TL;DR
This paper investigates how neural language models learn linguistic knowledge, specifically focusing on implicit causality (IC) verbs. They discover that models don't always exhibit human-like IC behavior due to competing linguistic constraints, particularly "pro-drop" phenomena in languages like Spanish and Italian. By strategically fine-tuning models, they demonstrate that this hidden IC knowledge can be "uncovered," suggesting that models need to learn not just constraints but also their relative importance.
What Problem Does This Paper Address?
Large language models (LLMs) have shown remarkable capabilities, leading researchers to explore what linguistic knowledge they acquire. While previous work has identified some grammatical knowledge, it's also clear that models rely on non-linguistic biases (like sentence length) that can lead to non-human-like behavior. This paper tackles a more nuanced problem: can competing linguistic phenomena within a language obscure underlying linguistic knowledge, even when models have learned it? They hypothesize that this isn't just about learning a constraint, but learning the hierarchy of constraints. They focus on implicit causality (IC) verbs, which influence pronoun resolution, and investigate this across four languages: English, Chinese, Spanish, and Italian.
Methods
The core methodology involves:
- Cross-linguistic comparison: Evaluating LLMs on IC verb bias in English, Chinese, Spanish, and Italian.
- Targeted Fine-Tuning: Modifying the models' behavior by fine-tuning them on specific datasets designed to either:
- Demote competing constraints: For Spanish and Italian, they fine-tuned models on data that discouraged pro-drop (i.e., forced overt pronouns in subject positions).
- Promote competing constraints: For English and Chinese, they fine-tuned models on data that encouraged pro-drop (i.e., removed subject pronouns).
- Behavioral Probing: Measuring the models' pronoun resolution preferences in response to IC verb sentences.
Datasets
- Training Data: The paper utilized pre-trained models (BERT, RoBERTa variants) that were trained on large, general-purpose text corpora.
- Evaluation Stimuli: Sentences were constructed based on existing psycholinguistic studies of IC verbs in English, Chinese, Spanish, and Italian. These stimuli were designed to disambiguate pronoun antecedents using gender.
- Fine-tuning Data:
- For demoting pro-drop in Spanish and Italian: Datasets derived from Universal Dependencies corpora (AnCora, ISDT, VIT) that were modified to include overt pronouns in subject positions.
- For promoting pro-drop in English and Chinese: Datasets derived from Universal Dependencies corpora (English Web Treebank, GSD, PUD) that were modified to remove subject pronouns.
Key Insights and Ideas
- Linguistic knowledge can be "hidden" by competing constraints: Models might possess knowledge of IC verbs, but this knowledge may not manifest in their behavior if another linguistic phenomenon (like pro-drop) is more dominant.
- Constraint ranking is crucial: It's not enough for a model to learn individual linguistic rules; it must also learn the relative importance or hierarchy of these rules to behave in a human-like manner.
- Cross-linguistic variation in models, not humans: While human IC bias is remarkably consistent across languages, the paper found significant differences in how LLMs exhibit this bias, particularly between English/Chinese and Spanish/Italian.
- Fine-tuning as a diagnostic tool: Targeted fine-tuning proved effective not only for uncovering hidden knowledge but also for demonstrating the impact of constraint competition.
Key Results
- English and Chinese Models: Showed a clear IC bias, consistent with prior research. Object-biased verbs increased the likelihood of object pronoun reference, and subject-biased verbs increased subject pronoun reference.
- Spanish and Italian Models: Exhibited weak or no IC bias in their baseline state. This was attributed to the strong "pro-drop" phenomenon in these languages, which disfavors overt subject pronouns and competes with IC.
- Fine-tuning Spanish/Italian: Demoting the pro-drop constraint (by fine-tuning on data with overt subject pronouns) significantly increased the IC bias in Spanish and Italian models, revealing previously hidden knowledge. For example, Italian BERT and GilBERTO showed a more than doubled subject-biased IC verb effect after fine-tuning.
- Fine-tuning English/Chinese: Promoting a pro-drop constraint (by fine-tuning on data with removed subject pronouns) significantly diminished the IC effect in English and Chinese models, even with very small amounts of fine-tuning data.
Limitations and Future Work
- RuPERTa's Lack of IC Knowledge: The paper notes that one Spanish model (RuPERTa) showed no IC knowledge even after fine-tuning, suggesting it might genuinely lack this knowledge, though further investigation is recommended.
- Generalizability of Findings: The study focused on a single phenomenon (IC verbs) and specific model architectures (BERT, RoBERTa). Future work could explore other linguistic phenomena and different model families (e.g., GPT-2).
- Representational Probing: The authors suggest pairing their behavioral analyses with representational probing to more explicitly link model representations to observed behavior.
- Constraint Competition in Other Models: Investigating how constraint competition affects models like GPT-2, which have shown impressive linguistic behavior, is a promising future direction.
Summary updated