Generics are statements that express generalizations and are used to communicate generalizable knowledge. While generics convey general truths (e.g., Birds can fly ), they often allow for exceptions (e.g., penguins do not fly). Nonetheless, generics form the basis of how we communicate our commonsense about the world [1, 2]. We explored the interpretation of generics in Masked Language Models (MLMs), building on psycholinguistic experimental designs. As this interpretation requires a comparison with overtly quantified sentences, we investigated i) the probability of quantifiers, ii) the internal representation of nouns in generic vs. quantified sentences, and iii) whether the presence of a generic sentence as context inuflences quantifiers' probabilities. The outcomes confirm that MLMs are insensitive to quantification; nevertheless, they appear to encode a meaning associated with the generic form, which leads them to reshape the probability associated with various quantifiers when the generic sentence is provided as context.
there is any bias towards some of them (for ex- 2.2. Genericity in NLP
ample, a generic overgeneralization efect such
as that found in humans by Leslie et al. [
the creation and annotation of resources for identifying generic expressions as opposed to non-generic ones and, based on these resources, on the development of automatic annotation systems [10, 11, 12, among others]
To the best of our knowledge, there are no studies investigating the interpretation of generalizations by LLMs, except for the recent work by Ralethe and Buys [13], which addresses the generic overgeneralization efect in BERT and RoBERTa. The authors argue that these models sufer from overgeneralization by assessing how many times one or more of all, every, most, some, few and many are predicted in a masked sentence like [MASK] lions have manes: the higher the rank of the quantifiers, the stronger the LM exhibits the GOG efect. However, the GOG efect refers to the acceptance of universally quantified sentences, not just quantified ones; therefore, we can speak of overgeneralization only when the preferred quantifier is the universal one ( all or every). For this reason, we first propose a similar task to evaluate the probability distribution of various quantifiers, distinguishing between them qualitatively.
are publicly available1.
Experimental evidence has revealed a generic bias for
which people tend to overgeneralize from the truth of a
generic to the truth of the corresponding universal state- 3. Materials and Methods
ment [
Cimpian et al. [7] conducted a series of studies inves- to a concept, thus is considered a defining characteristic tigating the relationship between genericity and preva- of the concept (e.g., triangles have three sides). In this lence and found an inferential asymmetry in the meaning type of sentences, the generic is de facto equivalent to of generics. People tend to judge a generic sentence about the corresponding universal quantified statement (e.g., a novel category as true even if they have been informed all triangles have three sides). In principled sentences, that only a certain percentage of the kind (on average, up the property has a strong association with the concept. to less than 70 percent) possess the property in question This category includes both properties that are viewed as (Truth Condition task). However, when asked to estimate inherent, or connected in a principled way with a concept how many members of the kind possess the property, (e.g., birds can fly. ), and properties that are uncommon given the generic (Implied Prevalence task), they tend to and often dangerous (e.g., sharks attack swimmers); this assign very high percentages (on average, very close to last case is the one that Leslie [16, 17] defines striking. Fi100 percent). This study indicates that generic sentences nally, characterizing sentences express a non-accidental require little evidence to be judged true but have substan- relationship between property and concept, based only tial implications, since the properties they predicate tend on absolute or relative prevalence among category memto be interpreted as applying to virtually all members of bers. These generics concern properties that are neither the category. deeply connected to the concept nor striking, but occur in the majority (majority characteristic generics for [16, 17], e.g., Cars have radios) or in the minority of members of 1https://github.com/claudiacollacciani/ the category (minority characteristic generics for [16, 17], Interpretation-of-Generalization-in-Masked-Language-Models e.g., Lions have manes).
probability distribution of quantifiers, if there is any bias
towards some of them (i.e., overgeneralization efect), and
possibly whether the belonging of sentences to diferent
categories impacts it (as observed for humans by Leslie
et al. [
207 generic sentences, picking only the ones in the bare plural form (e.g., Tigers are striped), excluding indefinite and definite singulars (e.g., A/The tiger is striped). All these syntactic forms can express generic meanings, but the bare plural is the only surface form in English that gives rise to a generic interpretation unambiguously [19]. For this and other reasons, this is considered as the paradigmatic case, and it is the one that has been used in the psycholinguistic experiments from which we draw inspiration.
for the base models, as the larger counterparts show a similar trend (all boxplots are in Appendix B). Overall, all models consider few the least likely. This outcome reflects our expectations: as the selected sentences are Models We experimented with BERT and RoBERTa, all generalizations, they are, in most cases, referable to two bidirectional Masked Language Models (MLMs) a substantial number of category members, rarely to based on the Transformer architecture. BERT [20] is ‘few’ members. Apart from that, BERT and RoBERTa trained both on a masked language modeling task and show diferent probability distributions of quantifiers, on a next sentence prediction task, as the model receives regarding some and all in particular. BERT models assign sentence pairs in input and has to predict whether the a higher probability to the existential and proportional second sentence is after the first one in the training data. quantifiers ( some, many, and most) than the universal BERT has been trained on the BookCorpus and the En- quantifier all, and some is overall the most expected. The glish Wikipedia for around 3300M tokens. We employed diferences among the quantifier scores are statistically the bert-base-uncased and bert-large-uncased significant 3, with few exceptions (see Appendix B). It is pre-trained versions, which difer in terms of parameters 3By relying on Wilcoxon Signed-Rank Test statistical test. (110M and 340M parameters, respectively). On the other hand, RoBERTa [21] has the same architecture as BERT; however, it introduces several parameter optimization choices, such as dynamic masking, a larger batch and vocabulary size, and the removal of the next sentence prediction objective. Another key diference is the larger training corpus: RoBERTa was trained on 160GB of texts.
We relied on the Huggingface’s Transformers2 Library to load the models and carry on our experiments.
most expected quantifier for the sentences in our dataset. Therefore, we modified the original generic sentences by placing the special token [MASK] at the beginning of each sentence, as in ‘[MASK] strawberries have a sweet lfavor .’ Then, we computed the conditional log probability of quantifiers few, some, many, most, and all in the masked position, following previous works in quantification [ 22, 13, 23]. The conditional log probability is defined as () = (|) (1) where are the words preceding and following the critical word in the sentence.
worth noticing that the reported distributions remain the same even when we separate the analysis by sentence categories. In other words, BERT models are not sensitive to the pragmatic diferences of the selected sentences.
Conversely, both all and some are the most expected
quantifiers for RoBERTa. Accordingly, the universal
quantifier all seems to be more expected by RoBERTa
than by BERT, but this distribution is not constant for
all three sentence conditions. In quasi-definitional and
characterizing sentences, some and all have the same
probability; alternatively, some is more expected in
principled sentences. We could draw that, for principled
sentences, the model prefers not to overgeneralize the
property to all members of the category. This behavior
seems to approximate Leslie et al. [
Regardless, the observed trends could be determined by the overall frequency of the quantifiers. As a sanity check, we extracted their frequency from a large corpus of English, enTenTen21 [24, 25]. We found that () > () > () > ( ) > () (frequencies are reported in Appendix A). This pattern confirms that few is not the less probable because of a frequency efect but for the properties of the sentence. Conversely, most is the less frequent but has a probability score similar to the more frequent many and all. Finally, some is overall the most frequent quantifier. This observation could partially reflect the probability outputs of BERT; however, it is not the case for RoBERTa scores.
words in Generics and Quantified
relation4.
This study has a twofold aim: i) Identify the quantiifer that shifts the noun representation closer to that of the generic statement (if possible), and ii) Localize the layers where quantification emerges. To the best of our knowledge, no previous work has explored the internal representations of quantified expressions in relation to genericity.
Results Figure 2 illustrates how the Spearman’s between the noun and the quantified version changes with respect to each hidden layer in BERT and RoBERTa base (see Appendix C for the plots of larger models and the ones reporting correlations by sentence categories). The ifrst layers do not show a diference among correlations, meaning that representations characterized by the diferent quantifiers are practically identical; this is expected, as the context is limitedly attended by the model in these layers. The following layers show a gradual change in correlation values, but BERT and RoBERTa show diferent patterns. For the first, we observe a slight decrease in scores from layers 3 to 9 (but the correlation values are still above 0.9). Conversely, the peak of the curve is at layer 5 for RoBERTa ( 0.76), while the other internal
dimensions contextual language models than cosine or Euclidean similarity measures.
formations of each token throughout the neural network.
Previous works in BERTology have reported that internal representations, also known as contextualized embeddings, encode syntactic and semantic properties in diferent hidden layers [26]. However, it is complex to localize semantic phenomena, as they spread across the entire model [27]. For our purposes, we decided to compare the contextualized embedding of a target token (strawberries) in the generic sentence (strawberries have a sweet flavor ) with the embedding of the target token in each of the corresponding quantified sentences (e.g., all strawberries have..). Following Timkey and van Schijndel [28], for each layer, we computed the similarity of the two contextualized embeddings by relying on Spearman’s cor- Figure 2: Spearman’s against model layers. layers have a constant value of around 0.73. In the last intermediate layers (8-10), we observe some drifting away between the values for the diferent quantifiers, indicating that the contribution of the quantifiers difers from one to another. For instance, in BERT-base and large the noun quantified by some is the most similar to the corresponding non-quantified. At the same time, RoBERTa has generics similar to all, which could be interpreted as an overgeneralization efect.
Intriguingly, the most similar representations are the most probable quantifiers of the previous experiment, thus confirming that the choice of the quantifier is not a frequency efect but is related to providing a representation closer to that of the generic statement. However, the values for the various quantifiers always remain close to each other and follow the same trend, so it is hard to disclose how the meaning of the quantifier afects the noun representation. Finally, all models worsened their performance at the last layer - the one producing the most context-specific representations [ 29], indicating that contextual information weakens the quantification signal.
As observed above, MLMs are not particularly sensitive to quantifiers, and the probability choices are independent of the sentence’s meaning. This outcome is mainly due to the fact that these models are agnostic to world knowledge. Therefore, we decided to test the relation between quantification and generalization from a more formal point of view: we examined how models interpret generalizations aside from their content, that is, whether they contain any linguistic information associated with the form of generics.
We reproduce the experimental design of Cimpian
et al. [7] Implied Prevalence task, in which people were
presented with a generic sentence about a novel animal
category and then asked to estimate how many members
of the category possess the characteristic predicated by
the generic (e.g., Information: Morseths have silver fur.
Question: What percentage of morseths do you think have
silver fur?). In this case, world knowledge is not called
into play, unlike in Leslie et al. [
We build the stimuli using the generic sentence as the premise in the following way: Strawberries have a sweet flavor means that [MASK] strawberries have a sweet lfavor. As in Experiment 1, we compute the log probability of quantifiers ( few, some, many, most, and all) in the masked position. This last study should answer the following question: Does the presence of a generic sentence impact the quantifier preference, compared to the a-contextualized version of the first analysis?
for the base models, as the larger counterparts show the same trend (all boxplots are in Appendix D). Surprisingly, the most expected quantifier is now all for all models5. BERT shows an inversion in the ratio between the probability associated with the universal quantifier all and that associated with the other quantifiers (apart from few), which in the first experiment were all more expected than all, whereas now are all less expected. The pattern exhibited by RoBERTa shows a less striking change than BERT; however, the probability of all with respect to the other quantifiers is still higher than in Experiment 1. As an additional check, we performed the same test with a very small dataset of 24 generic sentences about novel
cant; a few exceptions are listed in Appendix D.
categories (non-words) from Cella et al. [
The results of this experiment, when compared with the baseline for the choice of quantifiers in Experiment 1, suggest that the presence of a generic sentence as a premise does indeed have an impact on the preference of the quantifier by the models. When the generalization is provided as context, the preferred quantifier becomes all. This behavior mirrors that of people, observed in Cimpian et al. [7] experiments and later replications: people tend to estimate very high percentages (on average very close to 100 percent) in the Implicit Prevalence task.
The first two experiments raise questions about the codification of quantifiers, as it seems that the models do not substantially exhibit a strong sensitivity to quantifiers and do not encode a semantic diference in the representation of quantification. Altogether, our results suggest that the models do not appear to contain the commonsense knowledge required to interpret generics that difer in content through quantifiers. However, they seem to have encoded a meaning associated with the generic form, which leads them to reshape the probability associated with various quantifiers when the generic sentence is provided as context. In the last experiment, we observed that the models prefer the universal quantifier unanimously if preceded by a generic utterance. People behave similarly when tested on novel categories, that is, non-word categories for which subjects have no prior understanding. However, people can modulate their interpretations of generalizations in a real language setting through their world knowledge of real categories. Regardless, MLMs tend to treat real and invented categories equally, being agnostic to world knowledge. For this reason, this could be a potentially harmful bias.
cal implications. First, we observed that the language of generalization is a complex phenomenon that is hard to investigate in human processing and even more in LLMs, mostly because the investigation of generics’ interpretation makes use of quantifiers, and language
taining the same results. models often fail in tasks related to quantification. Another problem lies in the fact that it is dificult to test autoregressive models (e.g., GPT family) on tasks such as the one used in Experiment 1 because, as they do not have access to the right context, they do not have suficient information to modulate the probabilities associated with the various quantifiers accordingly. Finding ways to test autoregressive models in addition to MLMs would be desirable.
In this paper, we have not directly investigated the aspect of the attention in MLMs. Future research could address this aspect. Furthermore, future work could involve the definition of alternative tasks for investigating generalizations to make comparing models and human interpretations easier. Psycholinguistic tests on this phenomenon often rely on truth judgments, and we should be cautious about comparing human truth judgments with model outputs since they lack commonsense knowledge comparable to that of humans. Overall, further investigations are needed to clarify the interpretation of generics in language models.
AGREEMENT: ERC-2021-STG-101039777). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. We thank the anonymous reviewers for their useful suggestions. fect, Journal of Memory and Language 65 (2011) ence society. Austin, TX: Cognitive Science Society, 15–31. 2009. [7] A. Cimpian, A. C. Brandone, S. A. Gelman, Generic [19] G. N. Carlson, Reference to kinds in English., Unistatements require little evidence for acceptance versity of Massachusetts Amherst, 1977. but have powerful implications, Cognitive science [20] J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: 34 (2010) 1452–1482. Pre-training of Deep Bidirectional Transformers [8] S. Khemlani, S.-J. Leslie, S. Glucksberg, Inferences for Language Understanding, in: Proceedings of about members of kinds: The generics hypothesis, NAACL, 2019.
Language and Cognitive Processes 27 (2012) 887– [21] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, 900. O. Levy, M. Lewis, L. Zettlemoyer, V. Stoyanov, [9] S.-J. Leslie, S. A. Gelman, Quantified statements Roberta: A robustly optimized bert pretraining apare recalled as generics: Evidence from preschool proach, arXiv preprint arXiv:1907.11692 (2019). children and adults, Cognitive psychology 64 (2012) [22] M. Apidianaki, A. Garí Soler, ALL dolphins 186–214. are intelligent and SOME are friendly: Prob[10] N. Reiter, A. Frank, Identifying generic noun ing BERT for nouns’ semantic properties and phrases, in: Proceedings of the 48th annual meet- their prototypicality, in: Proceedings of the ing of the association for computational linguistics, Fourth BlackboxNLP Workshop on Analyzing 2010, pp. 40–49. and Interpreting Neural Networks for NLP, As[11] A. Friedrich, A. Palmer, M. P. Sørensen, M. Pinkal, sociation for Computational Linguistics, Punta Annotating genericity: a survey, a scheme, and a Cana, Dominican Republic, 2021, pp. 79–94. corpus, in: Proceedings of the 9th Linguistic Anno- URL: https://aclanthology.org/2021.blackboxnlp-1. tation Workshop, 2015, pp. 21–30. 7. doi:10.18653/v1/2021.blackboxnlp-1.7. [12] V. Govindarajan, B. V. Durme, A. S. White, Decom- [23] A. Gupta, Probing quantifier comprehension posing generalization: Models of generic, habitual, in large language models, arXiv preprint and episodic statements, Transactions of the As- arXiv:2306.07384 (2023). sociation for Computational Linguistics 7 (2019) [24] A. Kilgarrif, Kovář; v.; rychlý, p.; suchomel, v. the 501–517. tenten corpus family, in: 7th International Corpus [13] S. Ralethe, J. Buys, Generic overgeneralization in Linguistics Conference CL, 2013. pre-trained language models, in: Proceedings of the [25] V. Suchomel, Better web corpora for corpus lin29th International Conference on Computational guistics and nlp, Doctoral Theses. Brno: Masaryk Linguistics, International Committee on Compu- University, Faculty of Informatics (2020). tational Linguistics, Gyeongju, Republic of Korea, [26] A. Rogers, O. Kovaleva, A. Rumshisky, A 2022, pp. 3187–3196. URL: https://aclanthology.org/ primer in BERTology: What we know about 2022.coling-1.282. how BERT works, Transactions of the Associa[14] E. Allaway, J. D. Hwang, C. Bhagavatula, K. McK- tion for Computational Linguistics 8 (2020) 842– eown, D. Downey, Y. Choi, Penguins don’t fly: 866. URL: https://aclanthology.org/2020.tacl-1.54. Reasoning about generics through instantiations doi:10.1162/tacl_a_00349. and exceptions, arXiv preprint arXiv:2205.11658 [27] I. Tenney, D. Das, E. Pavlick, BERT rediscovers (2022). the classical NLP pipeline, in: Proceedings of [15] C. Bhagavatula, J. D. Hwang, D. Downey, R. Le Bras, the 57th Annual Meeting of the Association for X. Lu, L. Qin, K. Sakaguchi, S. Swayamdipta, P. West, Computational Linguistics, Association for ComY. Choi, I2D2: Inductive knowledge distillation putational Linguistics, Florence, Italy, 2019, pp. with NeuroLogic and self-imitation, in: Proceed- 4593–4601. URL: https://aclanthology.org/P19-1452. ings of the 61st Annual Meeting of the Association doi:10.18653/v1/P19-1452. for Computational Linguistics (Volume 1: Long [28] W. Timkey, M. van Schijndel, All bark and no bite: Papers), Association for Computational Linguis- Rogue dimensions in transformer language models tics, Toronto, Canada, 2023, pp. 9614–9630. URL: obscure representational quality, in: Proceedings https://aclanthology.org/2023.acl-long.535. of the 2021 Conference on Empirical Methods in [16] S.-J. Leslie, Generics and the structure of the mind, Natural Language Processing, Association for Com
Philosophical perspectives 21 (2007) 375–403. putational Linguistics, Online and Punta Cana, Do[17] S.-J. Leslie, Generics: Cognition and acquisition, minican Republic, 2021, pp. 4527–4546. URL: https:
Philosophical Review 117 (2008) 1–47. //aclanthology.org/2021.emnlp-main.372. doi:10. [18] S. Khemlani, S.-J. Leslie, S. Glucksberg, Generics, 18653/v1/2021.emnlp-main.372. prevalence, and default inferences, in: Proceedings [29] K. Ethayarajh, How contextual are contextualized of the 31st annual conference of the cognitive sci- word representations? Comparing the geometry
We extracted the frequencies from enTenTen21. The corpus is made up of texts collected from the Internet consisting of more than 60 billion tokens. The texts were downloaded in October–December 2021 and January 2022. We relied on the concordance tool provided by SketchEngine to extract the frequencies in the form ‘[quantifier][noun]’. some all many few most
N. hits
scores for each sentence category. While the trends are the same for the three conditions, the values have a slight diference in the means, with quasi-definitional sentences having a higher correlation than the other two types.
Model Group1