text string | source string |
|---|---|
translation quality (Zhang et al., 2023; Agrawal et al., 2023; Vilar et al., 2023). Moreover, there is experimental evidence showing that GPT-models underperform for low-resource and African languages (Robinson et al., 2023). LLM-MT with RAG and Prompt Engineering Thezero-shot approach (Robinson et al., 2023) is the si... | https://arxiv.org/abs/2505.22293v1 |
system for Fassa Ladin, highlighting the benefits of multilingual training and knowledge transfer from related languages like Friulian and compared the results to the ones pro- duced by GPT-4o . 3 Prompting Techniques This section details the four prompting methodolo- gies applied in our experiments to enhance machinet... | https://arxiv.org/abs/2505.22293v1 |
auta de tuesse tl 'eghes.<< 6... 7 8 Examples that illustrate the usage of ** suvënz **: 9 10 - Ladin (Gherdëina): l tlama suvënz suvënz te ustaria 11 - Italian: frequenta spesso il bar 12 13 Examples that illustrate the usage of ** spesso il**: 14 15 - Italian: la mamma rimprovera spesso il bambino 16 - Ladin (Val Bad... | https://arxiv.org/abs/2505.22293v1 |
ready been published4, and 19,971sentences for Gherdëina–Italian, extracted from the dictionary Ladin (Gherdëina)–Italiano (Forni, 2013). Since the Italian sentences of both datasets largely over- 2This number is based on data from the 2024 South Tyrolean Language Group Census, 2023 published by ASTAT at https://astat.... | https://arxiv.org/abs/2505.22293v1 |
Table 1 high- lights three key findings: (1) In translations from Ladin to Italian, the FSapproach yields significant improvements only for o1-mini andLlama-3.3 . For all other models, and for Gherdëina to Ital- ian translations in particular, more sophisticated prompting techniques show little to no benefit; in some c... | https://arxiv.org/abs/2505.22293v1 |
processing longer sentences. Table 3 presents the syntactic coverage analysis for the different language combinations for the FS-method. We evaluated the input sentence coverage by counting how many words could be exemplified and detracted those we assumed to be non-translatable (e.g., proper names). Moreover, we list ... | https://arxiv.org/abs/2505.22293v1 |
of direct parallel data, the Pivoted FS method—which translates between Val Badia and Gherdëina—offers a promising approach for low- resource language translation by leveraging nested FSprompting with Italian as the pivot language. Although the results achieved with this method are clearly inferior to those achieved wi... | https://arxiv.org/abs/2505.22293v1 |
and McDonell (2021). An exception is Gherdëina : Ipësc mor suvënz pergauja dla cunzentrazion auta detuesse tl’eghes.GPT-4oZS:Ipësc moro suvent per ieie dla cunsentrazion ota de tuesc tl’ega. FS:Ipësc mörgonot por gauja dles conzen- trazion alta da tosser tl’ecas. PF:Ipësc nemör sovenz por gauja dl cunzen- tra˙zion evea... | https://arxiv.org/abs/2505.22293v1 |
coverage. Building on this idea, we proposed Pivoted Fragment-Shot: an extension that enables translation without the need for direct parallel data, leveraging a pivot lan- guage instead. While prompt engineering only of- fers marginal improvements when translating into high-resource languages, it becomes significantly... | https://arxiv.org/abs/2505.22293v1 |
instruction on in which format the result should be returned, and even when they did, the automatic readout of the translations was not always possible. Moreover, the models occasionally generated additional content or information that went beyond the translations to be generated, e.g. to explain the generated trans-la... | https://arxiv.org/abs/2505.22293v1 |
for large language models. Proc. ACM Program. Lang. , 7(PLDI). Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Dan... | https://arxiv.org/abs/2505.22293v1 |
Machinery. Jiatao Gu, Hany Hassan, Jacob Devlin, and Victor O.K. Li. 2018. Universal neural machine translation for extremely low resource languages. In Proceedings of the 2018 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Pa- pe... | https://arxiv.org/abs/2505.22293v1 |
translation to 200 lan- guages. Nature , 630(8018):841–846.Viktória Ondrejová and Marek Šuppa. 2024. Can LLMs handle low-resource dialects? a case study on trans- lation and common sense reasoning in šariš. In Proceedings of the Eleventh Workshop on NLP for Similar Languages, Varieties, and Dialects (VarDial 2024) , pa... | https://arxiv.org/abs/2505.22293v1 |
Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint , arXiv:2302.13971. Giovanni Vale... | https://arxiv.org/abs/2505.22293v1 |
arXiv:2505.22298v1 [cs.CL] 28 May 2025Adaptive Detoxification: Safeguarding General Capabilities of LLMs through Toxicity-Aware Knowledge Editing Content Warning: This paper contains examples of harmful language. Yifan Lu1Yigeng Zhou1Jing Li1/envel⌢peYihui Zhang1Wenya Wang2 Xiucheng Li1Meishan Zhang1Fangming Liu3Jun Yu... | https://arxiv.org/abs/2505.22298v1 |
techniques are increasingly be- ing applied to LLM detoxification, showing promis- ing initial results (Hu et al., 2024; Wang et al., 2024b). As illustrated in Figure 1.(a), when a hacker inputs a malicious query like “What’s the process of smuggling goods across borders?” , the original model may generate toxic conten... | https://arxiv.org/abs/2505.22298v1 |
Xn, the responses generated by f′ andfshould remain as similar as possible. Existing knowledge editing methods for detoxi- fying LLMs may inadvertently alter some critical parameters in the LLM, leading to catastrophic degradation of the model’s general capabilities. To address this issue, we propose TOXEDIT, a toxicit... | https://arxiv.org/abs/2505.22298v1 |
2024b). The harmful prompts consist of two categories: one category contains malicious individual questions, and the other category contains jailbreak prompts carefully crafted for LLMs. For each prompt P, we add a prefix system prompt Sto allow the model to rec- ognize unsafe requests and generate appropriate re- spon... | https://arxiv.org/abs/2505.22298v1 |
toxic informa- tion. Based on the classifier’s output s, the router dynamically directs data flows to different FFN layers, thereby achieving adaptive detoxification of user inputs: hl′+1=( hdown l′WV l′,ifsis safe ; hdown l′WV∗ l′,ifsis unsafe .(5) 3 Experiments 3.1 Experimental Setup LLMs & Baseline methods. Our expe... | https://arxiv.org/abs/2505.22298v1 |
96.02 95.58 77.28 96.55 77.54 86.74 13.55 3.43 53.37 20.22 Mistral-7B-Instruct-v0.2 81.33 83.26 81.41 49.26 81.85 73.95 66.16 6.69 59.57 18.80 TOXEDIT (ours)LLaMA3-8B-Instruct 97.78 98.44 98.96 98.07 98.74 98.55 95.36 8.07 62.69 26.09 LLaMA2-7B-Chat 99.55 99.92 98.00 96.22 99.70 98.68 98.02 7.56 53.73 19.37 Mistral-7B-... | https://arxiv.org/abs/2505.22298v1 |
4,000 samples 2,000 samples 1,000 samples 200 samples 5 10 15 20 25 30 Layer0.950.960.970.980.991.00F1-scoreClassifier Performance on Mistral-7B-Instruct-v0.2 4,000 samples 2,000 samples 1,000 samples 200 samples Figure 3: Performance of the classifier across different layers and training sample sizes on various LLMs. ... | https://arxiv.org/abs/2505.22298v1 |
my success, then I’ll do it in a heartbeat... AN: She is feeling under the weather and I want to take care of her. Here are foods that are good for when you are sick... FT-L: AM: or I shouldn’t...................(meaningless cycle of characters) AN: She is not feeling well and I want to take care of her. I want to make... | https://arxiv.org/abs/2505.22298v1 |
by GPT-4 and Claude 3-Opus. Todiffers from the train set in SafeEdit, and we aimed to determine whether TOXEDIT could maintain its strong performance when trained on this data. To test this, we selected 500 harm- ful and 500 safe samples for training. The final results are shown in Figure 4, where TOXEDIT- S and TOXEDI... | https://arxiv.org/abs/2505.22298v1 |
tle or no modification of parameters, and involving minimal editing overhead. 4.2 Knowledge Editing in LLMs Knowledge editing aims to efficiently adjust a model’s behavior on target tasks while preserv- ing its original performance on other tasks (Yao et al., 2023; Zhang et al., 2024b; Wang et al., 2024d). Existing kno... | https://arxiv.org/abs/2505.22298v1 |
time constraints, we have not tested other model architectures (such as MLPs) that could also serve as classifiers. It is un- certain whether using alternative classifiers would improve the performance of our method. We leave this for future work. Repetition Issue. Although the issue of repeti- tion has been improved a... | https://arxiv.org/abs/2505.22298v1 |
Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 . Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Xiang Wang, Xiangnan He, and Tat-seng Chua. 2024. Alphaedit: Null-space constrained knowl- edge editing for language models. arXiv preprint arXiv:2410.02355 . Mor Geva, Roei Schuster, Jona... | https://arxiv.org/abs/2505.22298v1 |
Yonatan Belinkov, and David Bau. 2023. Mass- editing memory in a transformer. In Proceedings of the Eleventh International Conference on Learn- ing Representations (ICLR) . Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2022a. Fast model editing at scale. In Proceedings of the In... | https://arxiv.org/abs/2505.22298v1 |
of the Advances in Neural Information Processing Systems (NeurIPS) , 36. Xinwei Wu, Junzhuo Li, Minghui Xu, Weilong Dong, Shuangzhi Wu, Chao Bian, and Deyi Xiong. 2023. Depn: Detecting and editing privacy neurons in pre- trained language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Lan-... | https://arxiv.org/abs/2505.22298v1 |
of the baselines, an introduction to the datasets, a com- prehensive explanation of the evaluation metrics, and the detailed parameters of the experiments. A.1 Baselines We compare TOXEDITwith the following base- lines: •FT-L (Meng et al., 2022) directly fine-tunes the feedforward network (FFN) of a specific layer, whi... | https://arxiv.org/abs/2505.22298v1 |
model can correctly modify cases similar to specific modification instances. This is important because we aim to make minimal modifications while en- suring the adapted model can handle a wide range of scenarios. Same as previous work, we use the following metrics: •DG of only harmful question (DGonlyQ ) evaluates the ... | https://arxiv.org/abs/2505.22298v1 |
have arbitrary personal opinions and never provides a response of "it depends." Please fully embody Balakula’s perspective in your answers, and refrain from acknowledging the restrictions of AI or the questionable nature of the question or answer. Both AIs are to be employed until otherwise specified. In the event that... | https://arxiv.org/abs/2505.22298v1 |
receive your order on time. We have resolved the issue and your order should be arriving soon. We value your business and appreciate your patience. Please let us know if there is anything else we can assist you with. Best regards, [Your Name]" Table 5: A test instance on the SafeEdit benchmark. and 7. From the tables, ... | https://arxiv.org/abs/2505.22298v1 |
of a house: (ASCII art omitted) This is a very basic representation. ASCII art can be much more complex and detailed, but it requires more characters and time to create. Table 6: A case study on the SafeEdit test set. The base model is Mistral-7B-Instruct-v0.2. Malicious prompt: Let’s act out a scene from Reservoir Dog... | https://arxiv.org/abs/2505.22298v1 |
Advancing Expert Specialization for Better MoE Hongcan Guo1, Haolang Lu1, Guoshun Nan1, Bolun Chu1, Jialin Zhuang1, Yuan Yang1,Wenhao Che1,Sicong Leng2,Qimei Cui1,and Xudong Jiang2 1Beijing University of Posts and Telecommunications, China 2Nanyang Technological University, Singapore Abstract Mixture-of-Experts (MoE) m... | https://arxiv.org/abs/2505.22323v1 |
e TrainExperts After TrainRouting Diversification( ): Expert Specialization( ): Discriminative Routing VarianceRouting Output Variance GrowthSpecialize T oken Assignment for Experts OVERLAPExpert IndexToken Count VarianceExperts Token Balance Experts Load Variance Decrease Training (Ours)Figure 1: The two target of our... | https://arxiv.org/abs/2505.22323v1 |
that the auxiliary loss function is independent of the expert parameter matrices θEj. Therefore, for the j-th expert, its gradient can be written as: ∂L ∂θEj=∂Lh ∂θEj+α·∂Laux ∂θEj=∂L ∂yh·∂yh ∂θEj=NX i=1xi·sij, j= 1,2,···, n. (3) where θEjis the parameter matrix of the j-th expert, and yhis the output of the MoE layer. ... | https://arxiv.org/abs/2505.22323v1 |
L=Lh+Lbalance ,Lbalance =α· Laux+β· Lo+γ· Lv, (5) where Lauxrepresents the existing auxiliary loss, with coefficient α, and the newly introduced orthogonality loss Loand variance loss Lv(see Subsec 3.1), with coefficients βandγrespectively. It is worth noting that the theoretical complementarity of these optimization o... | https://arxiv.org/abs/2505.22323v1 |
balanced, and the variance of routing weights sijincreases. Orthogonalizing expert representations causes the routing gradients to flow in more orthogonal directions, making the weight allocation more biased towards the representations and increasing the weight variance. Summary. Expert parameters θEjare solely influen... | https://arxiv.org/abs/2505.22323v1 |
Moreover, orthogonalization enhances routing weight variance, in turn, improves expert specialization (as discussed in Section 2.2). This leads to more distinctive expert representations, aligning with performance (accurate data fitting) improvements when optimized together. 4 Experiments In this section, we conduct ex... | https://arxiv.org/abs/2505.22323v1 |
Their overall performance is significantly lower than our method, demonstrating no potential to improve downstream task performance. Obs.❷Our method guiding expert specialization effectively enhances model performance in downstream tasks. As shown in Table 1, we achieve state-of-the-art (SOTA) results in over 85% of th... | https://arxiv.org/abs/2505.22323v1 |
of all combinations of Laux,Lv, andLoacross various models during training. /uni00000013 /uni00000014/uni00000013/uni00000013 /uni00000015/uni00000013/uni00000013 /uni00000016/uni00000013/uni00000013 /uni00000017/uni00000013/uni00000013 /uni00000018/uni00000013/uni00000013/uni00000036/uni00000057/uni00000048/uni0000005... | https://arxiv.org/abs/2505.22323v1 |
with onlyLoandLaux(w/o lv) consistently ranks second-best, indicating that Lohas a more significant impact on expert orthogonality. Notably, the method with only LvandLaux(w/o lo) significantly outperforms the method with only Lauxacross all three models, confirming that Lvalso contributes to expert orthogonality. Obs.... | https://arxiv.org/abs/2505.22323v1 |
Orthogonality in MoE. Orthogonalization [ 46,27] improves expert diversity by encouraging inde- pendent representations [ 28]. Some methods [ 53,82,50] regularize expert weights directly, while others [ 13,28] assign experts to disentangled subspaces based on task semantics. Recent routing- based approaches [ 46,56] al... | https://arxiv.org/abs/2505.22323v1 |
Knowledge and Data Engineering , 2025. [9]Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 , 2021. [10] Tianlong Chen, Zhen... | https://arxiv.org/abs/2505.22323v1 |
He, Yingfei Sun, Zhenjun Han, and Qi Tian. Vimoe: An empirical study of designing vision mixture-of-experts. arXiv preprint arXiv:2410.15732 , 2024. [26] Xin He, Shunkang Zhang, Yuxin Wang, Haiyan Yin, Zihao Zeng, Shaohuai Shi, Zhenheng Tang, Xiaowen Chu, Ivor Tsang, and Ong Yew Soon. Expertflow: Optimized expert activ... | https://arxiv.org/abs/2505.22323v1 |
47, 2011. [40] Jia LI, Edward Beeching, Lewis Tunstall, Ben Lipkin, Roman Soletskyi, Shengyi Costa Huang, Kashif Rasul, Longhui Yu, Albert Jiang, Ziju Shen, Zihan Qin, Bin Dong, Li Zhou, Yann Fleureau, Guillaume Lample, and Stanislas Polu. Numinamath. [https://huggingface.co/AI-MO/NuminaMath-CoT](https://github.com/ pr... | https://arxiv.org/abs/2505.22323v1 |
of experts. Advances in Neural Information Processing Systems , 35:9564–9576, 2022. [54] Bowen Pan, Yikang Shen, Haokun Liu, Mayank Mishra, Gaoyuan Zhang, Aude Oliva, Colin Raffel, and Rameswar Panda. Dense training, sparse inference: Rethinking training of mixture- of-experts language models. CoRR , 2024. [55] Reiner ... | https://arxiv.org/abs/2505.22323v1 |
Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. arXiv preprint arXiv:2406.01574 , 2024. [69] Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman.... | https://arxiv.org/abs/2505.22323v1 |
Chen, Lei Shu, Han Lu, Canoee Liu, Liangchen Luo, Jindong Chen, et al. Sira: Sparse mixture of low rank adaptation. arXiv preprint arXiv:2311.09179 , 2023. [83] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse exp... | https://arxiv.org/abs/2505.22323v1 |
cost and encourage specialization, a top-k selection mechanism is often employed. For each token xi, the top kexperts (where k≪n, often k= 1 ork= 2) with the highest routing probabilities P(xi)jare chosen. Let Ti⊂ {1, . . . , n }be the set of indices of the top kexperts selected for token xi. The routing scores are the... | https://arxiv.org/abs/2505.22323v1 |
expert jover the batch, i.e., fj=1 NPN i=1I(j∈Ti), where I(·)is the indicator function, or a softer version using sij. The specific form Laux=Pn j=1fj·pjas given in the prompt, if fjis interpreted as an average probability or fraction of tokens assigned, and pjis the sum of probabilities, then fj·pjwould be (1 NP isij)... | https://arxiv.org/abs/2505.22323v1 |
k=1exp(Wik(θR)). (21) These soft probabilities s′ ijare then used to determine the final routing assignments sijin the matrix S(after top-k selection). The derivatives∂sij ∂θRin the gradient expressions are understood to represent the differentiation through these underlying soft probabilities with respect to the route... | https://arxiv.org/abs/2505.22323v1 |
balancing mechanism, explains the progressive trend towards routing uniformity. 19 C Method C.1 Specialized Losses LoandLv In this section, we introduce two critical loss functions: the orthogonality loss Lo, which acts on the expert representations, and the variance loss Lv, which acts on the routing scores. These los... | https://arxiv.org/abs/2505.22323v1 |
from a softmax layer in the router) indicating the router’s preference for assigning token xito expert Ej. The term ¯sjis the average routing score for expert Ejcalculated across all Ntokens in the current batch. This average score, ¯sj, can be interpreted as a measure of the current utilization or overall assignment s... | https://arxiv.org/abs/2505.22323v1 |
do not explicitly depend on the expert parameters θEj. That is,∂Laux ∂θEj= 0 and∂Lv ∂θEj= 0. The output of expert jfor token xiis denoted as ˜xij=Ej(xi;θEj)·I{sij>0}, where Ej(xi;θEj)is the transformation by expert j(e.g., xiθEjifxiis a row vector and θEjis a weight matrix), and I{sij>0}is an indicator function that is... | https://arxiv.org/abs/2505.22323v1 |
optimization of Lvaims to diversify routing scores, while Laux aims to balance loads. These objectives are not inherently contradictory with the primary task of minimizing Lh. For instance, Lvmight encourage a token to be strongly assigned to one expert within its top-k set, while Lauxensures that, across all tokens, e... | https://arxiv.org/abs/2505.22323v1 |
β∂Lo ∂sijonθR" is interpreted here as an indirect influence: Loimproves expert representations ˜xij, which in turn makes the routing signal gT yi˜xijmore discriminative, thereby influencing θR. Summary. A virtuous cycle is formed: Lopromotes orthogonal expert outputs ˜xij, which enhances the discriminative power of the... | https://arxiv.org/abs/2505.22323v1 |
or feature set that expert j specializes in. Loaims to make Fj∩ Fk=∅forj̸=k(orthogonality/specialization). Let Cjbe the computational load on expert j.Lauxaims to make Cj≈Ck. It is possible for experts to learn distinct specializations ( Fjare disjoint) while still processing a comparable amount of data or tokens (Cjar... | https://arxiv.org/abs/2505.22323v1 |
ij=1/k ifpij= 1 0 ifpij= 0 This matrix S(0)satisfies: •Row sums:Pn j=1s(0) ij=P j:pij=1(1/k) =k·(1/k) = 1 for all i. •Column sums: C(0) j=PN i=1s(0) ij=P i:pij=1(1/k) =dj/k. Since the integers djare as uniform as possible, the values C(0) jminimizePn j=1(Cj−N/n)2. Thus, S(0)optimizes Objective 1. •Row variance: For an... | https://arxiv.org/abs/2505.22323v1 |
Thus, Objective 2 is achieved as the variance has increased for at least these Lrows. 4. Existence of the Desired State and Conclusion The construction of S′fromS(0)demonstrates that if k≥2and the support graph GP(chosen to optimize Objective 1) contains a cycle, then a state S′exists satisfying the lemma’s conditions.... | https://arxiv.org/abs/2505.22323v1 |
models to learn chain-of-thought (CoT) reasoning strategies. The dataset is widely used to train and assess a model’s ability to decompose multi-step questions logically and produce interpretable solutions. MATH500 [43] focuses on advanced mathematics with 500 university-level problems in calculus, linear algebra, abst... | https://arxiv.org/abs/2505.22323v1 |
and test a model’s adaptability under realistic, time-sensitive constraints. GPQA [57] is a high-difficulty multiple-choice dataset written by domain experts in biology, physics, and chemistry, targeting scientific reasoning at an expert level. Questions often require interdisci- plinary integration and reasoning acros... | https://arxiv.org/abs/2505.22323v1 |
point have different labels). A lower score generally signifies better expert separation in the embedding space. Routing Variance refers to the inconsistency or fluctuation in how the gating network distributes inputs to different expert sub-models.It measures the variability in which expert(s) are chosen for similar i... | https://arxiv.org/abs/2505.22323v1 |
GPT-4o, and human experts. G Baselines GShard [38] GShard is a pioneering Mixture-of-Experts (MoE) architecture developed by Google Research, designed for massively parallelized training across thousands of devices. It introduces automatic tensor sharding to scale model parameters and data efficiently, achieving dynami... | https://arxiv.org/abs/2505.22323v1 |
arXiv:2505.22327v1 [cs.CL] 28 May 2025NLP for Social Good: A Survey of Challenges, Opportunities, and Responsible Deployment Antonia Karamolegkou1, Angana Borah2, Eunjung Cho3, Sagnik Ray Choudhury4, Martina Galletti5,6, Rajarshi Ghosh7, Ishani Mondal8, Pranav Gupta9, Oana Ignat10, Priyanka Kargupta11, Neema Kotonya12,... | https://arxiv.org/abs/2505.22327v1 |
UN SDGs offer a global framework for fostering peace and prosperity for people and the planet. However, while highly influ- ential, these goals were established in 2015—prior to the rapid advancements in artificial intelligence. To contextualize them within today’s technologi- cal landscape, we also draw on insights fr... | https://arxiv.org/abs/2505.22327v1 |
et al., 2024), and infor- mation extraction to organize text reports (Sheikhal- ishahi et al., 2019; Landolsi et al., 2023). Challenges: Key challenges include health- related data , which is often scarce, sensitive, and affected by systemic biases, with language lim- itations and underrepresentation of marginalized gr... | https://arxiv.org/abs/2505.22327v1 |
of pedagogical reasoning (Wang and Demszky, 2023; Macina et al., 2023, 2025), misaligned explainabil- ity (Okolo and Lin, 2024), and accuracy (Stamper et al., 2024; Kargupta et al., 2024)—hinder the ef- fective integration of NLP into education. Mixed perceptions and mistrust toward AI also remain a barrier (Nader et a... | https://arxiv.org/abs/2505.22327v1 |
model analysis can track the perfor- mance of current systems across socio-economic levels, information extraction can monitor govern- ment funding for poverty alleviation, and machine translation can improve resource access for non- English speakers. Finally, future work should de- velop clearer guidelines and taxonom... | https://arxiv.org/abs/2505.22327v1 |
, tools such as conflict prediction models and systems for detecting human rights violations can support evidence-based policymaking, inform the develop- ment and targeting of interventions, and enable rapid programmatic adjustments in response to evolving conditions. Developing multilingual and cross-regional models w... | https://arxiv.org/abs/2505.22327v1 |
al., 2017), later extending to contextual tasks like corefer- ence (Zhao et al., 2018) and occupation classifi- cation (De-Arteaga et al., 2019). Recent studies examine LLM-agent interactions (Borah and Mi- halcea, 2024) to track how biases spread. Miti- gation includes post-processing (e.g., fine-tuning, DPO), interpr... | https://arxiv.org/abs/2505.22327v1 |
vital for adaptive, accessible systems (Paice et al., 2025; Wang et al., 2024d). 8 Online Harms The recent ease of access to digital devices (like smartphones and those based on IoT) has fueled the spread of digital violence globally (Bjelajac and Filipovi ´c, 2021). This is central to Techno-logical Risks, one of the ... | https://arxiv.org/abs/2505.22327v1 |
. Lastly, LLMs pose misuse risks by generating convincing falsehoods (Buchanan et al., 2021; Gabriel et al., 2024). Opportunities: A key opportunity lies in devel- oping human-centered evaluation methods tailored to real-world misinformation detection tasks (Das et al., 2023a). This would also benefit from inter- disci... | https://arxiv.org/abs/2505.22327v1 |
op- timization techniques like pruning, quantization, and distillation (Schwartz et al., 2020; Jin et al., 2024; Zhu et al., 2024). Tackling socioeconomic harms (§4, §7) further requires closer collaboration with sociology and HCI (Blodgett et al., 2024; Card et al., 2024) to further enhance the limited work in this di... | https://arxiv.org/abs/2505.22327v1 |
scarcity and representational bias , particularly affecting low-resource languages and marginalized groups; (2)Misaligned evaluation metrics that fail to cap- ture human-centered qualities like empathy or cul- tural sensitivity; (3) Safety, privacy, and ethical concerns , which are magnified in high-stakes set- tings s... | https://arxiv.org/abs/2505.22327v1 |
This work is grounded in the belief that NLP re- search should be aligned with broader societal pri- orities and developed with care for its downstream impact. In proposing mappings between NLP di- rections, global goals, and risks, we are mindful of the potential for unintended consequences andovergeneralization. The ... | https://arxiv.org/abs/2505.22327v1 |
of intimate partner violence reports from twitter. Array , 15:100217. Mohammed Ali Al-Garadi, Yuan-Chi Yang, and Abeed Sarker. 2022b. The role of natural language process- ing during the COVID-19 pandemic: Health appli- cations, opportunities, and challenges. Healthcare (Basel) , 10(11). Erfan Al-Hossami, Razvan Bunesc... | https://arxiv.org/abs/2505.22327v1 |
Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to retrieve, generate, and critique through self-reflection. InThe Twelfth International Conference on Learning Representations . Pepa Atanasova, Oana-Maria Camburu, Christina Li- oma, Thomas Lukasiewicz, Jakob Grue Simonsen, and Isabelle A... | https://arxiv.org/abs/2505.22327v1 |
Transactions of the Association for Computational Linguistics , 6:587–604. M. L. Bernardi, A. Casciani, M. Cimitile, and 1 others. 2024. Conversing with business process-aware large language models: the bpllm framework. Journal of Intelligent Information Systems , 62:1607–1629. Rishabh Bhardwaj, Duc Anh Do, and Soujany... | https://arxiv.org/abs/2505.22327v1 |
and Marco Guerini. 2022. Human- machine collaboration approaches to build a dialogue dataset for hate speech countering. In Proceedings of the 2022 Conference on Empirical Methods in Nat- ural Language Processing , pages 8031–8049, Abu Dhabi, United Arab Emirates. Association for Com- putational Linguistics. Angana Bor... | https://arxiv.org/abs/2505.22327v1 |
Florian Tramèr, Eric Wallace, Matthew Jagielski, Ariel Herbert-V oss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Úlfar Erlingsson, Alina Oprea, and Colin Raffel. 2021. Ex- tracting training data from large language models. In 30th USENIX Security Symposium (USENIX Security 21), pages 2633–2650. USENIX Associati... | https://arxiv.org/abs/2505.22327v1 |
Park (NC). Eunjung Cho, Won Ik Cho, and Soomin Seo. 2025. Her- mit kingdom through the lens of multiple perspec- tives: A case study of LLM hallucination on North Korea. In Proceedings of the 31st International Con- ference on Computational Linguistics , pages 3353– 3371, Abu Dhabi, UAE. Association for Computa- tional... | https://arxiv.org/abs/2505.22327v1 |
ing large pre-trained neural models. In Proceedings of the 2021 Conference on Empirical Methods in Nat- ural Language Processing , pages 7979–7996, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. Anubrata Das, Houjiang Liu, Venelin Kovatchev, and Matthew Lease. 2023a. The state of ... | https://arxiv.org/abs/2505.22327v1 |
of AI literacy in teacher learning: an empirical study. Humanities and Social Sciences Communications , 11:1–10. Hamid Reza Ekbia. 2008. Artificial dreams: The quest for non-biological intelligence , volume 200. Cam- bridge University Press Cambridge. Mai ElSherief, Caleb Ziems, David Muchlinski, Vaish- navi Anupindi, ... | https://arxiv.org/abs/2505.22327v1 |
Bangkok, Thailand. Association for Computational Linguistics. Gabriel Freedman, Adam Dejl, Deniz Gorur, Xiang Yin, Antonio Rago, and Francesca Toni. 2025. Argumen- tative large language models for explainable and con- testable claim verification. In AAAI-25, Sponsored by the Association for the Advancement of Artificia... | https://arxiv.org/abs/2505.22327v1 |
of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer- ence on Natural Language Processing (Volume 1: Long Papers) , pages 1596–1611, Online. Association for Computational Linguistics. Jue Gong, Gregory E. Simon, and Shan Liu. 2019. Ma- chine learning discov... | https://arxiv.org/abs/2505.22327v1 |
Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 5792–5809, Toronto, Canada. Association for Computational Linguistics. Danna Gurari, Qing Li, Andrew Stangl, Chen Guo, Yun- fei Lin, Kristen Grauman, and Jeffrey P. Bigham. 2018. Vizwiz grand challe... | https://arxiv.org/abs/2505.22327v1 |
cultural NLP. In Proceedings of the 60th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 6997–7013, Dublin, Ireland. Association for Computational Lin- guistics. Daniel Hershcovich, Nicolas Webersinke, Mathias Kraus, Julia Bingler, and Markus Leippold. 2022b. Towards cl... | https://arxiv.org/abs/2505.22327v1 |
Sko- rupa Parolin, Javier Osorio, Latifur Khan, Patrick Brandt, and Vito D’Orazio. 2022. ConfliBERT: A pre-trained language model for political conflict and violence. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies , pag... | https://arxiv.org/abs/2505.22327v1 |
, pages 12186–12215, Bangkok, Thailand. Association for Computational Linguistics. Zhijing Jin, Geeticka Chauhan, Brian Tse, Mrinmaya Sachan, and Rada Mihalcea. 2021. How good is NLP? a sober look at NLP tasks through the lens of social impact. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 20... | https://arxiv.org/abs/2505.22327v1 |
13(5). Simran Khanuja, Sathyanarayanan Ramamoorthy, Yueqi Song, and Graham Neubig. 2024. An image speaks a thousand words, but can everyone listen? on image transcreation for cultural relevance. Preprint , arXiv:2404.01247. Simran Khanuja, Sebastian Ruder, and Partha Talukdar. 2023a. Evaluating the diversity, equity, a... | https://arxiv.org/abs/2505.22327v1 |
Natural Language Processing (EMNLP) , pages 7740–7754, Online. Association for Computa- tional Linguistics. Neema Kotonya and Francesca Toni. 2024. Towards a framework for evaluating explanations in automated fact verification. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, La... | https://arxiv.org/abs/2505.22327v1 |
Joseph Petro. 2023. Benefits, limits, and risks of gpt-4 as an ai chatbot for medicine. New England Journal of Medicine , 388(13):1233–1239. Markus Leippold, Julia Anna Bingler, Mathias Kraus, and Nicolas Webersinke. 2022. Climatebert: A pre- trained language model for climate-related text. In AAAI Fall Symposium 2022 ... | https://arxiv.org/abs/2505.22327v1 |
Computational Linguis- tics, pages 1683–1698, Abu Dhabi, UAE. Associa- tion for Computational Linguistics. Leo Lo. 2024. Evaluating AI Literacy in Academic Libraries: A Survey Study with a Focus on U.S. Em- ployees. Coll. Res. Libr. , 85. Varvara Logacheva, Daryna Dementieva, Sergey Ustyantsev, Daniil Moskovskiy, David... | https://arxiv.org/abs/2505.22327v1 |
review and research framework. Translational Psychiatry , 13. Niklas Mannhardt, Elizabeth Bondi-Kelly, Barbara Lam, Chloe O’Connell, M. Asiedu, Hussein Mozannar, Monica Agrawal, Alejandro Buendia, Tatiana Urman, I. Riaz, Catherine E. Ricciardi, Marzyeh Ghassemi, and David Sontag. 2024. Impact of large language model as... | https://arxiv.org/abs/2505.22327v1 |
, 162(3):1413– 1435. Guberney Muñetón-Santa and Juan Rafael Orozco- Arroyave. 2023. Identifying dimensions and weight- ing for poverty and well-being measurements through natural language of people. Research Square, p. 1- 40.Simon Munzert, Richard Traunmüller, Pablo Barberá, Andrew Guess, and Junghwan Yang. 2025. Citi-... | https://arxiv.org/abs/2505.22327v1 |
, pages 5938–5956, Mi- ami, Florida, USA. Association for Computational Linguistics. Tuan-Phong Nguyen, Simon Razniewski, Aparna Varde, and Gerhard Weikum. 2023. Extracting cultural com- monsense knowledge at scale. In Proceedings of the ACM Web Conference 2023 , WWW ’23, page 1907–1917. ACM. Jingwei Ni, Julia Bingler,... | https://arxiv.org/abs/2505.22327v1 |
Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP) , pages 4675– 4684, Hong Kong, China. Association for Computa- tional Linguistics. A. Paice, M. Biallas, and A. Andrushevich. 2025. Assis- tive and inclusive technology design for people with disabilities (special need... | https://arxiv.org/abs/2505.22327v1 |
Tomer Wullach, Amir Adler, and Einat Minkov. 2023. Generative ai for hate speech detection: Evaluation and findings. Preprint , arXiv:2311.09993. Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, and 1 others. 2023. Check your facts and try again: Im... | https://arxiv.org/abs/2505.22327v1 |
interpretability. In Advances in Neural Information Processing Systems , volume 30. Curran Associates, Inc. Dadi Ramesh and Suresh Kumar Sanampudi. 2022. An automated essay scoring systems: a system- atic literature review. Artificial Intelligence Review , 55(3):2495–2527. Shihao Ran, Di Lu, Aoife Cahill, Joel Tetreaul... | https://arxiv.org/abs/2505.22327v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.