Limits for learning with language models. Asher, N., Bhar, S., Chaturvedi, A., Hunter, J., & Paul, S. In Palmer, A. & Camacho-collados, J., editors, Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 236–248, Toronto, Canada, July, 2023. Association for Computational Linguistics. ![Limits for learning with language models [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex With the advent of large language models (LLMs), the trend in NLP has been to train LLMs on vast amounts of data to solve diverse language understanding and generation tasks. The list of LLM successes is long and varied. Nevertheless, several recent papers provide empirical evidence that LLMs fail to capture important aspects of linguistic meaning. Focusing on universal quantification, we provide a theoretical foundation for these empirical findings by proving that LLMs cannot learn certain fundamental semantic properties including semantic entailment and consistency as they are defined in formal semantics. More generally, we show that LLMs are unable to learn concepts beyond the first level of the Borel Hierarchy, which imposes severe limits on the ability of LMs, both large and small, to capture many aspects of linguistic meaning. This means that LLMs will operate without formal guarantees on tasks that require entailments and deep linguistic understanding.
@inproceedings{asher_limits_2023,
address = {Toronto, Canada},
title = {Limits for learning with language models},
url = {https://aclanthology.org/2023.starsem-1.22},
doi = {10.18653/v1/2023.starsem-1.22},
abstract = {With the advent of large language models (LLMs), the trend in NLP has been to train LLMs on vast amounts of data to solve diverse language understanding and generation tasks. The list of LLM successes is long and varied. Nevertheless, several recent papers provide empirical evidence that LLMs fail to capture important aspects of linguistic meaning. Focusing on universal quantification, we provide a theoretical foundation for these empirical findings by proving that LLMs cannot learn certain fundamental semantic properties including semantic entailment and consistency as they are defined in formal semantics. More generally, we show that LLMs are unable to learn concepts beyond the first level of the Borel Hierarchy, which imposes severe limits on the ability of LMs, both large and small, to capture many aspects of linguistic meaning. This means that LLMs will operate without formal guarantees on tasks that require entailments and deep linguistic understanding.},
urldate = {2024-11-02},
booktitle = {Proceedings of the 12th {Joint} {Conference} on {Lexical} and {Computational} {Semantics} (*{SEM} 2023)},
publisher = {Association for Computational Linguistics},
author = {Asher, Nicholas and Bhar, Swarnadeep and Chaturvedi, Akshay and Hunter, Julie and Paul, Soumya},
editor = {Palmer, Alexis and Camacho-collados, Jose},
month = jul,
year = {2023},
pages = {236--248},
}
Downloads: 0
{"_id":"azFfSeaFivR3FRWbM","bibbaseid":"asher-bhar-chaturvedi-hunter-paul-limitsforlearningwithlanguagemodels-2023","author_short":["Asher, N.","Bhar, S.","Chaturvedi, A.","Hunter, J.","Paul, S."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","address":"Toronto, Canada","title":"Limits for learning with language models","url":"https://aclanthology.org/2023.starsem-1.22","doi":"10.18653/v1/2023.starsem-1.22","abstract":"With the advent of large language models (LLMs), the trend in NLP has been to train LLMs on vast amounts of data to solve diverse language understanding and generation tasks. The list of LLM successes is long and varied. Nevertheless, several recent papers provide empirical evidence that LLMs fail to capture important aspects of linguistic meaning. Focusing on universal quantification, we provide a theoretical foundation for these empirical findings by proving that LLMs cannot learn certain fundamental semantic properties including semantic entailment and consistency as they are defined in formal semantics. More generally, we show that LLMs are unable to learn concepts beyond the first level of the Borel Hierarchy, which imposes severe limits on the ability of LMs, both large and small, to capture many aspects of linguistic meaning. This means that LLMs will operate without formal guarantees on tasks that require entailments and deep linguistic understanding.","urldate":"2024-11-02","booktitle":"Proceedings of the 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023)","publisher":"Association for Computational Linguistics","author":[{"propositions":[],"lastnames":["Asher"],"firstnames":["Nicholas"],"suffixes":[]},{"propositions":[],"lastnames":["Bhar"],"firstnames":["Swarnadeep"],"suffixes":[]},{"propositions":[],"lastnames":["Chaturvedi"],"firstnames":["Akshay"],"suffixes":[]},{"propositions":[],"lastnames":["Hunter"],"firstnames":["Julie"],"suffixes":[]},{"propositions":[],"lastnames":["Paul"],"firstnames":["Soumya"],"suffixes":[]}],"editor":[{"propositions":[],"lastnames":["Palmer"],"firstnames":["Alexis"],"suffixes":[]},{"propositions":[],"lastnames":["Camacho-collados"],"firstnames":["Jose"],"suffixes":[]}],"month":"July","year":"2023","pages":"236–248","bibtex":"@inproceedings{asher_limits_2023,\n\taddress = {Toronto, Canada},\n\ttitle = {Limits for learning with language models},\n\turl = {https://aclanthology.org/2023.starsem-1.22},\n\tdoi = {10.18653/v1/2023.starsem-1.22},\n\tabstract = {With the advent of large language models (LLMs), the trend in NLP has been to train LLMs on vast amounts of data to solve diverse language understanding and generation tasks. The list of LLM successes is long and varied. Nevertheless, several recent papers provide empirical evidence that LLMs fail to capture important aspects of linguistic meaning. Focusing on universal quantification, we provide a theoretical foundation for these empirical findings by proving that LLMs cannot learn certain fundamental semantic properties including semantic entailment and consistency as they are defined in formal semantics. More generally, we show that LLMs are unable to learn concepts beyond the first level of the Borel Hierarchy, which imposes severe limits on the ability of LMs, both large and small, to capture many aspects of linguistic meaning. This means that LLMs will operate without formal guarantees on tasks that require entailments and deep linguistic understanding.},\n\turldate = {2024-11-02},\n\tbooktitle = {Proceedings of the 12th {Joint} {Conference} on {Lexical} and {Computational} {Semantics} (*{SEM} 2023)},\n\tpublisher = {Association for Computational Linguistics},\n\tauthor = {Asher, Nicholas and Bhar, Swarnadeep and Chaturvedi, Akshay and Hunter, Julie and Paul, Soumya},\n\teditor = {Palmer, Alexis and Camacho-collados, Jose},\n\tmonth = jul,\n\tyear = {2023},\n\tpages = {236--248},\n}\n\n\n\n\n\n\n\n","author_short":["Asher, N.","Bhar, S.","Chaturvedi, A.","Hunter, J.","Paul, S."],"editor_short":["Palmer, A.","Camacho-collados, J."],"key":"asher_limits_2023","id":"asher_limits_2023","bibbaseid":"asher-bhar-chaturvedi-hunter-paul-limitsforlearningwithlanguagemodels-2023","role":"author","urls":{"Paper":"https://aclanthology.org/2023.starsem-1.22"},"metadata":{"authorlinks":{}},"html":""},"bibtype":"inproceedings","biburl":"https://bibbase.org/zotero/saurabhr","dataSources":["nxjWwW7fWbb5tfpKz"],"keywords":[],"search_terms":["limits","learning","language","models","asher","bhar","chaturvedi","hunter","paul"],"title":"Limits for learning with language models","year":2023}