Speaker
Description
Background and aim
Large language models generate fluent, dictionary-like and analytically plausible text, yet their linguistic knowledge is encoded implicitly in numerical representations rather than as inspectable lexical entries, paradigms, feature structures or constructions. Classical annotation frameworks make categories explicit, but often reduce complex linguistic behaviour to flat tags and relations. This creates a representational gap for lexicography: lexical knowledge must be searchable, correctable, reusable and explainable, not merely recoverable from model output. Drawing on Construction Grammar (Goldberg, 2006; Croft, 2022), this paper presents an experimental model for Albanian that restores the lexicon to the beginning of processing, while redefining it as a constructicon: a structured inventory of lexical and grammatical form-content pairings. The aim is not to replace large language models or Universal Dependencies (UD; de Marneffe et al., 2021), whose purposes differ, but to develop a linguistically accountable architecture for richer lexical resources and annotation.
Model and method
The model represents constructions as transparent nested Python structures that can later be exchanged through JSON, XML or RDF-oriented formats. Each entry contains a compound_set with three components: constituent_set, which records the internal morphosyntactic parts; pos_set, which assigns a category to the complete entry or phrase; and feature_sets, which stores one or more compatible feature bundles. For nouns and noun-associated categories, the current feature inventory comprises gender, case, definiteness and number. A value may be concrete or left open as "ok", allowing another construction to supply it during unification. Two lexicons are presently distinguished: lexical-category dictionaries for nouns, adjectives, pronouns, prepositions, verbs and related classes, and marker dictionaries for inflectional elements. In a deliberate refinement of broader Construction Grammar terminology, form is restricted here to the graphic or phonological matter that activates an entry, whereas lexical-grammatical information is organized within content. Processing matches textual forms to stored constructions, tests compatibility, and builds larger usage-level constructs through licensed unification.
Illustration and preliminary evaluation
The Albanian noun shef 'chief' illustrates the procedure. The lemma is a lexical construction which contains information about (masculine) gender but leaves case, number and definiteness open. Unification with the postposed marker -i produces shef-i 'the chief', analysed internally as a noun plus marker and externally as a noun phrase. The surface form initially licenses more than one feature analysis; in the larger construct shef-i i zbulim-it 'the chief of intelligence', contextual compatibility eliminates the unavailable reading, thus solving the ambiguity. The same mechanism models adjective and prepositional phrases, numerals, and the distinct processing roles of Albanian postposed and preceding articles. Grammatical agreement is thus represented as alignment information distributed across constructions, rather than as a label checked only after parsing. Initial domain-limited trials on nominal, adjectival, prepositional and selected verbal structures exceeded 90% accuracy and rose above 95% after missed constructions and compatibility patterns were corrected. These figures remain provisional because the test material, evaluation unit and error distribution have not yet been standardized.
Contribution and future development
The principal contribution is architectural. Annotation becomes the transparent result of recognizing stored constructions and projecting their information into constructs, rather than a separate layer of labels attached to bare tokens. This relocates POS and feature information within a richer representation and permits ambiguity without premature disambiguation. For Albanian lexicography, the approach supports a transition from digital dictionary to constructicon (Lyngfelt, 2018), encompassing lexical items, markers, phraseological patterns and larger constructions across the lexicon-grammar continuum. The present implementation is nevertheless incomplete: semantic frames, corpus-based frequency and productivity, broader verbal coverage, and external benchmark evaluation remain future tasks. Because its hierarchy is regular and explicit, the resource can also be mapped to linked-data models (Cimiano et al., 2016) and connected with corpus examples, ontologies, semantic frames and cross-linguistic correspondences. It therefore offers a practical basis for an interpretable, extensible and interoperable Albanian lexical resource.