Elamur Research⌕
☰
home / papers / 2024 / 生成 AI は含意関係認識ができるのか

生成 AI は含意関係認識ができるのか

Summary

The authors test whether GPT-4 can really recognize textual entailment (NLI) in Japanese without fine-tuning and without examples in the prompt. Their starting point is that high scores after fine-tuning may reflect learned surface patterns rather than logical reasoning. They first compared seven fine-tuned BERT-family models on JNLI with GPT-4 in few-shot and zero-shot settings. Error analysis showed that the model's weakest points are grammatical tense, quantifiers and numerals. For each phenomenon, the authors built a separate set of sentence pairs. They tested gpt-4-1106-preview in zero-shot with a prompt that imitates fact-checking. On the grammatical tense set, GPT-4's accuracy is almost equal to random guessing. On the quantifier set, the result is noticeably better but still not good enough for practical use. The model's answers also change sharply when a single word is replaced or parts of a sentence are rearranged. The work shows that it is still risky to rely on such models for fact-checking and fake news detection. It also suggests what should be improved in future language models.

Key points

  • Preliminary experiment on JNLI (valid-v1.1): the best fine-tuned model was deberta-v3-base, with accuracy 0.92. GPT-4 scored 0.73 in few-shot and 0.72 in zero-shot, so it fell behind all seven BERT-family models.
  • Two new datasets of 600 pairs each. The grammatical tense set has two labels, with 300 pairs each for "entailment" and "contradiction"; GPT-4 generated the base sentences. The quantifier and numeral set has three labels, with 200 pairs each.
  • Grammatical tense: accuracy 0.52, against a chance level of 0.50. The model very rarely answered "not a fact", so recall for this label was only 0.03. For "present tense → future tense" pairs, it got 0 out of 50 correct with both prompt variants. Replacing the words "fact / not a fact" with "correct / incorrect" did not change the picture.
  • Quantifiers and numerals: accuracy 0.74. In a typical error, the model assumes that "many people" entails "one person". It also accepts as fact a hypothesis that adds a quantitative detail absent from the premise.
  • Answer instability (10 runs each, temperature 0.1): when the verb "will decrease" was replaced with "will increase", the model gave 9 "neutral" answers and 1 "not a fact" where it had given 10 "fact" answers. Rearranging parts of the sentence produced a 5-to-5 split. When "Russia" was replaced with "the USA", a 7/3 split turned into 10 "not a fact" answers.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{2024ai,
  title = {{生成 AI は含意関係認識ができるのか}},
  author = {荒沢康平 and 狩野芳伸},
  booktitle = {言語処理学会 第30回年次大会 発表論文集 (NLP2024)},
  year = {2024},
  url = {https://www.anlp.jp/proceedings/annual_meeting/2024/#P6-13}
}

References

14 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions, 2023. 2311.05232 https://arxiv.org/abs/2311.05… ↗ our card
  2. 吉越卓見, 河原大輔, 黒橋禎夫. 機械翻訳を用いた自 然言語推論データセットの多言語化. 情報処理学 会 第 244 回自然言語処理研究会, Vol. 2020-NL-244, No. 6, pp. 1–8, 7 2020. find
  3. 谷中瞳, 峯島宏次. Jsick: 日本語構成的推論・類似度 データセットの構築. 人工知能学会全国大会論文集, Vol. JSAI2021, pp. 4J3GS6f02–4J3GS6f02, 2021. find
  4. 栗原健太郎, 河原大輔, 柴田知秀. Jglue: 日本語言語 理解ベンチマーク. 自然言語処理, Vol. 30, No. 1, pp. 63–87, 2023. find
  5. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. A large annotated corpus for learning natural language inference. In Lluı́s Màrquez, Chris Callison-Burch, and Jian Su, editors, Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing , pp… find
  6. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Jill Burstein, Christy Doran, and Thamar Solorio, editors, Proceedings of the 2019 Conference of the North American Chapter of the Association for Co… find
  7. Marco Marelli, Stefano Menini, Marco Baroni, Luisa Bentivogli, Raffaella Bernardi, and Roberto Zamparelli. A SICK cure for the evaluation of compositional distributional semantic models. In Nicoletta Calzolari, Khalid Choukri, Thierry Declerck, Hrafn Loftsson, Bente Maegaard, Joseph Mariani, Asuncio… find
  8. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , Vol. 30, , 2017. find
  9. OpenAI : Josh Achiam and et al. Gpt-4 technical report, 2023. 2303.08774 https://arxiv.org/abs/2303.08774. ↗ our card
  10. llm-jp-eval リ ー ダ ー ボ ー ド, （2024-1 閲 覧）. https://wandb.ai/llm-jp-eval/test-eval/reports/llm-jp-eval---Vmlldzo1NzE0NjA1?accessToken=s09hm7xrqg43ls8i25am6t0r7iwjpninw\\ zeelqqgbx53zivlm9s04ixfpv3xgiwm. ↗
  11. Akira Sasaki, Masato Hirakawa, Shintaro Horie, and Tomoaki Nakamura. Elyza-japanese-llama-2-7b, 2023. https://huggingface.co/elyza/ELYZA-japanese-Llama-2-7b. ↗
  12. Hugo Touvron and et al. Llama 2: Open foundation and fine-tuned chat models, 2023. 2307.09288 https://arxiv.org/abs/2307.09288. ↗
  13. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chri… find
  14. Pengcheng He, Jianfeng Gao, and Weizhu Chen. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. In Pro- find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/arasawa-2024-can-generative-ai-recognise/