in English, they are "grammatically logical" errors. They show the model has learned the general rules of suffixation (adding -ed to verbs or -ment to form nouns) even if it hasn't mastered every irregular exception.
- Selectional Restrictions: The model demonstrated an "awareness" of which suffixes belong to which word classes. For example, it correctly assigned a higher probability to the suffix -ment when following a verbal base than a nominal one.
Key Results
- Segmentation (C2M): The probe achieved an F1 score of 53.3. While lower than state-of-the-art supervised models, it confirms the LM encodes significant boundary information. Specifically, it reached a 90.9 F1 score for identifying actual word endings (EOW).
- POS Tagging (C2T): The model achieved 87.06% word-level accuracy. This is impressive because the model only had access to past context (no looking ahead), suggesting the LM is particularly good at encoding syntactic information within its hidden states.
- Suffix Logic: 11 out of 19 tested suffixes correctly "selected" their base category (e.g., deverbal suffixes preferring verbal bases) based on probability scores.
Limitations and Future Work
- English-Centricity: The authors acknowledge that English has "fairly poor" inflectional morphology. In languages like Turkish or Finnish, where word boundaries are less clear and morphology is more complex, a character-level LM might struggle to learn these same regularities.
- Model Luck: The authors note that different initializations or architectures might not always produce a unit as interpretable as their "word unit."
- Future Directions: The researchers suggest applying this analysis to morphologically richer languages to see if the "extrapolation from word boundaries" theory still holds up when the linguistic signals are noisier.
Summary updated