Representing interlingual meaning in lexical databases

Abstract

In today’s multilingual lexical databases, the majority of the world’s languages are under-represented. Beyond a mere issue of resource incompleteness, we show that existing lexical databases have structural limitations that result in a reduced expressivity on culturally-specific words and in mapping them across languages. In particular, the lexical meaning space of dominant languages, such as English, is represented more accurately while linguistically or culturally diverse languages are mapped in an approximate manner. Our paper assesses state-of-the-art multilingual lexical databases and evaluates their strengths and limitations with respect to their expressivity on lexical phenomena of linguistic diversity.

Tackling Language Modelling Bias in Support of Linguistic Diversity

“…Lexico-semantic and lexical diversity, in turn, are ex- pressed via 39k interlingual gaps from 744 languages and over 230k language-specific relations including derivation, antonymy, and metonymy. When computed on identical data (obtained from (Giunchiglia et al, 2023)) and in an identical manner as for the other four multilingual lexical databases in Section 3, the language modelling bias of the UKC lexical model is obtained to be zero. This is expected as the model was designed to solve interlingual mapping limitations present in other databases.…”

“…As a generalisation of bilingual dictionaries, the 2000s saw the appearance of multilingual lexical databases that map words, based on their meanings, across large numbers of languages. As shown in the survey (Giunchiglia et al, 2023), several of these multilingual databases interconnect words from hundreds of languages, mapping the words of each language to the 100 thousand English word meanings (socalled synsets) of Princeton WordNet (Miller, 1998). On the one hand, this choice makes practical sense, as among all similar resources, WordNet offers by far the widest coverage of word meanings.…”

“…On the one hand, this choice makes practical sense, as among all similar resources, WordNet offers by far the widest coverage of word meanings. On the other hand, it results in a strong bias towards the English language and Anglo-Saxon culture in general, as the expressivity of the database is limited to notions for which a word exists in English (Giunchiglia et al, 2023;Bella et al, 2022a). Figure 1 provides a simple example from the food domain, known to be culturally, and thus also linguistically, diverse.…”

“…Applying the definition from Section 2, we compute the language modelling bias of the lexical models of four multilingual lexical databases: the first and second versions of the Open Multilingual Wordnet (OMW, OMW2) (Bond and Foster, 2013;Bond et al, 2020), In-doWordNet (IWN) (Bhattacharyya, 2010), and BabelNet (BN) (Navigli and Ponzetto, 2010). The figures were not computed from the actual contents of the databasesthat are necessarily incomplete and thus are not representative of their structural properties-but rather from a gold-standard cross-lingual mapping dataset (Giunchiglia et al, 2023), covering a diverse set of nine languages from five phyla. Coverages are theoretical in the sense that they mean the percentage of cross-lingual mapping relationships that each model is structurally able to represent with respect to the gold standard mappings.…”

“…As a case study on the value-sensitive design (Friedman et al, 2002) of language technology for the preservation of linguistic and cultural diversity, we present the Universal Knowledge Core (UKC), a large-scale multilingual lexical database, described from a technical perspective in (Bella et al, 2022b;Giunchiglia et al, 2023). The UKC adopts a diversity-aware lexical concept space that is able to represent concepts that are culturally or linguistically specific to languages and communities, avoiding the type of bias illustrated in Figure 1.…”