Overview

In this lesson, we'll look at a definition of token and see some examples in English.

Outcomes

After completing this lesson, you should be able to …

• define token
• describe the distinction between token and type
• describe patterns for delimiting tokens in one or more natural languages

Before starting

• Read the following section of Manning et al.'s Introduction to Information Retrieval:
• https://nlp.stanford.edu/IR-book/html/htmledition/tokenization-1.html

Background

What is a token?

In 2.2.1 of Introduction to Information Retrieval, Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze offer the following definition:

A token is an instance of a sequence of characters in some particular document that are grouped together as a useful semantic unit for processing.

Notice that definition doesn't use word. It aligns with the notion of lexical item. By this definition, a token could be a single word, a sequence of words, or just a part (ex. morpheme) of a word.

Let's look at an example:

I like Valeria's shoes. $\rightarrow$ ["I", "like", "Valeria", "'s", "shoes", "."]

Given our definition of token, we can tokenize or split the sentence I like Valeria's shoes. in a number of different ways. Here are two possibilities:

["I", "like", "Valeria", "'s", "shoes", "."]
["I", "like", "Valeria's", "shoes", "."]

You might be wondering which representation is "correct" or preferable. As we'll see in later lessons, different problems require different solutions.

There are, however, some general considerations:

• compatibility with existing tools: statistical models for part of speech tagging, parsing, etc. are trained to expect a certain tokenization strategy. If you're using an off-the-shelf tool (ex. POS tagger) to process data you've tokenized with a strategy unfamiliar to the model¹, performance is likely to be poor.

• extending or supplementing existing datasets: if you want to supplement some existing dataset, you probably want your tokenization strategy to match.

• size: different tokenization strategies can increase or decrease the number of unique tokens encountered. When we learn about classification tasks, this will become important.

In a very large corpus, which of the two tokenization strategies presented above will results in fewer unique tokens (i.e., a more compact set of tokens)?

Types vs tokens

If tokens are roughly lexical items, then we can think of a token as an individual occurrence of a lexical item. For example, the word "the" may occur many times in a document.²

Each of those occurrences is a token.

We use type to reference the category represented by all of the occurrences of a single lexical item (ex. occurrences of "the" treated as a category).

• Identical tokens constitute a single type.
• The set of types or vocabulary of some text is its unique set of tokens.
• The type vs token distinction is that of unique vs non-unique strings.

Let's look at an example:

"his friends are also his enemies"

tokens = ["his", "friends", "are", "also", "his", "enemies"]
types  = {"his", "friends", "are", "also", "enemies"}

Tokens in English

If you examine tokenized English datasets, such as those made available by the Universal Dependencies, you'll notice that …

• tokens don't contain whitespace
• possessive markers are split into separate tokens
• ex. "Iggy's" $\rightarrow$ ["Iggy", "'s"]
• clitic negations are split into separate tokens
• ex. "shouldn't" $\rightarrow$ ["should", "n't"]
• acronyms are not split
• ex. "D.C." -> "D.C."

For further guidance, see https://universaldependencies.org/en/index.html.

Next steps

Practice

• Give an example of a token that is not a word
• Consider a non-English language that you've studied. Alternatively, search for a sample of writing in non-English language.
• Does the language use whitespace to delimit word boundaries? Are there other features that appear useful for delimiting words?
• Give one reason why splitting "couldn't" as ["could", "n't"] is preferable to ["couldn", "'t"]

Footnotes

1. "unfamiliar" here means data that does not match or resemble the model's training data (ex. the two datasets were processed using different tokenization strategies). ↩
2. To understand why, review the parts of speech lesson. ↩