liu.seSearch for publications in DiVA
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Understanding Large Language Models: Towards Rigorous and Targeted Interpretability Using Probing Classifiers and Self-Rationalisation
Linköping University, Department of Computer and Information Science, Artificial Intelligence and Integrated Computer Systems. Linköping University, Faculty of Science & Engineering.
2024 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Large language models (LLMs) have become the base of many natural language processing (NLP) systems due to their performance and easy adaptability to various tasks. However, much about their inner workings is still unknown. LLMs have many millions or billions of parameters, and large parts of their training happen in a self-supervised fashion: They simply learn to predict the next word, or missing words, in a sequence. This is effective for picking up a wide range of linguistic, factual and relational information, but it implies that it is not trivial what exactly is learned, and how it is represented within the LLM. 

In this thesis, I present our work on methods contributing to better understanding LLMs. The work can be grouped into two approaches. The first lies within the field of interpretability, which is concerned with understanding the internal workings of the LLMs. Specifically, we analyse and refine a tool called probing classifiers that inspects the intermediate representations of LLMs, focusing on what roles the various layers of the neural model play. This helps us to get a global understanding of how information is structured in the model. I present our work on assessing and improving the probing methodologies. We developed a framework to clarify the limitations of past methods, showing that all common controls are insufficient. Based on this, we proposed more restrictive probing setups by creating artificial distribution shifts. We developed new metrics for the evaluation of probing classifiers that move the focus from the overall information that the layer contains to differences in information content across the LLM. 

The second approach is concerned with explainability, specifically with self-rationalising models that generate free-text explanations along with their predictions. This is an instance of local understandability: We obtain justifications for individual predictions. In this setup, however, the generation of the explanations is just as opaque as the generation of the predictions. Therefore, our work in this field focuses on better understanding the properties of the generated explanations. We evaluate the downstream performance of a classifier with explanations generated by different model pipelines and compare it to human ratings of the explanations. Our results indicate that the properties that increase the downstream performance differ from those that humans appreciate when evaluating an explanation. Finally, we annotate explanations generated by an LLM for properties that human explanations typically have and discuss the effects those properties have on different user groups. 

While a detailed understanding of the inner workings of LLMs is still unfeasible, I argue that the techniques and analyses presented in this work can help to better understand LLMs, the linguistic knowledge they encode and their decision-making process. Together with knowledge about the models’ architecture, training data and training objective, such techniques can help us develop a robust high-level understanding of LLMs that can guide decisions on their deployment and potential improvements. 

Place, publisher, year, edition, pages
Linköping: Linköping University Electronic Press, 2024. , p. 81
Series
Linköping Studies in Science and Technology. Dissertations, ISSN 0345-7524 ; 2364
National Category
Computer and Information Sciences
Identifiers
URN: urn:nbn:se:liu:diva-201985DOI: 10.3384/9789180754712ISBN: 9789180754705 (print)ISBN: 9789180754712 (electronic)OAI: oai:DiVA.org:liu-201985DiVA, id: diva2:1848043
Public defence
2024-04-18, Ada Lovelace, B-building, Campus Valla, Linköping, 14:00 (English)
Opponent
Supervisors
Available from: 2024-04-02 Created: 2024-04-02 Last updated: 2024-04-02Bibliographically approved
List of papers
1. Classifier Probes May Just Learn from Linear Context Features
Open this publication in new window or tab >>Classifier Probes May Just Learn from Linear Context Features
2020 (English)In: Proceedings of the 28th International Conference on Computational Linguistics, 2020, Vol. 28, p. 5136-5146, article id 450Conference paper, Published paper (Refereed)
Abstract [en]

Classifiers trained on auxiliary probing tasks are a popular tool to analyze the representations learned by neural sentence encoders such as BERT and ELMo. While many authors are aware of the difficulty to distinguish between “extracting the linguistic structure encoded in the representations” and “learning the probing task,” the validity of probing methods calls for further research. Using a neighboring word identity prediction task, we show that the token embeddings learned by neural sentence encoders contain a significant amount of information about the exact linear context of the token, and hypothesize that, with such information, learning standard probing tasks may be feasible even without additional linguistic structure. We develop this hypothesis into a framework in which analysis efforts can be scrutinized and argue that, with current models and baselines, conclusions that representations contain linguistic structure are not well-founded. Current probing methodology, such as restricting the classifier’s expressiveness or using strong baselines, can help to better estimate the complexity of learning, but not build a foundation for speculations about the nature of the linguistic structure encoded in the learned representations.

Keywords
Natural Language Processing, Machine Learning, Neural Language Representations
National Category
Language Technology (Computational Linguistics) Computer Sciences
Identifiers
urn:nbn:se:liu:diva-175384 (URN)10.18653/v1/2020.coling-main.450 (DOI)
Conference
International Conference on Computational Linguistics (COLING), Barcelona, Spain (Online), December 8–13, 2020
Available from: 2021-04-30 Created: 2021-04-30 Last updated: 2024-04-02Bibliographically approved
2. Test Harder Than You Train: Probing with Extrapolation Splits
Open this publication in new window or tab >>Test Harder Than You Train: Probing with Extrapolation Splits
2021 (English)In: Proceedings of the Fourth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP / [ed] Jasmijn Bastings, Yonatan Belinkov, Emmanuel Dupoux, Mario Giulianelli, Dieuwke Hupkes, Yuval Pinter, Hassan Sajjad, Punta Cana, Dominican Republic, 2021, Vol. 5, p. 15-25, article id 2Conference paper, Published paper (Refereed)
Abstract [en]

Previous work on probing word representations for linguistic knowledge has focused on interpolation tasks. In this paper, we instead analyse probes in an extrapolation setting, where the inputs at test time are deliberately chosen to be ‘harder’ than the training examples. We argue that such an analysis can shed further light on the open question whether probes actually decode linguistic knowledge, or merely learn the diagnostic task from shallow features. To quantify the hardness of an example, we consider scoring functions based on linguistic, statistical, and learning-related criteria, all of which are applicable to a broad range of NLP tasks. We discuss the relative merits of these criteria in the context of two syntactic probing tasks, part-of-speech tagging and syntactic dependency labelling. From our theoretical and experimental analysis, we conclude that distance-based and hard statistical criteria show the clearest differences between interpolation and extrapolation settings, while at the same time being transparent, intuitive, and easy to control.

Place, publisher, year, edition, pages
Punta Cana, Dominican Republic: , 2021
Keywords
Natural Language Processing, Neural Language Models, Interpretability, Probing, BERT, Extrapolation
National Category
Language Technology (Computational Linguistics) Computer Sciences
Identifiers
urn:nbn:se:liu:diva-182166 (URN)10.18653/v1/2021.blackboxnlp-1.2 (DOI)
Conference
BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, November 11, 2021
Available from: 2022-01-10 Created: 2022-01-10 Last updated: 2024-04-02Bibliographically approved
3. Where Does Linguistic Information Emerge in Neural Language Models?: Measuring Gains and Contributions across Layers
Open this publication in new window or tab >>Where Does Linguistic Information Emerge in Neural Language Models?: Measuring Gains and Contributions across Layers
2022 (English)In: Proceedings of the 29th International Conference on Computational Linguistics / [ed] Nicoletta Calzolari, Chu-Ren Huang, Hansaem Kim, James Pustejovsky, Leo Wanner, Key-Sun Choi, Pum-Mo Ryu, Hsin-Hsi Chen, Lucia Donatelli, Heng Ji, Sadao Kurohashi, Patrizia Paggio, Nianwen Xue, Seokhwan Kim, Younggyun Hahm, Zhong He, Tony Kyungil Lee, Enrico Santus, Francis Bond, Seung-Hoon Na, 2022, p. 4664-4676, article id 1.413Conference paper, Published paper (Refereed)
Abstract [en]

Probing studies have extensively explored where in neural language models linguistic information is located. The standard approach to interpreting the results of a probing classifier is to focus on the layers whose representations give the highest performance on the probing task. We propose an alternative method that asks where the task-relevant information emerges in the model. Our framework consists of a family of metrics that explicitly model local information gain relative to the previous layer and each layer’s contribution to the model’s overall performance. We apply the new metrics to two pairs of syntactic probing tasks with different degrees of complexity and find that the metrics confirm the expected ordering only for one of the pairs. Our local metrics show a massive dominance of the first layers, indicating that the features that contribute the most to our probing tasks are not as high-level as global metrics suggest.

Keywords
NLP, AI, Language Technology, Computational Linguistics, Machine Learning
National Category
Language Technology (Computational Linguistics)
Identifiers
urn:nbn:se:liu:diva-191000 (URN)
Conference
COLING, October 12–17, 2022
Available from: 2023-01-12 Created: 2023-01-12 Last updated: 2024-04-02Bibliographically approved
4. Human Ratings Do Not Reflect Downstream Utility: A Study of Free-Text Explanations for Model Predictions
Open this publication in new window or tab >>Human Ratings Do Not Reflect Downstream Utility: A Study of Free-Text Explanations for Model Predictions
2022 (English)In: Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, 2022, Vol. 5, p. 164-177, article id 2022.blackboxnlp-1.14Conference paper, Published paper (Refereed)
Abstract [en]

Models able to generate free-text rationales that explain their output have been proposed as an important step towards interpretable NLP for “reasoning” tasks such as natural language inference and commonsense question answering. However, the relative merits of different architectures and types of rationales are not well understood and hard to measure. In this paper, we contribute two insights to this line of research: First, we find that models trained on gold explanations learn to rely on these but, in the case of the more challenging question answering data set we use, fail when given generated explanations at test time. However, additional fine-tuning on generated explanations teaches the model to distinguish between reliable and unreliable information in explanations. Second, we compare explanations by a generation-only model to those generated by a self-rationalizing model and find that, while the former score higher in terms of validity, factual correctness, and similarity to gold explanations, they are not more useful for downstream classification. We observe that the self-rationalizing model is prone to hallucination, which is punished by most metrics but may add useful context for the classification step.

Keywords
Large Language Models, Neural Networks, Transformers, Interpretability, Explainability
National Category
Language Technology (Computational Linguistics) Computer Sciences
Identifiers
urn:nbn:se:liu:diva-195615 (URN)
Conference
BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, December 8, 2022
Available from: 2023-06-22 Created: 2023-06-22 Last updated: 2024-04-02Bibliographically approved

Open Access in DiVA

fulltext(8083 kB)168 downloads
File information
File name FULLTEXT01.pdfFile size 8083 kBChecksum SHA-512
5e0c8fa82a5639ea77ce75ae076ce5d6f953feda784ce3436ab242a895582a010a25d23ca3f38a8ab9921baceddbbd800f99a4c50fa83dc36548ca1819ca8b82
Type fulltextMimetype application/pdf
Order online >>

Other links

Publisher's full text

Authority records

Kunz, Jenny

Search in DiVA

By author/editor
Kunz, Jenny
By organisation
Artificial Intelligence and Integrated Computer SystemsFaculty of Science & Engineering
Computer and Information Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 168 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

doi
isbn
urn-nbn

Altmetric score

doi
isbn
urn-nbn
Total: 1040 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • oxford
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf