신경망 언어 모델의 내부 표상: 탐침 분류기 기법을 중심으로
Investigating the Internal Representation of an Artificial Neural Language Model: Concentrating on the Method of Using a Probing Classifier to Assess the Result of Learning a Language
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper appraises the validity of using a probing classifier that serves as the most rigorous method in investigating the internal representation of a deep neural-network language model. Recently, such a model has been reported to display a high performance in undertaking various linguistic tasks. However, it is not easy to assess what kind of linguistic knowledge it acquires, and how robustly such knowledge is encoded in internal artificial neural networks. A probing classifier has been developed as a method to analyze the internal mechanism of such a model. We first evaluate the validity of this method in three respects by taking into account such control factors as task, function, and dataset that can cause defects when applying a probing classifier. Second, we discuss what kind of probing classifier should be applied, simple versus complex, and ponder on other alternative methods aside from using a probing classifier. Third, we consider the issues of correlation and causation in studying the relationship between probed linguistic properties and an original language model.
Abstract
This paper appraises the validity of using a probing classifier that serves as the most rigorous method in investigating the internal representation of a deep neural-network language model. Recently, such a model has been reported to display a high performance in undertaking various linguistic tasks. However, it is not easy to assess what kind of linguistic knowledge it acquires, and how robustly such knowledge is encoded in internal artificial neural networks. A probing classifier has been developed as a method to analyze the internal mechanism of such a model. We first evaluate the validity of this method in three respects by taking into account such control factors as task, function, and dataset that can cause defects when applying a probing classifier. Second, we discuss what kind of probing classifier should be applied, simple versus complex, and ponder on other alternative methods aside from using a probing classifier. Third, we consider the issues of correlation and causation in studying the relationship between probed linguistic properties and an original language model.
참고문헌 (18)
- Belinkov, Yonatan, "Probing Classifiers: Promises, Shortcomings, and Advances", Computational Linguistics, 48(1), pp.207~219 (2022)
- Clark, Kevin, "What Does Bert Look At? An Analysis of Bert’s Attention", arXiv preprint arXiv:1906.04341, ~ (2019)
- Elazar, Yanai, "Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals", Transactions of the Association for Computational Linguistics, 9, pp.160~175 (2021)
- Giulianelli, Mario, "Under the Hood: Using Diagnostic Classifiers to Investigate and Improve How Language Models Track Agreement Information", arXiv preprint arXiv:1808.08079, ~ (2018)
- Hewitt, J., "Designing and Interpreting Probes with Control Tasks", Proceedings of the 2019 Con (2019)
- Htut, Phu Mon, "Do Attention Heads in Bert Track Syntactic Dependencies?", arXiv preprint arXiv:1911.12246, ~ (2019)
- Hupkes, Dieuwke, "Visualisation And’diagnostic Classifiers’ Reveal How Recurrent and Recursive Neural Networks Process Hierarchical Structure", Journal of Artificial Intelligence Research, 61, pp.907~926 (2018)
- Liu, Nelson F, "Linguistic Knowledge and Transferability of Contextual Representations", arXiv preprint arXiv:1903.08855, ~ (2019)
- Lovering, Charles, "Predicting Inductive Biases of Pre-Trained Models", International Conference on Learning Representations (2020)
- Mareček, David, "From Balustrades to Pierre Vinken: Looking for Syntax in Transformer Self-Attentions", arXiv preprint arXiv:1906.01958, ~ (2019)
- Maudslay, Rowan Hall, "A Tale of a Probe and a Parser", arXiv preprint arXiv:2005.01641, ~ (2020)
- Pimentel, Tiago, "Pareto Probing: Trading OffAccuracy for Complexity", arXiv preprint arXiv:2010.02180, ~ (2020)
- Pimentel, Tiago, "Information-Theoretic Probing for Linguistic Structure", arXiv preprint arXiv:2004.03061, ~ (2020)
- Raganato, Alessandro, "An Analysis of Encoder Representations in Transformer-Based Machine Translation", Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (2018), The Association for Computational Linguistics
- Ravichander, Abhilasha, "Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?", arXiv preprint arXiv:2005.00719, ~ (2021)
- Vanmassenhove, Eva, "Investigating’aspect’in Nmt and Smt: Translating the English Simple Past and Present Perfect", Computational Linguistics in the Netherlands Journal (CLIN), 7, pp.109~128 (2017)
- Voita, Elena, "Information-Theoretic Probing with Minimum Description Length", arXiv preprint arXiv:2003.12298, ~ (2020)
- Zhou, Y., "DIRECTPROBE: Studying Representations without Classifiers", arXiv preprint arXiv : 2104. 05904, ~ (2021)