We present a scheme with which a lexeme on Wikidata consisting of multiple parts may be annotated to denote syntactic dependencies among its parts. The scheme is su ciently general to accommodate many dependency grammar frameworks and can take advantage of Wikidata lexemes' structure to reduce redundancy in representation while still being exible enough for further quali cation. While we note some challenges in adjustments to the scheme for particular phenomena, we contend that adopting this scheme will aid syntactic parsing e orts in other general domains as well as text generation systems for the Abstract Wikipedia project.
The projects under the Wikimedia Foundation's umbrella have frequently been
used for various natural language processing tasks, including disambiguating
word senses [
The under-construction Abstract Wikipedia project [
Copyright © 2021 for this paper by its author. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
For these building blocks to be useful, some mappings from concepts to
lexemes must rst exist, which presently consist of synonym and translation
linkages between senses and correspondences between Wikidata items and senses.
A concept that in one language is representable with one word may need
multiple words in another, however; depending on the sort of multi-part expression
used, the syntactic information needed for adjustment of that expression in
different contexts may di er. The English verb `evade' has a correspondence with
the South American Spanish phrase `hacer el quite' [
Databases of multi-component expressions have previously been developed
for individual languages [
We thus propose here a compact representation of syntactic dependencies within the structures of Wikidata lexicographical data, generally applicable to di erent avors of dependency grammar, but here demonstrated with respect to Universal Dependencies (UD). We contend that the marking up of this information is useful even for modeling structures of multi-part elements that may be regarded in some languages as words, and that it permits lexemes with syntax represented this way to form parts of other lexemes which, as single units, take part in other dependency relations. We recognize too that modi cations to handle special syntactic cases may not necessarily be immediately acceptable to those annotating relevant lexemes. We nevertheless believe that the greater portion of what may be annotated of multi-part lexemes with this representation will not only make those lexemes usable in the syntactic parsing of other texts, but will also considerably ease the generation of text through the manipulation of underlying dependency graphs. 2
What follows is an outline of the proposed dependency representation within Wikidata lexemes. RDF predicates for Wikidata properties and quali ers1 are 1 https://www.mediawiki.org/wiki/Wikibase/Indexing/RDF Dump Format (section \Full list of pre xes") lists the RDF pre xes used herein. provided in monospace, as are Wikidata items, lexemes, and their forms and senses, which are left unpre xed. `series ordinal'
what go around come around (L333986) (L3006) (L333609) (L3210) (L333609)
1 2 3 4 5 what goes around comes around `object form' (L333986-F1) (L3006-F2) (L333609-F1) (L3210-F2) (L333609-F1) `object sense' (Lt3h3a3t9w86h-icSh1) to(Lm30o0v6e-aSw2a)y (L3a3tp3lv6aa0cre9ise-dS1) t(oL3a2p1p0r-oSa1ch) (L3a3tp3lv6aa0cre9ise-dS1) 2 4 2 0 4 rela`thioenadship' (Qre5l6a8t7iv0i2z2er6) (Q1sc9ul7ba0uj8esc5et32) (Qa1ldo2vc7ae2rt4bio4ina8l0) (Q17ro57o0t74) (Qa1ldo2vc7ae2rt4bio4ina8l0) mark csubj obl:lmod root obl:lmod Each component of the surface form of a multi-part lexeme is represented by a use of the `combines' property (p:P5238) linking to a lexeme for said component. The component's representation in the lexeme is speci ed with a quali er noting the form which is the object of the `combines' statement ('object form', pq:P5548) on each statement, and the position of each part using a `series ordinal' quali er (pq:P1545). For completeness, the statement may also note the sense of the component being used with `object sense' (pq:P5980); languagespeci c machinery for text generation may nd utility in di erent treatment of a constituent based on the meaning it expresses, although we do not further consider uses of that quali er here.
An example of these quali ers in action is shown in the rst three rows of Figure 1 for the proverb \what goes around comes around". Note that `series ordinal' values need not be numeric; a separate ordering to handle in xes, circum xes, and other non-sequential phenomena may well be desired for some languages.
Many dependency treebanks optionally store information about the part of speech of a word token and the grammatical features it bears in the context in which it appears. In a Wikidata lexeme, the former of these is a top-level feature (via wikibase:lexicalCategory), while the latter of these reside on the individual forms of lexemes (via wikibase:grammaticalFeature). As a result this information for the components of a multi-part lexeme generally does not need to be reproduced on the multi-part lexeme itself; they may be programmatically retrieved from the parts themselves based on the main `combines' values and `object form' quali ers respectively. 2.2
A dependency relationship may be thought of as a directed edge between a dependent and a head, so that specifying both ends of the relationship and the type of relationship su ces to de ne the dependency, and so that the resulting set of dependencies for a multi-component lexeme resembles a tree. To de ne dependency relationships between parts of a lexeme within Wikidata's existing structure, a set of quali er properties are instead applied to the `combines' statement for a relationship's dependent.
The rst of these quali ers, `head position' (pq:P9764), indicates the `series ordinal' of the head of a dependency relationship. The second of these, `head relationship' (pq:P9763), indicates the type of said relationship. Both of these quali ers, having been created in late July 2021, are as yet little used beyond additions of these by the author on existing lexemes, and user scripts to better facilitate their addition to new lexemes on Wikidata are yet to be written. Documentation of appropriate `head relationship' values is slowly being developed, however2.
what
goes subjective clause around location adverbial comes root around location adverbial
The relationships in the proverb `what goes around comes around' are de ned in the fourth and fth rows of Figure 1, with UD equivalents below them. A potential diagram generated with `series ordinal' and the two new quali ers is 2 With respect to UD, see https://www.wikidata.org/wiki/
Wikidata:Lexicographical data/Universal Dependencies . shown in Figure 2. The link from \what" to \goes" is de ned by the edge from `series ordinal' value `1' to `series ordinal' value `2', where this latter value is speci ed via the `head position' value on \what". Most other relationships are marked up similarly. As a convention, the `head position' of the root part is `0' and the `head relationship' is the item for `root'.
The two quali ers, as two separate properties de ning parts of the same relationship, were proposed because Wikidata statements and quali ers cannot have statements or quali ers themselves as values, much less annotate connections between them, without major modi cations to Wikidata's Wikibase software and its Query Service. While such connections might also be stored in an entirely separate Wikibase instance, in the absence of an implicit federation system between Wikibase triple stores, querying such connections becomes more costly for end users and downstream applications than necessary. 3
With this dependency representation, a large number of syntactic phenomena in multi-part lexemes can be faithfully modeled. There nevertheless remain some situations that an application of this representation by itself would not satisfactorily handle, and for which the introduction of certain changes might be controversial. We outline some of these challenging aspects here. 3.1
In some parallel constructions, one may decide to omit in later parts of a phrase portions common to earlier parts of the same phrase. The sentence \I wrote the book, he wrote the story, and she wrote the poem" may be shortened without loss of understanding by omitting the latter two occurrences of `wrote'. The resulting subgraphs of the latter phrases may be marked di erently as well; in UD the special orphan relation would link `he' and `the story', as well as `she' and `the poem', in the shortened version of the sentence.
I write the book and he no value the story `series ordinal' 1 2 3 4 5 6 7 8 9 `head position' 2 0 4 2 7 7 2 9 7 `head relationship' subject root dmeitneerr- odbirjeecctt contijounnc- subject conjunct dmeitneerr- odbirjeecctt UD equivalent of `head relationship' nsubj root det obj cc nsubj conj det obj
UD alternatively de nes, in its speci cations of entirely optional `enhanced dependencies', the concept of `orphan nodes' which represent elided words and can take part in those dependency relationships originally substituted with orphan. To mimic this concept, elided words might be similarly indicated by setting the value of the `combines' statement to the special Wikibase value \no value" and otherwise marking up relationships with respect to that elided word. (In the interest of preserving some contextual information, the `object form' quali er might still refer to the form the elided part would take on were it still present in the lexeme.) An example of an elided word's use is shown in Figure 3.
The insertion of extra `nonexistent' components to a multi-part lexeme may appear to some to be a repurposing of the `combines' property, given that these components do not appear in the surface form of that lexeme and that any counts of components of such a lexeme may appear in ated. We might alternately contend that this is merely a syntactic analogue of when grammatical features added to a lexeme form may not necessarily change the form despite those features' importance (as, for example, when further in ecting a Hindustani adjective that is already in ected for female gender), and that in counting components ltering out \no value" statements is a small addition to a query. 3.2
`series ordinal' `head position'
She 1 4 `hUeDadeqreuliavtaiolennsthoipf' nsubj `object has role' and 2 3 cc
I 3 1 visited Japan 4 5 0 4 and 6 7 conj subject root obj cc `head relationship' subject contijounnc- conjunct root object contijounnc- conjunct conj object
A clause may contain multiple instances of the same type of element, such as two subjects or two objects, to which other components in the clause apply equally. \She and I visited Japan and Korea" contains two subjects, each of which applies to two objects, and \They washed and combed the dog" contains two predicates. The graph of the rst phrase may directly connect `she' to the predicate and connect `I' to `she' (the approach taken by UD), and the graph of the second phrase may alternately group the actions `washed' and `combed' into an umbrella predicate to which `they' and `the dog' directly attach. Either of these approaches, however, adds distance between components that we might regard as being close syntactically.
UD's `enhanced dependencies` also allow multiple relationships to share the same dependent, so that in the rst example the word `I' points both to `she' (as a conjunct) and `visited' (as a subject), turning the resulting syntactic tree into a general directed (possibly cyclic) graph. Since multiple `head position' and `head relationship' quali ers on a single `combines' statement cannot be separated to refer to di erent dependency relationships, one possible solution is to mark out the semantic roles more explicitly, and in some cases redundantly, using the `object has role' (pq:P3831) or `has quality' (pq:P1552) quali ers on the `combines' statements of conjuncts. An example using the rst phrase and the `object has role' quali er is shown in Figure 4.
The extra marking of syntactic roles may appear to some to simply duplicate information, especially since a conjunct of a particular part very frequently has the same role as that part. We might instead say that the explicit marking of the relationships expressed by conjuncts allows them to be queried more readily, so that traversing conjunct paths becomes unnecessary. 3.3
In many South Asian languages, it is a frequently productive process for a word and a nonce word rhyming with it, when taken together, to refer to something related to said word. In Bengali the noun \ranna" is the act of cooking, while \rannabanna" refers to `cooking and related activities'; in Hindi \samna" is `to encounter', while \amna samna" is the act of encountering. The extra echo components that result from this productive process do not themselves have any meaning on their own, however, and so creating lexemes for those components would not be appropriate.
ranna some value `series ordinal' 1 2 `head position' 0 1 `head relationship' root echo word UD equivalent of `head relationship' root compound:redup `stated as' `banna'
Just as Wikidata statements and quali ers can specify that \no value" exists as an object of their relationships, so too can they specify that the special Wikibase value \some value" exists; the implication desired here is that a component is present but that no separate Wikidata lexeme exists for it. Since, unlike the elided word case, something is still being realized in the surface form, the quali er `stated as' (pq:P1932) on the `combines' statement can provide that form. An example of \some value" with a `stated as' quali er is shown in Figure 5.
The addition of statements with \some value" may appear to some particularly in exible, given that individual lexeme forms can have pronunciation and grammatical information and be tied to usage examples, and that these rump `combines' statements might become particularly unwieldy if that information were admitted there. At the same time, for languages with multiple spelling conventions (where otherwise these might be handled with separate representations using di erent language codes), the use of the datatype that `stated as' expects, which does not allow attaching a language code, might lead to confusion when selecting which of a number of alternatives to use. While we cannot quite counter the latter beyond suggesting a new quali er with a new datatype exist, for the former we might contend instead that the contribution that these echo words have outside the scope of the lexeme is especially minimal and, if one so desires, can be derived from the word it echoes without considerable di culty. 4
Wikidata lexicographical data has the potential to support storing information about the syntactic structure of multi-part lexemes, and we have provided a scheme using existing Wikidata quali ers on an existing Wikidata property for this structured information, noting as well ways in which said scheme could be improved and potential issues in pursuing those ways. Although the examples provided herein used Universal Dependencies as a basis, this by no means is limited to that particular framework; conversion between frameworks to accommodate di erent use cases is just as possible with Wikidata lexemes as without them. We envision the possibility of full treebanks being constructed using Wikibase and some variant of this scheme as a starting point, as well as the introduction of new structured datatypes to better handle the sorts of connections and speci cations that have been handled by this scheme{all in addition to the downstream task improvements that have the potential to bene t from resources using this scheme.