home edit page issue tracker

This page pertains to UD version 2.

Tokenization and Word Segmentation

The UD annotation is based on a lexicalist view of syntax, which means that dependency relations hold between words. Hence, morphological features are encoded as properties of words and there is no attempt at segmenting words into morphemes. However, it is important to note that the basic units of annotation are syntactic words (not phonological or orthographic words), which means that we systematically want to split off clitics, as in Spanish dámelo = da me lo, and undo contractions, as in French au = à le. We refer to such cases as multiword tokens because a single orthographic token corresponds to multiple (syntactic) words. In exceptional cases, it may be necessary to go in the other direction, and combine several orthographic tokens into a single syntactic word. Starting from v2 of the UD guidelines, such multitoken words are allowed for a restricted class of phenomena, such as numerical expressions like 20 000 and abbreviations like e. g., as long as these phenomena are approved and clearly specified in the language-specific documentation. Note, however, that this technique should not be generalized to multiword expressions like in spite of and by and large (let alone to more flexible multiword expressions like compounds or particle verbs), which should instead be annotated using special dependency relations.

Since word segmentation in general is a non-trivial task in many languages, and since the usefulness of tools trained on treebank data ultimately depends on how well the word segmentation can be reproduced for new data, it is important to document the principles of word segmentation for each language. The nature of this documentation will vary from one language to the next, depending on properties of the language and the writing system. For languages where word segmentation can be performed by a simple script given white-space and punctuation, only the words need to be represented in the treebank. For languages not using white-space at all, such as Chinese and Japanese, a complex word segmentation algorithm has to be employed, but there is no need to represent the basic character sequence in the treebank since it is completely recoverable from the word representation. By contrast, in languages where the mapping between white-space delimited tokens and syntactic words is highly ambiguous, such as Arabic and Hebrew, we provide the option of including both tokens and words in the treebank using a two-level indexing scheme described in the CoNLL-U format section. The morphological and syntactic annotation is only defined at the word level, but a heuristic mapping to the token level can usually be provided.

Language-specific extensions to this documentation must describe how tokenization and word segmentation has been performed for each language (including references to standard tokenization schemes if any). In addition, it should answer the following questions: