🌐 fixed: fixed multiword expression
The fixed multiword expression relation (fixed) is the Universal Dependencies label for grammaticalized multiword expressions (MWEs): word sequences that have turned from free phrases into a single function word or adverbial, such as because of or as well as. The first word heads the expression and every other word attaches to it.
What it is
In Universal Dependencies (UD), most words are joined by relations that state their grammatical role toward a content-word head. Function words such as prepositions and auxiliaries depend on the content word they serve (Marneffe et al. 2021).
Some strings resist this analysis. English in spite of, French parce que and Portuguese apesar de behave like one preposition or conjunction. UD treats them as single units because their parts no longer combine productively. For such MWEs UD v2 offers the label fixed. It is one of the 37 universal relation types, and languages may extend that inventory only through subtypes (Nivre et al. 2020).
The convention, as presented in introductions to UD principles (Zeman 2021), is flat:
- The first word of the expression is the head.
- Every following word attaches to that first word with fixed. They do not form a chain.
- The whole unit then attaches to the rest of the sentence as its function demands, for example as case, mark, cc or advmod.
The words inside normally take no dependents of their own. The universal guidelines define no subtype of fixed.
The label replaces the v1 relation mwe, which belonged to the 40-type inventory of the first release (Nivre et al. 2016). That inventory itself grew out of the universal Stanford dependencies (Marneffe et al. 2014). The change also narrowed the scope: v2 reserves fixed for grammaticalized expressions, and other strings once labelled mwe receive flat or compound. In the English enhanced++ representation, mwe served for multiword prepositions such as in front of, yielding a single edge nmod:in_front_of (Schuster & Manning 2016).
Neighbouring relations
| Relation | Used for | Internal structure |
|---|---|---|
| fixed | grammaticalized function-word expressions | frozen, head = first word |
| flat | headless names, foreign phrases (optionally flat:foreign) | no syntactic head, first word chosen by convention |
| compound | lexical compounds such as phone book | real head–modifier structure |
Idioms such as kick the bucket are not fixed. They keep ordinary relations, because their internal syntax is regular (kicked the bucket inflects normally).
Examples
Arrows point from the head to the dependent:
- She left because of the rain: because → of (fixed). The unit attaches to rain as case.
- Bring bread as well as cheese: as → well (fixed), as → as (fixed). The unit attaches to cheese as cc.
- Il est parti parce qu'il pleuvait: parce → qu' (fixed). The unit attaches to pleuvait as mark.
A test for borderline strings is whether any inner word could be modified or replaced. If it can, the string is probably not fixed.
Across languages
How often the relation is used depends on how much a language has grammaticalized, and on annotation tradition. Common patterns include:
- English: a preposition, adjective or adverb heading a following preposition (out of, such as, instead of), adverb–adverb pairs such as as well, and preposition–noun units such as of course.
- French: an adverb heading a subordinating conjunction, as in bien que and alors que, along with preposition–noun and preposition–adverb pairs.
Such differences reflect both the languages themselves and where each team draws the line between fixed and compositional analysis.
In treebanks and research
Because fixed depends on lexical judgement, it is a known source of inconsistency across UD treebanks. Consistency checks of the corpora look for identical strings annotated differently in different sentences (Marneffe et al. 2017).
To make annotation systematic, Portuguese UD work gathered candidate fixed expressions from three separate sources before revising the treebank (Souza & Freitas 2023). This ties the relation to idioms, to multiword expressions in general and to lexicography.
In dependency parsing, fixed edges are short and generally predictable from the word forms. The harder decision is likely whether a string such as as well is frozen in a given sentence at all. UD served as the data for the CoNLL 2017 and 2018 shared tasks on parsing raw text (Marneffe et al. 2021). Such decisions therefore count toward attachment scores, although fixed edges form only a small part of any treebank.
Downstream, treating a fixed unit as one function word suits tasks that read relations between content words, as with the nmod:in_front_of edge above (Schuster & Manning 2016). Information extraction over shortest dependency paths is one such task (Wu & Weld 2010).
What fixed joins: the most frequent head tag → dependent tag
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: ADP → ADP | 368 |
| English: ADJ → ADP | 365 |
| English: ADV → ADP | 359 |
| English: ADV → ADV | 292 |
| English: ADP → NOUN | 210 |
| French: ADV → SCONJ | 2,116 |
| French: ADP → NOUN | 1,238 |
| French: ADP → ADV | 806 |
| French: ADP → DET | 482 |
| French: ADP → SCONJ | 416 |
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-fixed/