AI4LD Workshop Programme

Slides 

Lexical-semantic resources (LSRs), such as online lexicons or wordnets, are fundamental for natural language processing applications. In many languages, however, such resources suffer from quality issues: incorrect entries, incompleteness, but also, the rarely addressed issue of bias towards the English language and AngloSaxon culture. Such bias manifests itself in the absence of concepts specific to the language or culture at hand, the presence of foreign (Anglo-Saxon) concepts, as well as in the lack of an explicit indication of untranslatability, also known as cross-lingual lexical gaps, when a term has no equivalent in another language. This study proposes a novel crowdsourcing methodology for reducing bias in LSRs. Crowd workers compare lexemes from two languages, focusing on domains rich in lexical diversity, such as kinship or food. Our LingoGap crowdsourcing tool facilitates comparisons through microtasks identifying equivalent terms, language-specific terms, and lexical gaps across languages. We validated our method by applying it to two case studies focused on food-related terminology: (1) English and Arabic, and (2) Standard Indonesian and Banjarese. These experiments identified 2,140 lexical gaps in the first case study and 951 in the second. The success of these experiments confirmed the usability of our method and tool for future large-scale lexicon enrichment tasks.

Indonesia is a vast country with over 700 languages. These languages are varied in multiple aspects. In order for machines to be able to process these languages, they have to pay close attention in the language diversity. In our studies, we take a closer look at multiple Indonesian languages: Indonesian, Javanese, and Banjarese. With these languages, we perform various experiments on kinship terms, food terms, and basic-level categories. This talk will discuss the results, the challenges, and the possible directions this line of research can provide to the future.

My research explores the digital marginalization of low-resource languages through the case study of Jejueo, and how they become reliant on context-specific data. By examining the Korean UKC database, I identified and corrected the errors appeared in Standard Korean dataset, and offered a new Jejueo corpus using open API dictionary data. Through this process, I experienced how difficult it is for the marginalized low-resource languages to gather and display the data digitally. This project highlights the inherently political narrative of dominant NLP technologies and how the underlying epistemic injustice on low-resource languages interrupts information acquisition and NLP technologies. It is expected to contribute to creating a diversity-aware, inclusive Jejueo dataset that include local contexts. 

This research investigates how epistemic injustice arises from the inability of multilingual AI models to effectively translate culturally unique lexical gaps in the Irish language. Through a carefully selected set of 40 untranslatable marine expressions and interviews with participants, the study assesses the translation results produced by ChatGPT and Google Translate. With an overall average error rate of 55.6%, the findings suggest that the literal translation of lexical gaps was the primary cause of mistranslations. These inaccuracies not only reflect technical mistakes but also highlight more profound structural inequalities linked to the creation and application of language technologies. By incorporating concepts of epistemic injustice, language modelling bias, and decolonial design, the research advocates for a transition toward participatory and culturally aware AI systems that actively support the preservation of marginalised languages.

Slides 

This research explores the epistemological and computational challenges of cross-lingual sentiment classification by comparing two core resources: SentiWordNet 3.0, a semi-automatically constructed, valence-oriented sentiment lexicon based on WordNet 3.0, and Micro-WordNet-Opinion, a manually annotated gold standard used for its evaluation. My project begins by investigating the underlying structures of these two resources and their differing approaches to modelling sentiment. On one hand, SentiWordNet uses statistical propagation algorithms and numerical scoring. On the other, Micro-WordNet is constituted by human judgement and qualitative nuance. A key finding is that SentiWordNet displays a strong bias toward neutrality, treating sentiment as a stable, scalar dimension, while comparative analysis with machine-learning approaches, in my case BERT-based sentence representations, reveals a more organic, context-sensitive sentiment space. Using both PCA and t-SNE, I show that sentiment in SWN behaves like a linear axis embedded within a fixed semantic geometry, while in mBERT, sentiment appears emergent, scattered across multidimensional embedding space.

The second stage of this research focuses on adapting Micro-WordNet as the foundation for a cross-lingual evaluation framework, beginning with Italian. This requires aligning WordNet synset IDs with UKC concept IDs and developing a human annotation protocol to evaluate sentiment propagation quality across languages. By showing the results of a spearman correlation and the top ten score divergences in both positivity and negativity between SWN3 and Micro-WN, I observe relevant inconsistencies that indicate the necessity of a deeper cultural and qualitative scrutiny.

In conclusion, I try to give the outlines of my future line of research, of which the present speech aims to be the foundation, by advancing a theoretical interpretation of sentiment as a site of tension between two ontologies of meaning: a universal ontology grounded in the use of formal- symbolic logics and a regional ontology grounded in the particularism of specific semantic information. I further argue for a “meso-ontological” perspective that tries to locate computational approaches to sentiment within this “in-between” space, where symbolic structure and socio-affective context co-produce sentiment as a form of affectively-determined knowledge.

Slides 

Languages capture the world in strikingly diverse ways. This diversity manifests pervasively across lexicons through phenomena like lexical gaps and untranslatability—yet it remains poorly represented in computational resources, such as multilingual lexical databases. In this paper, we employ a systematic hybrid approach, combining expert-sourcing and crowdsourcing methods, to enrich computational lexicons with Turkish-language data that reflect linguistic diversity. Our experiments focus on two key domains: (1)

cognitively salient basic-level concepts (e.g., chair, dog, write) and (2) kinship terminology. Specifically, we: (1) Translate 1,000 basic-level concepts into Turkish, (2) Crowdsource lexical diversity in kinship systems by comparing English and Turkish kinship terms bidirectionally, and (3) Map language-independent kinship concepts to Turkish. Our results reveal substantial gaps and enrichments: in kinship terminology, we identify 194 lexical gaps, 47 terms, and 4 novel concepts, while the basic-level categories yield 953 translated words and 15 lexical gaps. 

Languages represent the world in remarkably diverse ways. Across different lexicons, this diversity manifests through phenomena such as lexical gaps and untranslatability. However, such richness is rarely reflected in computational resources like multilingual lexical databases. In this study, we present a series of experiments aimed at enriching computational lexicons with Persian linguistic data that highlights lexical diversity. Our approach employs a systematic hybrid of expert-sourcing and crowdsourcing methodologies to capture nuanced linguistic content. The experiments focus on cognitively salient basic-level concepts (e.g., chair, dog, write) and kinship terminology, a culturally rich and structurally complex domain. Specifically, we conduct the following: (1) Translating 1,000 basic-level concepts into Persian; (2) Bidirectional crowdsourced comparisons of kinship terms between Persian and English; (3) Bidirectional comparisons between Persian and Arabic kinship terms; and (4) Mapping language-independent kinship concepts to Persian equivalents.  Our findings include 903 Persian lexical entries and the identification of 65 lexical gaps within the basic-level category. In the kinship domain, we identified 228 lexical gaps, 70 culturally specific terms, and 3 previously undocumented kinship concepts.

Slides 

Most of the more popular Large Language Models (LLMs) use Byte Pair Encoding (BPE) to construct their vocabulary. This technique is very convenient and has its fair share of advantages, from being able to tokenize Out Of Vocabulary (OOV) words to build their meaning from their sub-words.

However, this approach can leave behind nuances in the meaning of words, which might vary depending on the context. The current solution is to let the model construct them internally with not much clarity nor control. This limitation becomes even more problematic in low-resource languages, where insufficient data might not allow the model to develop these internal mechanisms. A lot of the meaning is lost, not only while interpreting user inputs, but also while producing relevant responses to them. In particular, the problem we are facing is the polycentricity of words, leading to possibly vastly different interpretation of the same word, which the currently widely used decoder- only transformer LLMs have no formal way of solving.

This is not a new problem; it has been around for a long time and several solutions have been proposed like: WordNet and hub-language mappings. However, these solutions are not perfect and face inherent limitations, notably lexical gaps—cases where specific meanings in one language have no direct translation in another. Other solutions, such as the Universal Knowledge Core (UKC), circumvent this problem by taking a concept-centric approach, where the languages are mapped to a language-independent domain; where concepts from multiple languages meet and are connected by semantic relations, such as: hypernym, hyponym, …

We want to make use of these formal resources to pre-embed the meaning of word senses. To do so, we propose the use of a Graph encoder, which ingests semantic relations in the UKC concept space and produce for each node (concept) a relevant embedding. Preliminary experiments using Word Sense Disambiguation (WSD) tasks indicate promising results. The current work is still limited and faces challenges in multilingual generalization and computational efficiency, with efforts target zero-shot learning and enhanced multilingual capabilities.

This would allow the models to abstract from the current lexicalization-based language initial encoding, paving the way for LLMs to deeply understand word senses without explicit sense-specific training. Recent work https://arxiv.org/abs/2504.06036 supports this, showing how by using disambiguated word embeddings smaller models can match the performance of bigger ones. Additionally, from the way our architecture builds meanings, updating the meaning of words for a language should not require full re- training of the model. Instead, it should suffice to feed our concept encoder with the updated graph to get a correct and semantically relevant embedding.

Importantly, we also believe that by using a language-agnostic representation such as the UKC concepts, it’s possible to build models which can understand the “long tail” of low - resource languages with much less effort than is currently required.