Google

A significant drawback of Masked Language Modeling is its reliance on a special $$[\mathrm{MASK}]$$ token during the training phase. Because this artificial token is not present in natural text during testing or real-world inference, it creates a discrepancy between how the model is trained and how it operates in practice at test time.

Drawback of Masked Language Modeling: The [MASK] Token Discrepancy

A language model is pre-trained using an objective where 15% of the input tokens in a sentence are replaced with a special, artificial token. The model's goal is to predict the original tokens that were replaced. After this pre-training phase, the model is to be fine-tuned for two different downstream applications:

1.  **Text Summarization:** The model takes a long document as input and must generate a short, coherent summary.
2.  **Sentiment Classification:** The model takes a movie review as input and must classify it as 'positive', 'negative', or 'neutral'.

Analyze how the mismatch between the pre-training objective (which uses the special token) and the fine-tuning data (which contains only natural language) could uniquely impact the model's performance on each of these two tasks. Which task do you predict would be more negatively affected by this discrepancy, and why?

Impact of Pre-training/Fine-tuning Mismatch on Downstream Tasks

An engineer is developing a language model. First, they train it on a large corpus of text where 15% of the words in each sentence are replaced with a special `[BLANK]` symbol; the model's objective is to predict these original words. After this initial training, they adapt the model for a new task: classifying customer reviews as 'positive' or 'negative'. This new task uses complete, unaltered customer reviews. The engineer notices that the model's performance on the classification task is lower than anticipated. Based on this information, identify and explain the fundamental mismatch between the initial training phase and the final task that is likely causing this issue.

Troubleshooting a Pre-trained Model's Performance

Permuted Language Modeling (PLM) is a pre-training approach designed to resolve specific issues found in Masked Language Modeling, such as the mismatch between training and inference and the independence assumption among masked tokens. While it is a sequential prediction task, the actual order of tokens in the original text remains completely unchanged. Instead, the model is trained to predict the tokens sequentially according to an arbitrarily determined, permuted order.

Permuted Language Modeling (PLM)

A language model is first trained on a large text corpus where some words in each sentence are replaced with a special `[MASK]` symbol, and the model's goal is to predict the original words. Subsequently, this pre-trained model is adapted for a specific task, such as sentiment analysis, using a new dataset of complete, un-masked sentences. Which of the following statements best analyzes the primary architectural conflict that arises between these two phases?

Learn Before

Related