What do tokens know about their characters and how do they know it?
Pre-trained language models (PLMs) that use subword tokenization schemes can succeed at a variety of language tasks that require character-level information, despite lacking explicit access to the character composition of tokens. Here, studying a range of models (e.g., GPT-J, BERT, RoBERTa, GloVe), we probe what word pieces encode about character-level information by training classifier to predict the presence or absence of a particular alphabetical character in an English-language token, based on its embedding (e.g., probing whether the model embedding for "cat" encodes that it contains the character "a"). We find that these models robustly encode character-level information and, in general, larger models perform better at the task. Through a series of experiments and analyses, we investigate the mechanisms through which PLMs acquire character information during training and argue that this knowledge is acquired through multiple phenomena, including a systematic relationship between particular characters and particular parts of speech, as well as natural variability in the tokenization of related strings.
Code (1)
Methods 이 논문이 사용한 방법론
Similar Papers 제목 키워드 기반
What do tokens know about their characters and how do they know it?
Pre-trained language models (PLMs) that use subword tokenization schemes can succeed at a variety of language tasks that require character-level information, despite lacking explicit access to the character composition o…
Understanding the Role of Input Token Characters in Language Models: How Does Information Loss Affect Performance?
Understanding how and what pre-trained language models (PLMs) learn about language is an open challenge in natural language processing. Previous work has focused on identifying whether they capture semantic and syntactic…
“Politeness, you simpleton!” retorted [MASK]: Masked prediction of literary characters
What is the best way to learn embeddings for entities, and what can be learned from them? We consider this question for the case of literary characters. We address the highly challenging task of guessing, from a sentence…
Entity EmbeddingsSentenceCUTE: Measuring LLMs' Understanding of Their Tokens
Large Language Models (LLMs) show remarkable performance on a wide variety of tasks. Most LLMs split text into multi-character tokens and process them as atomic units without direct access to individual characters. This …
Zipf's Law and the Frequency of Characters or Words of Oracles
The article discusses the frequency of characters of Oracle,concluding that the frequency and the rank of a word or character is fit to Zipf-Mandelboit Law or Zipf's law with three parameters,and figuring out the paramet…