Multilingual Verb Valence Lexicon
--Lars Hellan 20:28, 20 October 2014 (UTC)
This webdemo offers verb valence information in a uniform format for four languages (listed in the order of their inclusion), which can be activated in any combination:
Norwegian Ga Spanish Bulgarian
The web demo:
The demo offers 5 drop-down search menus and one write-in field; any combination of them can be activated in a search:
Write-in-field: Search menus: Verb lexeme Syntactic Arguments Function Situation Aspect Type
The field ’Syntactic Arguments’ – henceforth ‘SAS’, illustrated in the list snippet (1),
NP+INF NP+INF:equiSBJ NP+INF:raisingSBJ NP+NP NP+NP+APpred
is based on so-called ‘formal’ syntactic categories, reflecting common analytic assumptions. For Norwegian, the set of possible SAS specifications is currently 158, which is close to being exhaustive at this level of specification. The symbol ‘+’ in (1) stands for linear order.
The field ’Function’ – henceforth ‘FCT’ - relates to a more traditional type of descriptive terms, such as ‘intransitive’, ‘transitive’, ‘transitive with oblique’, etc.. They provide less detail in differentiation than the SAS field, thus, for Norwegian, there are currently only 88 FCT term. In contrast to the SAS list, the FCT terms say nothing about linear order.
The fields ‘Situation’ and ‘Aspect’ contain situation type and aspectual properties of situations expressed, thus both representing semantic information. Currently these fields
In ‘Verb lexeme’ one writes a verb of the language(s) selected, or an initial substring of a verb, in combination with as many parameter specifications as one wishes. No matter which fields are specified, the result of a search is a number of verbs from the language(s) activated satisfying the conditions given. For each verb in this list, one can prompt a further specification of that verb’s properties: these will include the conditions specified, but also all other properties associated with the verb in the resource present in the database. We illustrate with the typed string hoppe; 'SHOW' is the prompt button:
Verb lexeme Syntactic Arguments Function Situation Aspect Type hoppe
Search Result no SHOW hoppe_intrdir no SHOW hoppe_secpred no SHOW hoppe_secpred-refl
‘no’ here means ‘Norwegian’, and ‘hoppe_intrdir ‘ is the identifier of the lexical entry in the Norwegian lexical resource. When pushing on SHOW, one sees the full amount of specification stored in the databse for that verb entry; e.g., for hoppe_intrdir:
(3) Lexicon Instance
Language no Verb Id hoppe_intrdir SAS NP FCT intransitive SIT directedMotion Aspect Verb Type v-intr-suDir Example of type gutten løper Orthography < "hoppe" > Phon Engl-gloss Example Gloss Free-transl TypeCraft URLs Example from TypeCraft
When you click on the link in the last line, you get to an annotated token with hoppe in the TC editor, with information including what is shown in the figure below:
So far rather few of the 12 000 Norwegian verb entries have an annotated example in TypeCraft, but more will come. For Spanish and Bulgarian there are at the moment none, while for Ga, corresponding information is given for each entry under the lines Phon, Engl-gloss, Example, Gloss, Free-transl, imported from a ToolBox project.
For each language, the information offered is based on a computational grammar of the language including a large lexicon; in each case, the grammatical framework used is HPSG, and the implementation platform is the Language Knowledge Builder (LKB; cf. Copestake 2002). The basic design and implementation was developed by members of the Research Group in Digital Linguistics at NTNU, Trondheim (see http://www.ntnu.edu/digital-linguistics ). Main contributors to the present version are, in alphabetic order: Dorothee Beermann, Tore Bruland, Mary Esther Kropp Dakubu (from the Ga grammar), Lars Hellan (from the Norwegian grammar), Montserrat Marimon (from the Spanish grammar), Petya Osenova (from the Bulgarian grammar); see: Hellan, L., D. Beermann, T. Bruland, M.E.K. Dakubu, and M. Marimon (2014) MultiVal: Towards a multilingual valence lexicon. LREC 2014.
In principle, a multilingual valence database should consist of the following:
The Languages: A selection of languages L1... Ln;
The Parameters: A set of specification parameters defined across all the languages (i.e., common parameters, in the sense of being independent of any particular language, although not in the sense of necessarily being relevant for all of the languages);
The Valence-profiles: For each language, an inventory of its valence types characterized in terms of the parameters available, called its valence-profile;
The Valence-type suites: For each language, a list of sentences instantiating each of its valence types, indexed according to the types;
The Valence Lexicons: For each language, a verb lexicon where each verb entry is classified according to its valence type (in addition to other lexical information);
The Valence Corpora: For each language, a sentence corpus instantiating each verb in each of the valence frames it can support.
The notion valence represents a perspective from the verb and thus from the Lexicon, whereas from the viewpoint of the sentence and the Corpora, the most closely corresponding term is argument structure, as when we talk about ‘the argument structure of a sentence’; since both perspectives are represented here, we use both terms. The sentential perspective is necessary when not just a single verb determines the argument structure of a sentence, such as when it is determined by a verb plus a secondary predicate, or resides in a series of verbs – the argument structure then results from the interplay between the valence of the constituent items and constructional factors. To widen the scope of the database to fully recognize constructional factors in such cases, we may think of it as a database of argument structure constructions, and use Construction-profile as an alternative to Valence-profile, and Construction-type suites as an alternative to Valence-type suites.
In general, the parameters selected for inclusion in the database must in the first place be amenable to formalization for a relational database, and in the second place accessible in an understandable form to those who input data and search for data. The latter point is connected to the need for a flexible inventory of terms, on the one hand accommodating terminologies of various frameworks, on the other hand being based on consistent conversion systems between terminologies.
From the viewpoint of standard linguistic adequacy, the following factors may be expected in the representation of argument structure:
a. syntactic argument structure, i.e., whether there is a subject, an object, a second/indirect object, etc., referred to as grammatical functions, and the formal categories carrying them; b. semantic argument structure, that is, how many participants are present in the situation depicted, and which roles they play (such as ‘agent’, ‘patient’, etc.); c. linkage between syntactic and semantic argument structure, i.e., which grammatical functions express which roles, and possible roles not expressed; here also belong identity relations, part-whole relations, etc., between arguments; d. aspect and Aktionsart, that is, properties of a situation expressed by a sentence with the valence in question in terms of whether it is dynamic/stative, continuous/instantaneous, completed/ongoing, etc.; e. type of the situation expressed, in terms of some classificatory system.
Some, but not all, of these factors are represented in the present database. It is derived from independently existing resources, so that for each language in the database the information stored depends on the resources already existing for that language.
The current content of the database is constructed from implemented HPSG grammars using the DELPH-IN grammar engineering resources including the computational platform LKB (Copestake 2002). The construction takes into account the lexicon files of these grammars, and a conversion script with rewrite rules of the form in (5) below. The leftmost item in this rule is a lexical type, which, as standard in a typed feature structure system, reflects both grammatical and lexical properties. The rule rewrites the type symbol ‘v-ditr’, which essentially means ‘ditransitive headed by verb’, into the SAS counterpart ‘NP+NP+NP’ and the semantic specification of a three-place relation.
v-ditr => SAS: “NP+NP+NP” FCT: ditrans SIT: ternaryRel
This rewrite instruction will be part of the conversion rule for v-ditr in all of the languages; however, for each language, reflecting its grammar and lexicon, and the resources available in the database, the instructions supplementing the part in (5) may be different. In the Ga LKB grammar, for instance, verb entries are of the form instantiated in (6), reflecting their provenance from a Toolbox project:
bɔle_85 := v-ditr & [STEM <"bɔle">, PHON <"bɔ̀lè">, ENGL-GLOSS <"expect">, SYNSEM.LKEYS.KEYREL.PRED "_bɔle_v-ditr_rel", EXAMPLE "Wɔ-bɔle-ee bo nakai", GLOSS "1P.AOR-go.around-NEG.IMPERF 2S that", FREE-TRANSL "we didn't expect it of you, that you would behave in that manner."].
Entries in the other grammars lack PHON, ENGL-GLOSS and the three lines corresponding to a glossed example.
A difference between the Spanish and the Norwegian grammar is that the former employs optionality marking in its lexical types, so that one and the same entry in the Spanish lexical resource can represent, e.g., both a transitive and an intransitive frame, whereas in the Norwegian (and Ga) resource, an entry represents only one frame. This is reflected in the SAS and FCT inventories of terms, in that for some Spanish verbs, the entry identifier of one single verb entry may be matched to two or more SASs, possibly corresponding to a FCT label including the part 'Opt', whereas for Norwegian and Ga, a given entry identifier corresponds to only one SAS.
The current list of conversion rules for Norwegian is given in ConversionListNo ; the list of SAS specifications for Norwegian is given in SAS types No, and the list of FCT specifications for Norwegian is given in Funct types No.
The valence profile for Norwegian is given in Valence Profile Norwegian, and the valence profile for Ga in Ga Valence Profile. For both languages, the system used for defining Type is the Construction Labeling system, cf. Verbconstructions cross-linguistically - Introduction.
Download information from the Norwegian Valence project.
Download Lars Hellan and Dorothee Beermann 's presentation on Infinitive constructions in Norwegian, in a comparative perspective
TO BE CONTINUED