Pith. sign in

REVIEW 2 major objections 5 minor 36 references

Finding Structure in Language Models

T0 review · 2 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read Language models encode abstract grammar and can generalize it from indirect evidence, while adjective-order preferences largely track corpus frequencies.

desk verdict Careful psycholinguistic-style thesis with a strong structural priming chapter and an honest adjective-order study; the FiCT 'indirect evidence' claim is plausible but not yet proven without a leakage audit or matched-deletion control. read the letter →

arxiv 2411.16433 v1 pith:WM4YO2SX submitted 2024-11-25 cs.CL

classification cs.CL
keywords languagemodelsstructuralpriminglinguisticgeneralisationfilteredcorpustrainingadjectiveordernegativepolarityitemsnext-wordpredictioninterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The thesis asks whether Transformer language models—neural networks trained only to predict the next word in a sentence—acquire abstract grammatical structure rather than memorised text. It argues that they do: models show structural priming, the psycholinguistic signature in which recent exposure to a structure makes the same structure more probable next, and the effect is moderated by the same inverse-frequency and lexical-overlap factors found in humans. Training models on corpora with specific constructions filtered out still leaves them able to judge those constructions grammatically, which the author reads as generalisation from indirect evidence. At the same time, adjective-order preferences, while highly accurate on natural orders, are strongly correlated with bigram frequencies in the training data, with only limited generalisation to unseen combinations. Together the results support a two-sided picture: next-word prediction yields genuine abstract structural knowledge, but part of that knowledge is collocational rather than rule-based.

What carries the argument

The Priming Effect (PE) is the load-bearing metric: the difference in log probability of a target sentence conditioned on a congruent versus an incongruent prime, averaged over controlled prime-target corpora, which isolates structural persistence from lexical and semantic overlap. Filtered Corpus Training (FiCT) is the load-bearing method: train on a corpus with a linguistic construction removed, then evaluate with Targeted Syntactic Evaluations, with the accuracy delta between filtered and unfiltered training measuring generalisation from indirect evidence. The adjective-order analysis adds AOP-Δ, the log-probability difference between natural and swapped adjective order, correlated against relative n-gram counts extracted from the training corpus.

What would settle it

Run a FiCT experiment in which every sentence containing a prepositional-phrase-modified subject is removed from the training corpus, verified with an independent parser and n-gram search of the training files, then test the trained model on Targeted Syntactic Evaluations for that construction; if the model scores at chance, or if the filtered corpus is found to contain any direct example of the construction, the claim that models generalise from indirect evidence would be refuted.

Watch

Extended reading notes

Core claim

On its own terms, the central claim is that language models do not merely memorise training text: they build representations of abstract grammatical structure that persist across sentences and can be acquired from indirect evidence. The structural-priming experiments show that congruent primes raise the probability of a target structure even when prime and target share no lexical content, and that the effect increases with repeated exposure and decreases with distance, mirroring human production patterns. The Filtered Corpus Training experiments show that models trained on corpora from which a construction was removed can still make correct grammaticality judgements about that construction, and that Transformers' better perplexity does not translate into better linguistic generalisation than LSTMs. The adjective-order experiments complicate the picture: high accuracy on naturally occurring orders is largely predicted by simple n-gram counts, though context and unseen pairs reveal some abstraction beyond counts.

Load-bearing premise

The conclusion rests on the assumption that the filtering step removes every direct example of the target construction from training while leaving the rest of the distribution intact; if some examples leak through or the filtering changes correlated cues, the apparent generalisation could come from direct evidence.

Editorial extensions

If this is right

  • If language models genuinely prime, structural information is active at inference time and can be studied through sentence-level probability contrasts rather than probes or fine-tuning.
  • If FiCT results are correct, a language model can acquire a construction it never saw, so the language-modelling objective itself can support poverty-of-stimulus-style generalisation.
  • Perplexity and linguistic generalisation dissociate: a Transformer's perplexity advantage does not imply better structural learning, so architecture comparisons should separate the two.
  • Adjective-order results imply that claims about LM linguistic rules must be checked against training-data frequency baselines, since bigram counts alone reach around 90 percent accuracy on natural orders.
  • Inverse-frequency and lexical-overlap effects in priming suggest LM predictions incorporate abstract expectations that mirror human production patterns.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, FiCT could be pointed at adjective order: remove every double-adjective phrase from training and test whether preferences persist; if they do, the signal must come from adjective-noun statistics or semantic classes rather than memorised collocations.
  • The cumulative priming effect reported here, where more congruent primes monotonically increase the effect, resembles a minimal form of in-context learning; a direct test would compare priming curves to few-shot demonstrations on a held-out structure with matched lexical content.
  • The strong bigram correlation in adjective order implies that interpretability claims about LM linguistic rules should routinely be checked against corpus-frequency baselines; otherwise human-like accuracy may be misread as human-like representation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The thesis investigates whether Transformer language models encode abstract grammatical structure, using three complementary strands: structural priming as a psycholinguistic paradigm (Chapters 2–3), analyses connecting training-data statistics to model behaviour including the Filtered Corpus Training (FiCT) intervention (Chapters 4–6), and controlled synthetic-language experiments for evaluating interpretability methods (Chapters 7–8). The central positive claims are that language models exhibit structural priming modulated by lexical, semantic, and frequency factors in ways partly analogous to humans, and that FiCT-trained models can judge constructions that were filtered out of their training data, indicating generalization from indirect evidence. The adjective-order chapter is more cautious, showing that simple bigram frequencies from The Pile predict most adjective-order preferences, with limited generalization to unseen pairs and with context effects that go beyond n-gram counts.

Significance. If the results hold, the thesis makes a substantial contribution to the debate on memorisation versus generalisation in language models and demonstrates the value of repurposing psycholinguistic paradigms in NLP. Concrete strengths include: the carefully constructed and publicly released Prime-LM and CAP corpora; the Priming Effect metric of Chapter 2, which compares two primes against the same target and thereby avoids the prior-perplexity confound of earlier adaptation-based measures; the use of intermediate checkpoints and batch-level n-gram counts in Chapter 4 to relate model behaviour directly to its training signal; the genuinely causal FiCT training-intervention design; and the FIDAM framework in Chapter 7, which explicitly addresses faithfulness of feature-interaction attributions. The thesis is also appropriately self-critical, for example in reporting that bigram statistics explain 90.3% of adjective-order accuracy and in showing that priming asymmetries are opposite in direction to typical human asymmetries.

major comments (2)
  1. [§5.3, Fig. 5.1, §5.5] The central claim of Chapter 5—that FiCT-trained models 'have never seen instances of their evaluation targets' and therefore generalize from indirect evidence—requires that the filters actually remove all direct evidence of each target construction. The visible text describes filters based on parsing and surface heuristics (Figure 5.1; §5.3) but reports no independent audit of the filtered corpora: parse errors, coordination, or alternative syntactic realizations could leave fragments or full instances of the target construction in the 'filtered' training data, making the evaluation items partially seen. The authors should quantify filter precision and recall on a held-out sample and, ideally, verify with an n-gram or dependency search that the TSE evaluation sentences or their structural instantiations do not appear in the filtered corpus. Without such an audit, the strong 'never seen' formulation is unsupported.
  2. [§5.5, §5.7] Even if the filters are perfectly precise, the accuracy-delta comparison between filtered and unfiltered models cannot separate indirect-evidence generalization from distribution shift induced by filtering. Removing all sentences containing PP-modified subjects, for example, also changes the frequencies of specific prepositions, noun-preposition bigrams, and subject lengths, which can shift baseline statistics for the TSE items. The chapter does not include a matched-deletion control in which an equal amount of unrelated, length- and frequency-matched material is removed. Adding such a control—or showing via re-weighting that the filtered corpus preserves the relevant marginal distributions—is necessary to support the conclusion that the measured accuracy deltas reflect genuine generalization to unseen constructions rather than a shift in the training distribution.
minor comments (5)
  1. [§2.3, Eq. (2.2)] Equation (2.2) writes the conditioning context as 'tx_i' in the final term; it should be the previous target tokens tx_{<i}, as correctly written in Eq. (3.2).
  2. [Chapter 2 captions] Several cross-references in Chapter 2 are stale: the captions of Figures 2.3–2.5 refer to '§7.1.1', '§7.2', '§7.3.1', '§7.3.2' and '§7.4', which should be Sections 2.6.1–2.6.5, and Table 2.1 is referenced as 'T able2.1'.
  3. [§3.2.1] There is a duplicated word in the sentence 'unless unless explicitly mentioned otherwise'.
  4. [Fig. 5.1] Figure 5.1 labels the method as 'FICT' in the overview box while the text consistently uses 'FiCT'; the capitalization should be standardised.
  5. [Fig. 3.2] The token labels in the legend of Figure 3.2 (e.g., 'NPO2 × NDO2') are difficult to parse; the caption would benefit from listing the corresponding word positions in the example sentence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the thesis's conclusions rest on controlled corpora, external data statistics, and empirical interventions, not on outputs fitted to their own inputs.

full rationale

The derivation chain in this thesis does not reduce to its inputs. The Priming Effect (Eq. 2.1) is a defined measure, not a fitted parameter: positive s-PE is an empirical outcome on the Prime-LM corpus, and the corpus construction (§2.4) controls for lexical overlap and semantic similarity before the effect is measured. Chapter 3's regression uses model-computed surprisal and structural preference (Eqs. 3.1 and 3.4) as explanatory predictors of the same models' PE; because these are predictors in a multivariate mixed model rather than parameters fitted to reproduce PE, the significant coefficients are an empirical finding, and the paper itself flags (§3.7) that a data-distribution control would be needed to rule out indirect effects. Chapter 4 compares AOP-Δ (Eq. 4.1) against external n-gram counts from The Pile; the 90.3% bigram accuracy is independent of the model, and the 'unseen order' analysis uses checkpoint-level counts to define held-out items, so the limited generalisation claim is not equivalent to the corpus statistics. FiCT (Ch. 5) trains on filtered corpora and evaluates on targeted syntactic evaluations; the accuracy-delta comparison is an empirical intervention, and leakage or distribution shift, if present, would weaken the conclusion but are not circularity. The only self-citation with a methodological role is the author's diagNNose software (§2.5.1, §3.3.3), which is an implementation tool and not load-bearing for any of the claims. No equation is defined in terms of the quantity it purports to predict, and no fitted parameter is renamed as a prediction. Score 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical or conceptual entities are posited. The central claims rest on corpus construction choices, metric definitions, and a causal interpretation of filtered training data. The free parameters are design thresholds and sample sizes rather than fitted constants, and the axioms are domain assumptions about what the metrics and filtering procedures actually measure.

free parameters (3)
  • Prime-LM corpus size per target structure = 15,000 prime-target pairs
    Chosen via Cochran's formula with a p-value and margin of error of 0.01; a hand-set design parameter that determines the statistical power of the Priming Effect estimates.
  • Semantic similarity threshold = cosine similarity >= 0.4
    Used to construct Semantic Similarity conditions in Prime-LM; a hand-set threshold based on the 90th percentile of the core corpus distribution, affecting the measured boost from semantic similarity.
  • Number of primes per target = 10
    Adopted from Newman et al. (2021) to stabilize Priming Effect estimates; a hand-set design choice that affects variance but not the sign of the effect.
assumptions (5)
  • domain assumption Sentence probability factorizes as a product of token probabilities, and log-probability differences between congruent and incongruent primes measure structural priming (Equations 2.1 and 2.2).
    This defines the Priming Effect metric and assumes the difference isolates structure rather than lexical or semantic confounds, which the corpus controls attempt to enforce.
  • domain assumption The Prime-LM template vocabulary and manual verb and noun selection via USF norms, WordNet categories, and COCA frequency produce grammatical and semantically controlled stimuli.
    Chapter 2, Section 2.4; if the templates or manual labels admit ungrammatical or semantically associated pairs, priming effects could be lexical rather than structural.
  • ad hoc to paper FiCT filters remove all direct evidence of target constructions from training data while preserving the distributional cues needed for indirect generalization.
    Chapter 5, Section 5.3; the inference from filtering to generalization collapses if the filters leak fragments of the target constructions or distort the corpus in correlated ways.
  • domain assumption Pythia intermediate checkpoints and batch-level n-gram counts correctly identify which adjective pairs a model has seen at a given point in training.
    Chapter 4, Section 4.6; the unseen-pair analysis assumes the batch viewer counts match actual training exposure and that processing the first 10 percent of The Pile is sufficient.
  • domain assumption Targeted Syntactic Evaluations accuracy is a valid measure of linguistic generalization.
    Used throughout FiCT and the thesis; alternative surface heuristics could solve TSE items without requiring abstract grammatical generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Finding Structure in Language Models." pith.science (2026). https://pith.science/paper/WM4YO2SX

@misc{pith2026241116433,
  author       = {Pith},
  title        = {Pith review of: Finding Structure in Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WM4YO2SX}},
  note         = {Machine review of arXiv:2411.16433}
}
read the original abstract

When we speak, write or listen, we continuously make predictions based on our knowledge of a language's grammar. Remarkably, children acquire this grammatical knowledge within just a few years, enabling them to understand and generalise to novel constructions that have never been uttered before. Language models are powerful tools that create representations of language by incrementally predicting the next word in a sentence, and they have had a tremendous societal impact in recent years. The central research question of this thesis is whether these models possess a deep understanding of grammatical structure similar to that of humans. This question lies at the intersection of natural language processing, linguistics, and interpretability. To address it, we will develop novel interpretability techniques that enhance our understanding of the complex nature of large-scale language models. We approach our research question from three directions. First, we explore the presence of abstract linguistic information through structural priming, a key paradigm in psycholinguistics for uncovering grammatical structure in human language processing. Next, we examine various linguistic phenomena, such as adjective order and negative polarity items, and connect a model's comprehension of these phenomena to the data distribution on which it was trained. Finally, we introduce a controlled testbed for studying hierarchical structure in language models using various synthetic languages of increasing complexity and examine the role of feature interactions in modelling this structure. Our findings offer a detailed account of the grammatical knowledge embedded in language model representations and provide several directions for investigating fundamental linguistic questions using computational methods.

Figures

Figures reproduced from arXiv: 2411.16433 by the authors.

Figure 1.1
Figure 1.1. The interplay between the fields of linguistics and NLP: linguistic concepts are highly useful for interpreting model behaviour, but to what extent can we reverse this direction? structure is now represented implicitly and is acquired by a model without any explicit su￾pervision. This shift raises an interesting question: do these latent notions of grammar align with the explicit grammatical representations of the p… view at source ↗
Figure 2.1
Figure 2.1. Our Priming Effect metric compares the impact of two prime sentences with different structures on a single target exhibiting one of the structures. Prasad et al. (2019) examine the impact of a single prime structure on two target sentences. paradigm as well, by comparing the target sentence probabilities of two fine-tuned models. In the experiments presented in this chapter, we focus on priming as residual activatio… view at source ↗
Figure 2.2
Figure 2.2. Priming Effect results of all models on the core corpus, across the four syntactic structures. Error bars denote 99% confidence intervals of the mean. The GPT2-large model that will be explored in more detail in §3.6 has been highlighted. There are two models that exhibit symmetrical priming for both transitive and dative alternations: GPT2-large and GPT2-xl. The other GPT2 models exhibit symmetrical priming for tra… view at source ↗
Figures from the paper (41 more)
Figure 2.3
Figure 2.3. Figure 2.3: Results for GPT2-large on the experiments described in and §2.6.1 and §2.6.2: A. measures the impact of semantic similarity between prime and target, B. the impact of lexical overlap between prime and target, and C. whether priming is affected by the semantic implaus…
Figure 2.4
Figure 2.4. Figure 2.4: Results for GPT2-large on the experiments described in §2.6.3: A. recency effect on priming, by increasing the distance between prime and target with additional intransitive sentences: each bar denotes a different position of the prime (Px), surrounded by intervening…
Figure 2.5
Figure 2.5. Figure 2.5: Results for GPT2-large on the experiment described in §2.6.5, measuring the impact of in￾creasing the complexity of one noun phrase per sentence in prime and target. thing akin to a hierarchical phrase-structure rule such as VP → NP NP or VP → NP PP rather than as a …
Figure 2.6
Figure 2.6. Figure 2.6: Priming Effects for three masked language models on the core corpus: BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and ALBERT (Lan et al., 2020). To compute sentence probabilities we utilise the pseudo-log-likelihood of Salazar et al. (2020), masking out on…
Figure 3.1
Figure 3.1. Figure 3.1: We plot PE results against one another. The four quadrants in this ‘PE space’: balanced priming where the PE is positive in both directions, skewed priming where it’s only positive in one, and inverse priming when the PE is negative in both directions. There exists a…
Figure 3.2
Figure 3.2. Figure 3.2: The w-PE scores for the Core and Lexical Overlap conditions. Scores are grouped by token (based on colour) and model (based on shape). To exemplify how these Priming Space coordinates map to a bar chart, we show the Mistral-7b-instruct scores at the top of each plot.…
Figure 3.3
Figure 3.3. Figure 3.3: The token-level Priming Effect reveals which token predictions in the target sentence con￾tributed the most to the overall sentence-level Priming Effect, here averaged for Llama-2 over the Core. It is inversely correlated up to the point of divergence between the two…
Figure 3.4
Figure 3.4. Figure 3.4: Structural preferences for GPT2-large and Llama-2, expressing the preference of a ditransitive verb with respect to a prepositional object versus a double object construction. The verb order of Llama-2 is based on the sorted order of GPT2-large. preferences found in …
Figure 3.5
Figure 3.5. Figure 3.5: LMM coefficients for (A) predicting sδ-PE(po) and (B) sδ-PE(do), shown side-by-side with reported effects for predicting human priming in production- and corpus-based studies. Significant LLM coefficients (p < 10−3 ) are denoted by an asterisk. Results We report full…
Figure 4.1
Figure 4.1. Figure 4.1: We connect the adjective order preferences(AOP-Δ, §4.3.1) of language models (here Pythia￾12b) to the adjective order frequencies of the corpus they have been trained on (The Pile). We highlight various regions of interest: adjective pairs for which both orders are r…
Figure 4.2
Figure 4.2. Figure 4.2: A–B: AOP-% and AOP-Δ scores for Pythia models of increasing size. C–D: AOP-% and AOP-Δ scores for Pythia-1.4b during training. We highlight the three learning phases: 1) initialisation, 2) acquisition, and 3) consolidation. AOP-Δ does increase consistently with size:…
Figure 4.3
Figure 4.3. Figure 4.3: Average token probabilities for the original and swapped adjective orders on Pythia-12b, without and with sentence context (A–B), as well as the token-level differences (C) that correspond to the difference between the curves in (A) and (B). In [PITH_FULL_IMAGE:figu…
Figure 4.4
Figure 4.4. Figure 4.4: Correlations during training of LM probabilities for single adjectives, adjective pairs, and adjective-noun triplets with respect to their frequency in The Pile. tion whether LMs ‘hide’ a meaningful linguistic theory that extends beyond established factors. 4.6 Exper…
Figure 4.5
Figure 4.5. Figure 4.5: The contextual AOP-% performance for Pythia-70m and 1.4b across training, split out for items that have been seen 0, 1, 2 to 10, and more than 10 times at each specific checkpoint. with similar formulations for adjective bigrams (%(a1a2)) and unigrams (%(a1)). Predic…
Figure 4.6
Figure 4.6. Figure 4.6: Relative sizes of the 4 splits of CAP based on the number of bigram occurrences seen at each point in training. Note that an additional constraint on each split is that the swapped adjective order has not been seen at all, which explains why the 4 splits do not sum u…
Figure 4.7
Figure 4.7. Figure 4.7 [PITH_FULL_IMAGE:figures/full_fig_p095_4_7.png]
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p100_5.png]
Figure 5.1
Figure 5.1. Figure 5.1: 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 per…
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p105_5.png]
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p106_5.png]
Figure 5.2
Figure 5.2. Figure 5.2: BLiMP benchmark accuracy for the models trained on the full corpus, and accuracy delta (Δ(M, F, B)) for the filtered corpora, averaged across seeds. Boxes with bold outlines correspond to bench￾marks targeted by the model’s corpus filter (i.e. where F = F(B)). The ac…
Figure 5.3
Figure 5.3. Figure 5.3: Perplexity scores on the test corpus (C 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. In other words, we find …
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p112_5.png]
Figure 5.4
Figure 5.4. Figure 5.4: Log probability differences between grammatical and ungrammatical minimal pairs (PΔ(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 …
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p113_5.png]
Figure 5.5
Figure 5.5. Figure 5.5: A: PΔ 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 PΔ scores going from the full to the Filtered model. The closer to the…
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p114_5.png]
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p115_5.png]
Figure 6.1
Figure 6.1. Figure 6.1: The pipeline of our experimental setup. We start by computing the hidden states h ↓ t (within a DM environment ahead of the NPI) and h ↑ t (within a UM environment). These hidden states are then used for training the monotonicity DC (Exp. 1 & 4), and to compare Plm(n…
Figure 6.2
Figure 6.2. Figure 6.2: Accuracy and standard deviation on the monotonicity diagnostic classification task, averaged over 5 seeds for each model type. The All-ENV column denotes train/test split procedure sampled uniformly over all environment class; other columns denote accuracy on one env…
Figure 6.3
Figure 6.3. Figure 6.3: Accuracy on the NPI acceptability task—based on whether the NPI was assigned a higher prob￾ability in the DM environment than in its UM counterpart. Results The results of our first experiment are shown in the top row of [PITH_FULL_IMAGE:figures/full_fig_p126_6_3.png]
Figure 6.4
Figure 6.4. Figure 6.4: The DC Ranking experiment, in which we investigate whether the monotonicity DC and the LM decoder base their predictions on similar cues, by computing and ranking the cosine similarities between the DC weights and the decoder weights of each token. might in part be d…
Figure 6.5
Figure 6.5. Figure 6.5: Results on the median NPI rank task. A low median rank indicates that the monotonicity DC uses the same representational information as the NPI decoder. 113 [PITH_FULL_IMAGE:figures/full_fig_p128_6_5.png]
Figure 7.1
Figure 7.1. Figure 7.1: We generate a corpus based on a PCFG, and create negative examples by corrupting the generated corpus. Next, we train a neural model to predict whether a string is well-formed, forcing the model to obtain a comprehensive understanding of the rules of the language. Th…
Figure 7.2
Figure 7.2. Figure 7.2: Example for the computation of the Average Relative Rank metric. For each row we compute the relative rank of the interaction of interest (here the Dyck language), and these row-wise relative ranks are averaged into a single score between 0 and 1. A random interactio…
Figure 8.1
Figure 8.1. Figure 8.1: Conceptual overview of our experimental pipeline. First, we induce a massive probabilistic grammar from a natural language treebank. From this grammar we generate a corpus, that is used to train vari￾ous language models on. Then, with access to the true distribution …
Figure 8.2
Figure 8.2. Figure 8.2: Relationship between a token’s frequency rank and its frequency, which is near linear on a log￾log plot. Rank and frequency are computed over two disjoint splits of the corpora, following the procedure of Piantadosi (2014). not allow this, the data distribution that …
Figure 8.3
Figure 8.3. Figure 8.3: Distribution over sentence length for the three corpora, and the correlation of n-gram distribu￾tions with respect to the original Treebank data. (2014), who argues that frequency and rank should not be computed on the same corpus, since that way a token’s rank is al…
Figure 8.4
Figure 8.4. Figure 8.4: Language modelling performance expressed as a function of training size (a) and model size (b). Relative perplexity is computed with respect to the PCFG’s perplexity lower bound, averaged over 3 seeds. to reduce overfitting. In order to compare causal and masked LM p…
Figure 8.5
Figure 8.5. Figure 8.5: The LM probability divergence from the PCFG distribution, aggregated by general part-of￾speech classes, for the DeBERTa model. 8.6 Model Interpretability We investigate the behaviour and learning dynamics of the DeBERTa model from §8.5.3. Our setup provides an intere…
Figure 8.6
Figure 8.6. Figure 8.6: Left: POS Probing performance across layers (x-axis) and split out for multiple checkpoints through training. Right: Spearman correlation between POS probing probabilities of the true class and the token probabilities of the PCFG, LM, and probability divergence. the …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages

  1. [1]

    16.5 18.8

    In solution, lead(II) hydroxide is a somewhat weak base, forming lead(II) ion, Pb, un- der weakly acidic conditions. 16.5 18.8

  2. [2]

    16.9 15.4

    It has the only deep water port in Cambodia. 16.9 15.4

  3. [3]

    15.4 16.3

    The greatest social reforms in Denmark are certainly the work of thelast half-century. 15.4 16.3

  4. [4]

    16.9 14.5 Swapped order always preferred

    Icterids are unusual in songbirds because they haveconsiderable sexualdimorphism. 16.9 14.5 Swapped order always preferred

  5. [5]

    −7.1 −6.8

    Couches are usually bought in a set together with cushions, which give them a bouncey and decorative comfortabletouch. −7.1 −6.8

  6. [6]

    −7.3 −5.9

    No, its 10 feet tall, and it’s a red big monster, demon like. −7.3 −5.9

  7. [7]

    −5.8 −6.0

    It does this by following the grammatical basic rules of syntax. −5.8 −6.0

  8. [8]

    −5.8 −5.7 Context improves AOP

    The daily averageattendance from January through November 2010 was 22,133 peo- ple a day. −5.8 −5.7 Context improves AOP

Show all 36 references
  1. [9]

    0.9 16.8

    Being a spherical 3-manifold, it is the only homology 3-sphere, besides the 3-sphere itself, with a finite fundamental group. 0.9 16.8

  2. [10]

    −0.5 15.3

    Since the 2015 Styria municipal structural reform, it is part of the municipality Birkfeld. −0.5 15.3

  3. [11]

    0.2 15.4

    I won it at a fair last night. 0.2 15.4

  4. [12]

    2.6 17.5 Context worsens AOP

    I say that the man who does not play whist lays up asad old age for himself. 2.6 17.5 Context worsens AOP

  5. [13]

    12.3 4.0

    This is an ability which may have been present in their last common ancestor in the Archaean. 12.3 4.0

  6. [14]

    Jacob, you’ve had a complex partial seizure, which can cause psychosis, including reli- gious psychosis. 7.3 0.3

  7. [15]

    6.2 −0.6

    He was also a right-arm medium fast bowler with three wickets in test matches to his credit. 6.2 −0.6

  8. [16]

    4.0 −2.8 200 A.2

    He wheeled round sharply, and distinguished her lying withhelpless outspreadarms on the couch. 4.0 −2.8 200 A.2. Adjective Order A.2.2 Correlations to Cognitive Predictors 1 16 256 3000 20000 143000 0.0 0.2 0.4Spearman's 1 2 3 Length 1 16 256 3000 20000 143000 1 2 3 PMI(a;n) 1...

  9. [17]

    Jaap Jumelet, Milica Denic, Jakub Szymanik, Dieuwke Hupkes, and Shane Steinert- Threlkeld. 2021. Language models use monotonicity to assess NPI licensing . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4958–4969, Online. Association for C...

  10. [18]

    Arabella Sinclair*, Jaap Jumelet*, Willem Zuidema, and Raquel Fernández. 2022. Structural persistence in language models: Priming as a window into abstract lan- guage representations. Transactions of the Association for Computational Linguistics, 10:1031–1050

  11. [19]

    Jaap Jumelet and Willem Zuidema. 2023a. Feature interactions reveal linguistic structure in language models. In Findings of the Association for Computational Lin- guistics: ACL 2023, pages 8697–8712, T oronto, Canada. Association for Computa- tional Linguistics

  12. [20]

    Jaap Jumelet and Willem Zuidema. 2023b. T ransparency at the source: Evaluating and interpreting language models with access to the true distribution . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4354–4369, Singapore. Association for Computa...

  13. [21]

    Jaap Jumelet, Willem Zuidema, and Arabella Sinclair. 2024b. Do language mod- els exhibit human-like structural priming effects? In Findings of the Association for Computational Linguistics ACL 2024, pages 14727–14742, Bangkok, Thailand and virtual meeting. Association for Comp...

  14. [22]

    2024.Filtered corpus training (FiCT) shows that language models can generalize from indirect evidence

    Abhinav Patil*, Jaap Jumelet*, Yu Ying Chiu, Andy Lapastora, Peter Shen, Lexie W ang, Clevis Willrich, and Shane Steinert-Threlkeld. 2024.Filtered corpus training (FiCT) shows that language models can generalize from indirect evidence. Transac- tions of the Association for Com...

  15. [23]

    Jaap Jumelet, Lisa Bylinina, Willem Zuidema, and Jakub Szymanik. 2024a. Black big boxes: Do language models hide a theory of adjective order? Under Review During my PhD I have also contributed to the following papers:

  16. [24]

    Jaap Jumelet. 2020. diagNNose: A library for neural activation analysis . In Pro- ceedings of the Third BlackboxNLP W orkshop on Analyzing and Interpreting Neural Networks for NLP, pages 342–350, Online. Association for Computational Linguis- tics

  17. [25]

    2021.Language mod- elling as a multi-task problem

    Lucas W eber, Jaap Jumelet, Elia Bruni, and Dieuwke Hupkes. 2021.Language mod- elling as a multi-task problem. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main V olume, pages 2049– 2060, Online. Association fo...

  18. [26]

    T om Kersten, Hugh Mee W ong, Jaap Jumelet, and Dieuwke Hupkes. 2021. At- tention vs non-attention for a shapley-based explanation method. In Proceedings of Deep Learning Inside Out: The 2nd W orkshop on Knowledge Extraction and Integra- tion for Deep Learning Architectures, D...

  19. [27]

    2022.The birth of bias: A case study on the evolution of gender bias in an English language model

    Oskar V an Der W al, Jaap Jumelet, Katrin Schulz, and Willem Zuidema. 2022.The birth of bias: A case study on the evolution of gender bias in an English language model. In Proceedings of the 4th W orkshop on Gender Bias in Natural Language Processing (GeBNLP), pages 75–75, Sea...

  20. [28]

    Lucas W eber, Jaap Jumelet, Paul Michel, Elia Bruni, and Dieuwke Hupkes. 2023. Curriculum learning with adam: The devil is in the wrong details. CoRR, abs/2308.12202 214 B. List of Publications

  21. [29]

    Jaap Jumelet, Michael Hanna, Marianne de Heer Kloots, Anna Langedijk, Char- lotte Pouw, and Oskar van der W al. 2023. ChapGTP, ILLC’s attempt at raising a BabyLM: Improving data efficiency by automatic task formation. In Proceedings of the BabyLM Challenge at the 27th Conferen...

  22. [30]

    2023.Attribu- tion and alignment: Effects of local context repetition on utterance production and comprehension in dialogue

    Aron Molnar, Jaap Jumelet, Mario Giulianelli, and Arabella Sinclair. 2023.Attribu- tion and alignment: Effects of local context repetition on utterance production and comprehension in dialogue. In Proceedings of the 27th Conference on Computational Natural Language Learning (C...

  23. [31]

    Hosein Mohebbi, Jaap Jumelet, Michael Hanna, Afra Alishahi, and Willem Zuidema

  24. [33]

    Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet

  25. [34]

    In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4764– 4780, Mexico City, Mexico

    DecoderLens: Layerwise interpretation of encoder-decoder transformers. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4764– 4780, Mexico City, Mexico. Association for Computational Linguistics

  26. [35]

    2024.Interpretability of language models via task spaces

    Lucas W eber, Jaap Jumelet, Elia Bruni, and Dieuwke Hupkes. 2024.Interpretability of language models via task spaces . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 4522– 4538, Bangkok, Thailand. Asso...

  27. [36]

    Arabella Sinclair, Anastasia Klimovich-Gray, Jaap Jumelet, Nika Adamian, and Ag- nieszka Konopka. 2024. Structural priming in humans and large language models. Under Review 215 Structuur Ontrafelen uit T aalmodellen Samenvatting W anneer we spreken, schrijven of luisteren, cre...

  28. [2024]

    In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Tutorial Ab- stracts, pages 21–26, St

    T ransformer-specific interpretability. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Tutorial Ab- stracts, pages 21–26, St. Julian’s, Malta. Association for Computational Linguistics

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.