Year: 2022
Type: dissertation
Abstract: Recently, the performance of pre-trained language models has improved significantly, and their use in the real world is rapidly increasing.It is known that pre-trained language models acquire from their training data not only linguistic knowledge but also common sense and knowledge about real-world entities, which can then be used for various tasks.On the other hand, the output of language models often generates fluent sentences that contain erroneous content, which increases risks such as spreading misinformation and inducing errors in decision-making.For this reason, a mechanism is urgently needed to assess the knowledge stored in language models and to detect their potentially erroneous output.This study proposes an evaluation framework of the knowledge stored in language models.Our work is built on a benchmark for language model evaluation, the LAMA probe, which employs cloze tasks to assess the amount of knowledge stored in language models.The LAMA probe evaluation has issues, such as overestimating the models' knowledge due to lucky guesses caused by biases in the model predictions and not considering the discriminability of correct and erroneous answers.To address these issues, we introduce the selective prediction framework to the LAMA probe.Selective prediction assumes a system that calculates a confidence score for each output of a language model and evaluates not only the number of correct answers the model can make but also whether the system can detect when there is a high chance of incorrect output.In the first study, we evaluated multiple masked language models using the LAMA probe under the selective prediction setting.Experiments suggest that selective prediction-based evaluation can reduce the effect of overestimation of models iv knowledge due to biases in model prediction and evaluation data compared to the conventional evaluation based on prediction accuracy.In the second study, we focused on whether the performance of confidence estimation can be improved by using more information and designed confidence measures based on training data used during language model pre-training.The experiments using the BERT model showed that the performance of confidence estimation can be improved by combining the confidence measures using training data with those not using training data.
Authors: 吉川, 和, Yoshikawa, Hiyori, 岡崎, 直観, Okazaki, Naoaki
Language: Japanese
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7144816587