Just Say No
Trials and Tribulations of Teaching Large Language Models to
Understand Negation
Tereza Vrabcová
xvrabcov@fi.muni.cz
Faculty of Informatics, Masaryk University
November 7, 2024
Contents
What is negation?
Areas of interest
Large language models vs Negation
Problems
Common approaches
More problems
My research plan
T. Vrabcová ·Just Say No ·November 7, 2024 2 / 16
Do you use negation in your day to day life?
T. Vrabcová ·Just Say No ·November 7, 2024 3 / 16
Do you use negation in your day to day life?
trick question
Better question:
How often do you use negation in your day to day life?
T. Vrabcová ·Just Say No ·November 7, 2024 3 / 16
What Is Negation?
Negation in Human Communication
key component of human communication
one of the basic tools for understanding the world
philosophy: core part of individualism
creating bounds between self and the rest of the world
early in human development, 2 – 3 years of age
colloquially known as The Terrible Twos
T. Vrabcová ·Just Say No ·November 7, 2024 4 / 16
Areas of Interest
Natural Language Processing
as a part of natural language, we want to process it
negation is a non-trivial problem
number of ways to express negation
more than no, not, and n’t
negative adverbs – never, neither, rarely, barely
negative pronouns – nothing, none, nowhere
double negatives
partial vs total negation
She didn’t do it out of love.
She didn’t do it, out of love.
grammatical vs lexical negation
T. Vrabcová ·Just Say No ·November 7, 2024 5 / 16
Areas of Interest
Large Language Models (LLMs)
as a part of natural language, we want to use it to communicate
its non-triviality causes problems for LLMs
what problems?
let’s take a look
T. Vrabcová ·Just Say No ·November 7, 2024 6 / 16
LLMs vs Negation
Loves Me, Loves Me Not I
accuracy of LLM on cloze task
Allyson Ettinger (2019) [2]
A sparrow is a ____.
A sparrow is not a ____.
Nora Kassner, Hinrich Schutze (2020) [5]
Birds can ____.
Birds cannot ____.
Thinh Hung Truong et al. (2023) [7]
Paracetamol isn’t a kind of ____.
T. Vrabcová ·Just Say No ·November 7, 2024 7 / 16
LLMs vs Negation
Loves Me, Loves Me Not II
lack of accuracy of LLM on cloze task
Allyson Ettinger (2019) [2]
A sparrow is a bird.
A sparrow is not a bird.
Nora Kassner, Hinrich Schutze (2020) [5]
Birds can fly.
Birds cannot fly.
Thinh Hung Truong et al. (2023) [7]
Paracetamol isn’t a kind of medicine.
T. Vrabcová ·Just Say No ·November 7, 2024 8 / 16
LLMs vs Negation
Common Approaches
method of Reinforcement Learning with Human Feedback
(RLHF)
used by big companies such as OpenAI and Microsoft [1, 4]
problem: data is not open-source, not easily reproducible
modifying prompts
replacing words with antonyms [6]
adding more negation [3]
prepending the negative version of the prompt
T. Vrabcová ·Just Say No ·November 7, 2024 9 / 16
LLMs vs Negation
More Problems
LLM does not reason well with negation
problem of misinformation, hallucinations
the model does not know what is and what is not true
further experiments – letting LLM know the prompt can be
false [8]
T. Vrabcová ·Just Say No ·November 7, 2024 10 / 16
Research Plan
current methods focus on tackling the problem at the end
fine-tuning
prompt modification
my goal:
start at the beginning
training data
different ratios of positive and negatives examples
different processing methods to enhance negation tokens
modification of the LLM architecture to boost negation
possible pathway to enable more complex reasoning in
LLMs
T. Vrabcová ·Just Say No ·November 7, 2024 11 / 16
Research Plan
current methods focus on tackling the problem at the end
fine-tuning
prompt modification
my goal:
start at the beginning
training data
different ratios of positive and negatives examples
different processing methods to enhance negation tokens
modification of the LLM architecture to boost negation
possible pathway to enable more complex reasoning in
LLMs
Thank you for your attention
T. Vrabcová ·Just Say No ·November 7, 2024 11 / 16
Bibliography
Bibliography I
[1] David Burch. OpenAI on Reinforcement Learning With Human
Feedback (RLHF). Arize AI, May 2023. URL:
https://arize.com/blog/openai-on-rlhf/ (visited
on 10/17/2024).
[2] Allyson Ettinger. “What BERT Is Not: Lessons from a New Suite
of Psycholinguistic Diagnostics for Language Models”. In:
Transactions of the Association for Computational Linguistics 8
(2020). Ed. by Mark Johnson, Brian Roark, and Ani Nenkova,
pp. 34–48. DOI: 10.1162/tacl_a_00298. URL:
https://aclanthology.org/2020.tacl-1.3.
T. Vrabcová ·Just Say No ·November 7, 2024 12 / 16
Bibliography
Bibliography II
[3] Md Mosharaf Hossain and Eduardo Blanco. “Leveraging
Affirmative Interpretations from Negation Improves Natural
Language Understanding”. In: Proceedings of the 2022 Conference
on Empirical Methods in Natural Language Processing. Association
for Computational Linguistics, 2022, pp. 5833–5847. DOI:
10.18653/v1/2022.emnlp-main.393. URL: https:
//doi.org/10.18653/v1/2022.emnlp-main.393.
[4] Alyssa Hughes. Learning from interaction with Microsoft Copilot
(web). Microsoft Research, Mar. 2024. URL:
https://www.microsoft.com/en-
us/research/blog/learning-from-interactionwith-microsoft-copilot-web/
(visited on 10/17/2024).
T. Vrabcová ·Just Say No ·November 7, 2024 13 / 16
Bibliography
Bibliography III
[5] Nora Kassner and Hinrich Schütze. “Negated and Misprimed
Probes for Pretrained Language Models: Birds Can Talk, But
Cannot Fly”. In: Proceedings of the 58th Annual Meeting of the
Association for Computational Linguistics. Association for
Computational Linguistics, 2020. DOI:
10.18653/v1/2020.acl-main.698. URL: http:
//dx.doi.org/10.18653/v1/2020.acl-main.698.
[6] Izunna Okpala et al. “A Semantic Approach to Negation
Detection and Word Disambiguation with Natural Language
Processing”. In: Proceedings of the 2022 6th International
Conference on Natural Language Processing and Information
Retrieval. Vol. 1. NLPIR 2022. ACM, Dec. 2022, pp. 36–43. DOI:
10.1145/3582768.3582789. URL:
https://doi.org/10.1145/3582768.3582789.
T. Vrabcová ·Just Say No ·November 7, 2024 14 / 16
Bibliography
Bibliography IV
[7] Thinh Hung Truong et al. “Language models are not naysayers:
an analysis of language models on negation benchmarks”. In:
Proceedings of the 12th Joint Conference on Lexical and
Computational Semantics (*SEM 2023). Ed. by Alexis Palmer and
Jose Camacho-collados. Toronto, Canada: Association for
Computational Linguistics, July 2023, pp. 101–114. DOI:
10.18653/v1/2023.starsem-1.10. URL:
https://aclanthology.org/2023.starsem-1.10.
[8] Neeraj Varshney et al. Investigating and Addressing Hallucinations
of LLMs in Tasks Involving Negation. 2024. DOI:
10.48550/ARXIV.2406.05494. URL:
https://arxiv.org/abs/2406.05494.
T. Vrabcová ·Just Say No ·November 7, 2024 15 / 16