Adapting large language models to specialized domains remains challenging due to the computational cost of full finetuning and the limited availability of domain-specific parallel data. We present a systematic framework for parameter-efficient domain adaptation using Low-Rank Adaptation (LoRA) geared towards efficient learning in low-resource scenarios. Our method combines data-scaling analysis, dual-track hyperparameter optimization, and competitive benchmarking. We evaluate our approach on the low-resource English–Norwegian petroleum translation domain using a distilled version of NLLB and parallel data from the Norwegian Petroleum Directorate. Our adapted model achieves 61.48 BLEU (+24.62 over the base model) and 0.9298 COMET, while updating <0.4% of parameters. Our results provide a reproducible and computationally efficient blueprint for domain adaptation in neural machine translation, particularly for specialized and resource-constrained domains.

Shamsuddeen Hassan Muhammad, Nedjma Ousidhoum, Idris Abdulmumin, Jan Philip Wahle, Terry Ruas, Meriem Beloucif, Christine de Kock, Nirmal Surange, Daniela Teodorescu, Ibrahim Said Ahmad, David Ifeoluwa Adelani, Alham Fikri Aji, Felermino D. M. A. Ali, Ilseyar Alimova, Vladimir Araujo, Nikolay Babakov, Naomi Baes, Ana-Maria Bucur, Andiswa Bukula, Guanqun Cao, Rodrigo Tufiño, Rendi Chevi, Chiamaka Ijeoma Chukwuneke, Alexandra Ciobotaru, Daryna Dementieva, Murja Sani Gadanya, Robert Geislinger, Bela Gipp, Oumaima Hourrane, Oana Ignat, Falalu Ibrahim Lawan, Rooweither Mabuya, Rahmad Mahendra, Vukosi Marivate, Alexander Panchenko, Andrew Piper, Charles Henrique Porto Ferreira, Vitaly Protasov, Samuel Rutunda, Manish Shrivastava, Aura Cristina Udrea, Lilian Diana Awuor Wanzare, Sophie Wu, Florian Valentin Wunderlich, Hanif Muhammad Zhafran, Tianhui Zhang, Yi Zhou, Saif M. Mohammad. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Language is a form of symbolic capital that affects people's lives in many ways (Bourdieu, 1977(Bourdieu, , 1991)).As a powerful means of communication, it reflects identities, cultures, traditions, and societies more broadly.Therefore, data in a given language should be regarded as more than just a collection of tokens.Rigorous data collection and labeling practices are essential for developing more human-centered and socially aware technologies.Although there has been growing interest in under-resourced languages within the NLP community, work in this area faces unique challenges, such as data scarcity and limited access to qualified annotators.In this paper, we collect feedback from individuals directly involved in and impacted by NLP artefacts for medium-and low-resource languages.We conduct both quantitative and qualitative analyses of their responses and highlight key issues related to: (1) data quality, including linguistic and cultural appropriateness; and (2) the ethics of common annotation practices, such as the misuse of participatory research.Based on these findings, we make several recommendations for creating high-quality language artefacts that reflect the cultural milieu of their speakers, while also respecting the dignity and labor of data workers.

The recent rise in popularity of large language models (LLMs) has prompted considerable concerns about their moral capabilities. Although considerable effort has been dedicated to aligning LLMs with human moral values, existing benchmarks and evaluations remain largely superficial, typically measuring alignment based on final ethical verdicts rather than explicit moral reasoning. In response, this paper aims to advance the investigation of LLMs’ moral capabilities by examining their capacity to function as Artificial Moral Assistants (AMAs), systems envisioned in the philosophical literature to support human moral deliberation. We assert that qualifying as an AMA requires more than what state-of-the-art alignment techniques aim to achieve: not only must AMAs be able to discern ethically problematic situations, they should also be able to actively reason about them, navigating between conflicting values outside of those embedded in the alignment phase. Building on existing philosophical literature, we begin by designing a new formal framework of the specific kind of behaviour an AMA should exhibit, individuating key qualities such as deductive and abductive moral reasoning. Drawing on this theoretical framework, we develop a benchmark to test these qualities and evaluate popular open LLMs against it. Our results reveal considerable variability across models and highlight persistent shortcomings, particularly regarding abductive moral reasoning. Our work connects theoretical philosophy with practical AI evaluation while also emphasising the need for dedicated strategies to explicitly enhance moral reasoning capabilities in LLMs.

Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Abinew Ali Ayele, David Ifeoluwa Adelani, Ibrahim Said Ahmad, Saminu Mohammad Aliyu, Paul Röttger, Abigail Oppong, Andiswa Bukula, Chiamaka Ijeoma Chukwuneke, Ebrahim Chekol Jibril, Elyas Abdi Ismail, Esubalew Alemneh, Hagos Tesfahun Gebremichael, Lukman Jibril Aliyu, Meriem Beloucif, Oumaima Hourrane, Rooweither Mabuya, Salomey Osei, Samuel Rutunda, Tadesse Destaw Belay, Tadesse Kebede Guge, Tesfa Tegegne Asfaw, Lilian Diana Awuor Wanzare, Nelson Odhiambo Onyango, Seid Muhie Yimam, Nedjma Ousidhoum. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

In this paper, we leverage existing resources, such as WordNet and dependency parsing, to build the first Galician dataset for training semantic role labeling systems in an effort to expand available NLP resources.Additionally, we introduce Verbal Indexing, a new preprocessing method, which helps increase the performance when semantically parsing highly complex sentences.We use transfer learning to test both the resource and the Verbal Indexing method.Our results show that the effects of Verbal Indexing were amplified in scenarios where the model was both pre-trained and finetuned on datasets utilizing the method, but improvements are also noticeable when only used during fine-tuning.The best-performing Galician SRL model achieved an f1 score of 0.74, introducing a baseline for future Galician SRL systems.We also tested our method on Spanish where we achieved an f1 score of 0.83, outperforming the baseline set by the 2009 CoNLL Shared Task by 0.025, showing the merits of our Verbal Indexing method for pre-processing.

Shamsuddeen Muhammad, Idris Abdulmumin, Abinew Ayele, Nedjma Ousidhoum, David Adelani, Seid Yimam, Ibrahim Ahmad, Meriem Beloucif, Saif Mohammad, Sebastian Ruder, Oumaima Hourrane, Alipio Jorge, Pavel Brazdil, Felermino Ali, Davis David, Salomey Osei, Bello Shehu-Bello, Falalu Lawan, Tajuddeen Gwadabe, Samuel Rutunda, Tadesse Belay, Wendimu Messelle, Hailu Balcha, Sisay Chala, Hagos Gebremichael, Bernard Opoku, Stephen Arthur. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Comparative Question Answering (cQA) is the task of providing concrete and accurate responses to queries such as: “Is Lyft cheaper than a regular taxi?” or “What makes a mortgage different from a regular loan?”. In this paper, we propose two new open-domain real-world datasets for identifying and labeling comparative questions. While the first dataset contains instances of English questions labeled as comparative vs. non-comparative, the second dataset provides additional labels including the objects and the aspects of comparison. We conduct several experiments that evaluate the soundness of our datasets. The evaluation of our datasets using various classifiers show promising results that reach close-to-human results on a binary classification task with a neural model using ALBERT embeddings. When approaching the unsupervised sequence labeling task, some headroom remains.

Frame-semantic annotations exist for a tiny fraction of the world’s languages, Wikidata, however, links knowledge base triples to texts in many languages, providing a common, distant supervision signal for semantic parsers. We present WikiBank, a multilingual resource of partial semantic structures that can be used to extend pre-existing resources rather than creating new man-made resources from scratch. We also integrate this form of supervision into an off-the-shelf frame-semantic parser and allow cross-lingual transfer. Using Google’s Sling architecture, we show significant improvements on the English and Spanish CoNLL 2009 datasets, whether training on the full available datasets or small subsamples thereof.

Previous attempts at injecting semantic frame biases into SMT training for low resource languages failed because either (a) no semantic parser is available for the low resource input language; or (b) the output English language semantic parses excise relevant parts of the alignment space too aggressively. We present the first semantic SMT model to succeed in significantly improving translation quality across many low resource input languages for which no automatic SRL is available —consistently and across all common MT metrics. The results we report are the best by far to date for this type of approach; our analyses suggest that in general, easier approaches toward including semantics in training SMT models may be more feasible than generally assumed even for low resource languages where semantic parsers remain scarce. While recent proposals to use the crosslingual evaluation metric XMEANT during inversion transduction grammar (ITG) induction are inapplicable to low resource languages that lack semantic parsers, we break the bottleneck via a vastly improved method of biasing ITG induction toward learning more semantically correct alignments using the monolingual semantic evaluation metric MEANT. Unlike XMEANT, MEANT requires only a readily-available English (output language) semantic parser. The advances we report here exploit the novel realization that MEANT represents an excellent way to semantically bias expectationmaximization induction even for low resource languages. We test our systems on challenging languages including Amharic, Uyghur, Tigrinya and Oromo. Results show that our model influences the learning towards more semantically correct alignments, leading to better translation quality than both the standard ITG or GIZA++ based SMT training models on different datasets.

We propose an approach in which we inject a crosslingual semantic frame based objective function directly into inversion transduction grammar (ITG) induction in order to semantically train spoken language translation systems. This approach represents a follow-up of our recent work on improving machine translation quality by tuning loglinear mixture weights using a semantic frame based objective function in the late, final stage of statistical machine translation training. In contrast, our new approach injects a semantic frame based objective function back into earlier stages of the training pipeline, during the actual learning of the translation model, biasing learning toward semantically more accurate alignments. Our work is motivated by the fact that ITG alignments have empirically been shown to fully cover crosslingual semantic frame alternations. We show that injecting a crosslingual semantic based objective function for driving ITG induction further sharpens the ITG constraints, leading to better performance than either the conventional ITG or the traditional GIZA++ based approaches.

We present a novel model, Freestyle, that learns to improvise rhyming and fluent responses upon being challenged with a line of hip hop lyrics, by combining both bottomup token based rule induction and top-down rule segmentation strategies to learn a stochastic transduction grammar that simultaneously learns both phrasing and rhyming associations.In this attack on the woefully under-explored natural language genre of music lyrics, we exploit a strictly unsupervised transduction grammar induction approach.Our task is particularly ambitious in that no use of any a priori linguistic or phonetic information is allowed, even though the domain of hip hop lyrics is particularly noisy and unstructured.We evaluate the performance of the learned model against a model learned only using the more conventional bottom-up token based rule induction, and demonstrate the superiority of our combined token based and rule segmentation induction method toward generating higher quality improvised responses, measured on fluency and rhyming criteria as judged by human evaluators.To highlight some of the inherent challenges in adapting other algorithms to this novel task, we also compare the quality of the responses generated by our model to those generated by an out-ofthe-box phrase based SMT system.We tackle the challenge of selecting appropriate training data for our task via a dedicated rhyme scheme detection module, which is also acquired via unsupervised learning and report improved quality of the generated responses.Finally, we report results with Maghrebi French hip hop lyrics indicating that our model performs surprisingly well with no special adaptation to other languages.