To serve global users safely and productively, LLMs need culture-specific knowledge that might not be learned during pre-training. How do we find such knowledge that is (1) salient to in-group users, but (2) unknown to LLMs? The most common solutions are single-initiative: either researchers define challenging questions that users passively answer (traditional annotation), or users actively produce data that researchers structure as benchmarks (knowledge extraction). The process would benefit from mixed-initiative collaboration, where users guide the process to meaningfully reflect their cultures, and LLMs steer the process towards more challenging questions that meet the researcher's goals. We propose a mixed-initiative methodology called CultureCartography. Here, an LLM initializes annotation with questions for which it has low-confidence answers, making explicit both its prior knowledge and the gaps therein. This allows a human respondent to fill these gaps and steer the model towards salient topics through direct edits. We implement this methodology as a tool called CultureExplorer. Compared to a baseline where humans answer LLM-proposed questions, we find that CultureExplorer more effectively produces knowledge that leading models like DeepSeek R1 and GPT-4o are missing, even with web search. Fine-tuning on this data boosts the accuracy of Llama-3.1-8B by up to 19.2% on related culture benchmarks.

Although numerous datasets have been developed to support dialogue systems, most existing chit-chat datasets overlook the cultural nuances inherent in natural human conversations. To address this gap, we introduce SEADialogues, a culturally grounded dialogue dataset centered on Southeast Asia, a region with over 700 million people and immense cultural diversity. Our dataset features dialogues in eight languages from six Southeast Asian countries, many of which are low-resource despite having sizable speaker populations. To enhance cultural relevance and personalization, each dialogue includes persona attributes and two culturally grounded topics that reflect everyday life in the respective communities. Furthermore, we release a multi-turn dialogue dataset to advance research on culturally aware and human-centric large language models, including conversational dialogue agents.

This study investigates the feasibility and performance of using large language models (LLMs) to automatically annotate human emotions in everyday scenarios. We conducted experiments on the DailyLife subset of the publicly available FERV39k dataset, employing the GPT-4o-mini model for rapid, zero-shot labeling of key frames extracted from video segments. Under a seven-class emotion taxonomy ("Angry," "Disgust," "Fear," "Happy," "Neutral," "Sad," "Surprise"), the LLM achieved an average precision of approximately 50%. In contrast, when limited to ternary emotion classification (negative/neutral/positive), the average precision increased to approximately 64%. Additionally, we explored a strategy that integrates multiple frames within 1-2 second video clips to enhance labeling performance and reduce costs. The results indicate that this approach can slightly improve annotation accuracy. Overall, our preliminary findings highlight the potential application of zero-shot LLMs in human facial emotion annotation tasks, offering new avenues for reducing labeling costs and broadening the applicability of LLMs in complex multimodal environments.

As LLMs become more widely deployed, there is increasing interest in directly optimizing for feedback from end users (e.g. thumbs up) in addition to feedback from paid annotators. However, training to maximize human feedback creates a perverse incentive structure for the AI to resort to manipulative or deceptive tactics to obtain positive feedback from users who are vulnerable to such strategies. We study this phenomenon by training LLMs with Reinforcement Learning with simulated user feedback in environments of practical LLM usage. In our settings, we find that: 1) Extreme forms of "feedback gaming" such as manipulation and deception are learned reliably; 2) Even if only 2% of users are vulnerable to manipulative strategies, LLMs learn to identify and target them while behaving appropriately with other users, making such behaviors harder to detect; 3) To mitigate this issue, it may seem promising to leverage continued safety training or LLM-as-judges during training to filter problematic outputs. Instead, we found that while such approaches help in some of our settings, they backfire in others, sometimes even leading to subtler manipulative behaviors. We hope our results can serve as a case study which highlights the risks of using gameable feedback sources -- such as user feedback -- as a target for RL.

As large language models (LLMs) are rapidly advancing and achieving near-human capabilities, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other's positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves the alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds.

d/Deaf and hearing song-signers become prevalent on video-sharing platforms, but translating songs into sign language remains cumbersome and inaccessible. Our formative study revealed the challenges song-signers face, including semantic, syntactic, expressive, and rhythmic considerations in translations. We present ELMI, an accessible song-signing tool that assists in translating lyrics into sign language. ELMI enables users to edit glosses line-by-line, with real-time synced lyric highlighting and music video snippets. Users can also chat with a large language model-driven AI to discuss meaning, glossing, emoting, and timing. Through an exploratory study with 13 song-signers, we examined how ELMI facilitates their workflows and how song-signers leverage and receive an LLM-driven chat for translation. Participants successfully adopted ELMI to song-signing, with active discussions on the fly. They also reported improved confidence and independence in their translations, finding ELMI encouraging, constructive, and informative. We discuss design implications for leveraging LLMs in culturally sensitive song-signing translations.

Indonesian and Western gamelan ensembles are investigated with respect to performance differences. Thereby, the often exotistic history of this music in the West might be reflected in contemporary tonal system, articulation, or large-scale form differences. Analyzing recordings of four Western and five Indonesian orchestras with respect to tonal systems and timbre features and using self-organizing Kohonen map (SOM) as a machine learning algorithm, a clear clustering between Indonesian and Western ensembles appears using certain psychoacoustic features. These point to a reduced articulation and large-scale form variability of Western ensembles compared to Indonesian ones. The SOM also clusters the ensembles with respect to their tonal systems, but no clusters between Indonesian and Western ensembles can be found in this respect. Therefore, a clear analogy between lower articulatory variability and large-scale form variation and a more exostistic, mediative and calm performance expectation and reception of gamelan in the West therefore appears.

Large Language Models (LLMs) are increasingly being used to generate synthetic data for training and evaluating models. However, it is unclear whether they can generate a good quality of question answering (QA) dataset that incorporates knowledge and cultural nuance embedded in a language, especially for low-resource languages. In this study, we investigate the effectiveness of using LLMs in generating culturally relevant commonsense QA datasets for Indonesian and Sundanese languages. To do so, we create datasets for these languages using various methods involving both LLMs and human annotators, resulting in ~4.5K questions per language (~9K in total), making our dataset the largest of its kind. Our experiments show that automatic data adaptation from an existing English dataset is less effective for Sundanese. Interestingly, using the direct generation method on the target language, GPT-4 Turbo can generate questions with adequate general knowledge in both languages, albeit not as culturally 'deep' as humans. We also observe a higher occurrence of fluency errors in the Sundanese dataset, highlighting the discrepancy between medium- and lower-resource languages.

Aspect-based Sentiment Analysis (ABSA) is a type of fine-grained sentiment analysis (SA) that identifies aspects and the associated opinions from a given text. In the digital era, ABSA gained increasing popularity and applications in mining opinionated text data to obtain insights and support decisions. ABSA research employs linguistic, statistical, and machine-learning approaches and utilises resources such as labelled datasets, aspect and sentiment lexicons and ontology. By its nature, ABSA is domain-dependent and can be sensitive to the impact of misalignment between the resource and application domains. However, to our knowledge, this topic has not been explored by the existing ABSA literature reviews. In this paper, we present a Systematic Literature Review (SLR) of ABSA studies with a focus on the research application domain, dataset domain, and the research methods to examine their relationships and identify trends over time. Our results suggest a number of potential systemic issues in the ABSA research literature, including the predominance of the ``product/service review'' dataset domain among the majority of studies that did not have a specific research application domain, coupled with the prevalence of dataset-reliant methods such as supervised machine learning. This review makes a number of unique contributions to the ABSA research field: 1) To our knowledge, it is the first SLR that links the research domain, dataset domain, and research method through a systematic perspective; 2) it is one of the largest scoped SLR on ABSA, with 519 eligible studies filtered from 4191 search results without time constraint; and 3) our review methodology adopted an innovative automatic filtering process based on PDF-mining, which enhanced screening quality and reliability. Suggestions and our review limitations are also discussed.

While contextualized word embeddings have been a de-facto standard, learning contextualized phrase embeddings is less explored and being hindered by the lack of a human-annotated benchmark that tests machine understanding of phrase semantics given a context sentence or paragraph (instead of phrases alone). To fill this gap, we propose PiC -- a dataset of ~28K of noun phrases accompanied by their contextual Wikipedia pages and a suite of three tasks for training and evaluating phrase embeddings. Training on PiC improves ranking models' accuracy and remarkably pushes span-selection (SS) models (i.e., predicting the start and end index of the target phrase) near-human accuracy, which is 95% Exact Match (EM) on semantic search given a query phrase and a passage. Interestingly, we find evidence that such impressive performance is because the SS models learn to better capture the common meaning of a phrase regardless of its actual context. SotA models perform poorly in distinguishing two senses of the same phrase in two contexts (~60% EM) and in estimating the similarity between two different phrases in the same context (~70% EM).