rahaaskari/gricean_baby

Pragmatic Evaluation of Language Models based on Their Sensitivity to Violations of Conversational Maxims

Dataset Desctiption

This is an evaluation dataset to study language models' pragmatic plausibility based on conevrsational maxims of Grice (1975). It contains of 2250 short conversations. Each conversation includes a question and two answers; one of the answers violates a conversational maxim and is therefore not pragmatically suitable. The aim is to see whether, given the question, the model assigns a higher probability score for the answer that follows the maxim. We furthermore share the results of our experiment in which we tested language models of different training data sizes. Please refer to our paper for more info, and to the Github link for reproducing the experiment.

• Curated by: Raha Askari & Judith Sieker
• Language(s) (NLP): English

Dataset Sources

• Repository: Github
• Paper: ACL Anthology

Dataset Structure

The conversational items are in the file "eval_dataset.csv". The file contains 4 columns; the first one is the maxim that the conversation belongs to, followed by the column "Question". The items of "Follower" column are the answers to the question that follow that maxim, while the items of "Violator" column are the answers that violate the maxim. There are in total 2250 conversations.

Dataset Creation

The dataset is an extended version of a psycholinguistic study by Surian et al. (1996) with 25 conversation examples to assess children's pragmatic competence based on Gricean maxims. Through automatic generation, we first generated 350 similar conversations, and then curated them manually based on the purpose of our study. After finalizing the curated version, we further added rotating speakers' names to create more items and make the data more natural. The final results consists of 2250 conversations.

References

Grice, H. P. (1975). Grice-Logic and Conversation. London: University College London Press.

Luca Surian, Simon Baron-Cohen, and Heather van der Lely. 1996. Are children with autism deaf to gricean maxims? Cognitive neuropsychiatry, 1:55–72.

Citation

If you use this dataset, please cite:

@inproceedings{askari-etal-2025-babylms,
    title = "Are {B}aby{LM}s Deaf to {G}ricean Maxims? A Pragmatic Evaluation of Sample-efficient Language Models",
    author = {Askari, Raha  and
      Zarrie{\ss}, Sina  and
      Alacam, {\"O}zge  and
      Sieker, Judith},
    editor = "Charpentier, Lucas  and
      Choshen, Leshem  and
      Cotterell, Ryan  and
      Gul, Mustafa Omer  and
      Hu, Michael Y.  and
      Liu, Jing  and
      Jumelet, Jaap  and
      Linzen, Tal  and
      Mueller, Aaron  and
      Ross, Candace  and
      Shah, Raj Sanjay  and
      Warstadt, Alex  and
      Wilcox, Ethan Gotlieb  and
      Williams, Adina",
    booktitle = "Proceedings of the First BabyLM Workshop",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.babylm-main.4/",
    pages = "52--65",
    ISBN = "TODO",
    abstract = "Implicit meanings are integral to human communication, making it essential for language models to be capable of identifying and interpreting them. Grice (1975) proposed a set of conversational maxims that guide cooperative dialogue, noting that speakers may deliberately violate these principles to express meanings beyond literal words, and that listeners, in turn, recognize such violations to draw pragmatic inferences.Building on Surian et al. (1996){'}s study of children{'}s sensitivity to violations of Gricean maxims, we introduce a novel benchmark to test whether language models pretrained on {\ensuremath{<}}10M and {\ensuremath{<}}100M tokens can distinguish maxim-adhering from maxim-violating utterances. We compare these BabyLMs across five maxims and situate their performance relative to children and a Large Language Model (LLM) pretrained on 3T tokens.We find that overall, models trained on {\ensuremath{<}}100M tokens outperform those trained on {\ensuremath{<}}10M, yet fall short of child-level and LLM competence. Our results suggest that modest data increases improve some aspects of pragmatic behavior, leading to finer-grained differentiation between pragmatic dimensions."
}