Filtered Corpus Training (FiCT) Shows that Language Models can Generalize from Indirect Evidence

TL;DR

This work introduces Filtered Corpus Training (FiCT), a causal ablation approach to measure linguistic generalization by training language models on corpora from which specific constructions are filtered out. By applying FiCT to LSTMs and Transformer models and evaluating with BLiMP using metrics like $acc\Delta$ and $P\Delta$, the authors demonstrate that both architectures can generalize from indirect evidence even when direct examples are removed, though Transformers achieve better perplexity. A key finding is the dissociation between perplexity and linguistic generalization, implying that lower perplexity does not imply stronger grammatical generalization. The results provide evidence against purely memorization-based learning in LMs and offer a framework for more fine-grained evaluation of linguistic competence, with implications for model development and evaluation practices.

Abstract

This paper introduces Filtered Corpus Training, a method that trains language models (LMs) on corpora with certain linguistic constructions filtered out from the training data, and uses it to measure the ability of LMs to perform linguistic generalization on the basis of indirect evidence. We apply the method to both LSTM and Transformer LMs (of roughly comparable size), developing filtered corpora that target a wide range of linguistic phenomena. Our results show that while transformers are better qua LMs (as measured by perplexity), both models perform equally and surprisingly well on linguistic generalization measures, suggesting that they are capable of generalizing from indirect evidence.

Figures (7)

• Figure 1: Overview of the Filtered Corpus Training methodology (FICT). For a linguistic construction of interest (e.g. prepositionally modified subjects), we filter out sentences containing that construction and train a new language model on the filtered corpus. We measure performance on targeted syntactic evaluations to assess the capacity of the LM to generalize from related constructions to this novel, unseen construction.
• Figure 2: BLiMP benchmark accuracy for the models trained on the full corpus, and accuracy delta ($\Delta(M, F, B)$) for the filtered corpora, averaged across seeds. Boxes with bold outlines correspond to benchmarks targeted by the model's corpus filter (i.e. where $F = F(B)$). The accuracy scored by a given model on a given benchmark trained on a filtered corpus can be recovered by adding its delta to the accuracy score in the "full" column of the same row.
• Figure 3: Perplexity scores on the test corpus ($C^{\textit{test}}$) and the grammatical and ungrammatical BLiMP sentences ($s^+$ & $s^-$). BLiMP scores for the full models are averaged over all benchmarks, and for the Filtered models for their corresponding benchmark.
• Figure 4: Log probability differences between grammatical and ungrammatical minimal pairs ($P\Delta(M, F)(s)$), with Transformer performance plotted against LSTM performance. Individual points are the averaged scores across the five model seeds. The four quadrants indicate the cases where i) both architectures got a correct prediction (green), ii) only one architecture got a correct prediction (orange), and iii) neither architecture was right (red). It can be seen that corpus filtering results in probability differences moving closer to the origin, and that the magnitude of the difference of the full models can create a sufficient margin for the model to generalize in the filtered cases as well.
• Figure 5: A: $P\Delta$ scores for the full Transformers and LSTMs for each BLiMP paradigm. The more positive this score, the more certain a model is in its grammaticality judgment. B: Paradigm-level differences in $P\Delta$ scores going from the full to the Filtered model. The closer to the origin, the less impact the filtering procedure had on model behavior. C: Pearson correlation of $P\Delta$ scores between the full and Filtered models. A detailed table with these results per paradigm is provided in Figure \ref{['app:fig:pdelta-table']} in Appendix \ref{['sec:app-blimp-acc']}.