🌐 goeswith: goes with
goeswith is a Universal Dependencies relation that reunites the pieces of one word which a text wrongly splits with spaces, as in "with out" for "without". The first piece is the head, and every later piece attaches to it.
What it is
In Universal Dependencies (UD), every token of a sentence is attached to another word by a typed relation, forming a dependency tree. The basic unit of this tree is the syntactic word, following the principle of lexical integrity (Marneffe et al. 2021). Real text does not always respect word boundaries. Writers type "some thing" for "something", or a scanned page breaks a name in two. Annotators may not silently fix the text, because UD requires that the original tokens stay recoverable (Nivre et al. 2016). The goeswith relation records that two or more space-separated tokens are really one word.
The convention is strictly head-initial. The first fragment is the head. It carries the dependency that the whole word would have in the sentence, and it usually carries the lemma, part of speech and features of the reconstructed word. Every later fragment attaches directly to the first with goeswith, not in a chain. Non-first fragments normally get the universal tag X. They may also be marked with the morphological feature Typo, which was added in UD v2 (Nivre et al. 2020). goeswith is one of the 37 universal relation types of v2 (Nivre et al. 2020). No widely used subtypes of the form goeswith:subtype exist.
Neighbouring relations
| Relation | What it joins | Example |
|---|---|---|
| goeswith | pieces of one misspelt word | with ← out |
| flat | parts of a multiword name | Marie ← Curie |
| fixed | a fixed grammatical multiword expression | because ← of |
| compound | two real words forming a compound | ice ← cream |
| reparandum | a disfluent word that is then corrected | go → went |
The test is simple. If the tokens would be a single word in correct spelling, the relation is goeswith. If they are separate words that function together, one of the other relations applies.
Examples
In each example, the arrow points from the head to the dependent.
- English: "I did it with out help." The word with is the head and takes the relation case to help. The fragment out attaches as with → out (goeswith).
- English: "She moved to Bar celona." Bar is a proper-noun head, and celona attaches as Bar → celona (goeswith), tagged X.
- French: "Il part par ce qu'il pleut" ("He leaves because it rains"), where "parce que" has been split. The head is par, with par → ce (goeswith). The contraction qu' keeps its own analysis as part of the conjunction.
In the first two examples, the head receives the lemma of the whole word: "without" and "Barcelona".
Across languages
How often goeswith appears depends mostly on the genre of the text, not on grammar. Edited newswire seldom needs it. Web forums, reviews, learner essays and OCR output need it more.
In English UD 2.18, goeswith is rare overall. Its most frequent pairs are a proper-noun head with an X fragment, followed by nouns, adverbs, verbs and adjectives. This pattern reflects split names and split compounds in web text. In French it is even rarer, and noun heads clearly dominate.
Writing systems also matter. UD v2 allows spaces inside words outright in scripts that separate syllables with spaces, such as Vietnamese (Nivre et al. 2020). For those languages, regular spaces never call for goeswith. The relation is reserved for genuine errors, so treebanks of morphologically rich or agglutinative languages use it only where their sources are noisy.
In treebanks and research
goeswith belongs to the "loose joining" corner of the scheme. It is a practical device that keeps the content-word-headed tree well formed despite messy input (Marneffe et al. 2021). Tutorials on UD treat it alongside the other mechanisms for segmentation and tokenization (Zeman 2021). Studies of annotation consistency across UD corpora show that such low-frequency labels are an easy place for treebanks to diverge (Marneffe et al. 2017).
The enhanced representation adds edges on top of the basic tree but keeps the basic tree intact (Schuster & Manning 2016). As a result, goeswith edges pass unchanged into enhanced graphs.
For dependency parsing, goeswith is a challenge of rarity and of tokenization. The CoNLL shared tasks of 2017 and 2018 asked systems to parse raw text across many languages (Marneffe et al. 2021). In that setting, the parser must notice on its own that two whitespace tokens form one word. Because the relation is sparse, parsers often confuse it with flat or compound, and per-label scores for it are unstable. The X tag on the fragments also interacts with part-of-speech tagging.
Work on noisy and learner text, such as grammatical error correction, can use goeswith annotations to locate segmentation errors. Researchers comparing English and French treebanks should read differences in goeswith frequency as differences in genre and editing, not in syntax. The UD framework itself grew out of universal Stanford dependencies and other earlier initiatives (Nivre et al. 2016).
What goeswith joins: the most frequent head tag → dependent tag
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: PROPN → X | 84 |
| English: NOUN → X | 58 |
| English: ADV → X | 31 |
| English: VERB → X | 25 |
| English: ADJ → X | 24 |
| French: NOUN → X | 29 |
| French: ADV → X | 6 |
| French: ADJ → X | 5 |
| French: PRON → X | 3 |
| French: NUM → X | 1 |
Key papers
Related concepts
- Universal Dependencies — Universal Dependencies is a framework and an open collection of treebanks that annotates the grammar of many languages with one shared…
- Dependency parsing — Dependency parsing is the task of recovering a sentence's syntactic structure as directed, labelled links between its words. Each word is…
More papers
Topics: 🌐 Universal Dependencies · 🌳 Dependency parsing
Shelves: 🗣️ Language and Linguistics
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/ud-goeswith/