Derry Wijaya

Afra Feyza Akyürek, Muhammed Yusuf Kocyigit, Sejin Paik et al.

Researchers have devised numerous ways to quantify social biases vested in pretrained language models. As some language models are capable of generating coherent completions given a set of textual prompts, several prompting datasets have been proposed to measure biases between social groups -- posing language generation as a way of identifying biases. In this opinion paper, we analyze how specific choices of prompt sets, metrics, automatic tools and sampling strategies affect bias results. We find out that the practice of measuring biases through text completion is prone to yielding contradicting results under different experiment settings. We additionally provide recommendations for reporting biases in open-ended language generation for a more complete outlook of biases exhibited by a given language model. Code to reproduce the results is released under https://github.com/feyzaakyurek/bias-textgen.

Afra Feyza Akyürek, Sejin Paik, Muhammed Yusuf Kocyigit et al.

Large language models trained on a mixture of NLP tasks that are converted into a text-to-text format using prompts, can generalize into novel forms of language and handle novel tasks. A large body of work within prompt engineering attempts to understand the effects of input forms and prompts in achieving superior performance. We consider an alternative measure and inquire whether the way in which an input is encoded affects social biases promoted in outputs. In this paper, we study T0, a large-scale multi-task text-to-text language model trained using prompt-based learning. We consider two different forms of semantically equivalent inputs: question-answer format and premise-hypothesis format. We use an existing bias benchmark for the former BBQ and create the first bias benchmark in natural language inference BBNLI with hand-written hypotheses while also converting each benchmark into the other form. The results on two benchmarks suggest that given two different formulations of essentially the same input, T0 conspicuously acts more biased in question answering form, which is seen during training, compared to premise-hypothesis form which is unlike its training examples. Code and data are released under https://github.com/feyzaakyurek/bbnli.

Anietie Andy, Siyi Liu, Daphne Ippolito et al.

While popular televised events such as presidential debates or TV shows are airing, people provide commentary on them in real-time. In this paper, we propose a simple yet effective approach to combine social media commentary and videos to create a multimedia summary of televised events. Our approach identifies scenes from these events based on spikes of mentions of people involved in the event and automatically selects tweets and frames from the videos that occur during the time period of the spike that talk about and show the people being discussed.

Ge Gao, Jongin Kim, Sejin Paik et al.

Predicting emotions elicited by news headlines can be challenging as the task is largely influenced by the varying nature of people's interpretations and backgrounds. Previous works have explored classifying discrete emotions directly from news headlines. We provide a different approach to tackling this problem by utilizing people's explanations of their emotion, written in free-text, on how they feel after reading a news headline. Using the dataset BU-NEmo+ (Gao et al., 2022), we found that for emotion classification, the free-text explanations have a strong correlation with the dominant emotion elicited by the headlines. The free-text explanations also contain more sentimental context than the news headlines alone and can serve as a better input to emotion classification models. Therefore, in this work we explored generating emotion explanations from headlines by training a sequence-to-sequence transformer model and by using pretrained large language model, ChatGPT (GPT-4). We then used the generated emotion explanations for emotion classification. In addition, we also experimented with training the pretrained T5 model for the intermediate task of explanation generation before fine-tuning it for emotion classification. Using McNemar's significance test, methods that incorporate GPT-generated free-text emotion explanations demonstrated significant improvement (P-value < 0.05) in emotion classification from headlines, compared to methods that only use headlines. This underscores the value of using intermediate free-text explanations for emotion prediction tasks with headlines.

Zilu Tang, Mayank Agarwal, Alex Shypula et al.

This work explores the use of self-generated natural language explanations as an intermediate step for code-to-code translation with language models. Across three types of explanations and 19 programming languages constructed from the MultiPL-E dataset, we find the explanations to be particularly effective in the zero-shot case, improving performance by 12% on average. Improvements with natural language explanations are particularly pronounced on difficult programs. We release our dataset, code, and canonical solutions in all 19 languages.

Afra Feyza Akyürek, Eric Pan, Garry Kuwanto et al.

Even the most advanced language models remain susceptible to errors necessitating to modify these models without initiating a comprehensive retraining process. Model editing refers to the modification of a model's knowledge or representations in a manner that produces the desired outcomes. Prior research primarily centered around editing factual data e.g. "Messi plays for Inter Miami" confining the definition of an edit to a knowledge triplet i.e. (subject, object, relation). However, as the applications of language models expand, so do the diverse ways in which we wish to edit and refine their outputs. In this study, we broaden the scope of the editing problem to include an array of editing cases such as debiasing and rectifying reasoning errors and define an edit as any natural language expression that solicits a change in the model's outputs. We are introducing DUnE-an editing benchmark where edits are natural language sentences and propose that DUnE presents a challenging yet relevant task. To substantiate this claim, we conduct an extensive series of experiments testing various editing approaches to address DUnE, demonstrating their respective strengths and weaknesses. We show that retrieval-augmented language modeling can outperform specialized editing techniques and neither set of approaches has fully solved the generalized editing problem covered by our benchmark.

Zilu Tang, Muhammed Yusuf Kocyigit, Derry Wijaya

Data augmentation techniques have been proven useful in many applications in NLP fields. Most augmentations are task-specific, and cannot be used as a general-purpose tool. In our work, we present AugCSE, a unified framework to utilize diverse sets of data augmentations to achieve a better, general purpose, sentence embedding model. Building upon the latest sentence embedding models, our approach uses a simple antagonistic discriminator that differentiates the augmentation types. With the finetuning objective borrowed from domain adaptation, we show that diverse augmentations, which often lead to conflicting contrastive signals, can be tamed to produce a better and more robust sentence representation. Our methods achieve state-of-the-art results on downstream transfer tasks and perform competitively on semantic textual similarity tasks, using only unsupervised data.

Muhammed Yusuf Kocyigit, Jiho Lee, Derry Wijaya

Quality Estimation (QE) models have the potential to change how we evaluate and maybe even train machine translation models. However, these models still lack the robustness to achieve general adoption. We show that State-of-the-art QE models, when tested in a Parallel Corpus Mining (PCM) setting, perform unexpectedly bad due to a lack of robustness to out-of-domain examples. We propose a combination of multitask training, data augmentation and contrastive learning to achieve better and more robust QE performance. We show that our method improves QE performance significantly in the MLQE challenge and the robustness of QE models when tested in the Parallel Corpus Mining setup. We increase the accuracy in PCM by more than 0.80, making it on par with state-of-the-art PCM methods that use millions of sentence pairs to train their models. In comparison, we use a thousand times less data, 7K parallel sentences in total, and propose a novel low resource PCM method.

Tahsina Hashem, Weiqing Wang, Derry Tanti Wijaya et al.

Knowledge Graph (KG)-to-Text generation aims at generating fluent natural-language text that accurately represents the information of a given knowledge graph. While significant progress has been made in this task by exploiting the power of pre-trained language models (PLMs) with appropriate graph structure-aware modules, existing models still fall short of generating faithful text, especially when the ground-truth natural-language text contains additional information that is not present in the graph. In this paper, we develop a KG-to-text generation model that can generate faithful natural-language text from a given graph, in the presence of noisy reference text. Our framework incorporates two core ideas: Firstly, we utilize contrastive learning to enhance the model's ability to differentiate between faithful and hallucinated information in the text, thereby encouraging the decoder to generate text that aligns with the input graph. Secondly, we empower the decoder to control the level of hallucination in the generated text by employing a controllable text generation technique. We evaluate our model's performance through the standard quantitative metrics as well as a ChatGPT-based quantitative and qualitative analysis. Our evaluation demonstrates the superior performance of our model over state-of-the-art KG-to-text models on faithfulness.

Garry Kuwanto, Eno-Abasi E. Urua, Priscilla Amondi Amuok et al.

Low-resource languages often face challenges in acquiring high-quality language data due to the reliance on translation-based methods, which can introduce the translationese effect. This phenomenon results in translated sentences that lack fluency and naturalness in the target language. In this paper, we propose a novel approach for data collection by leveraging storyboards to elicit more fluent and natural sentences. Our method involves presenting native speakers with visual stimuli in the form of storyboards and collecting their descriptions without direct exposure to the source text. We conducted a comprehensive evaluation comparing our storyboard-based approach with traditional text translation-based methods in terms of accuracy and fluency. Human annotators and quantitative metrics were used to assess translation quality. The results indicate a preference for text translation in terms of accuracy, while our method demonstrates worse accuracy but better fluency in the language focused.

Jongin Kim, Byeo Rhee Bak, Aditya Agrawal et al.

This paper introduces a multilingual dataset of COVID-19 vaccine misinformation, consisting of annotated tweets from three middle-income countries: Brazil, Indonesia, and Nigeria. The expertly curated dataset includes annotations for 5,952 tweets, assessing their relevance to COVID-19 vaccines, presence of misinformation, and the themes of the misinformation. To address challenges posed by domain specificity, the low-resource setting, and data imbalance, we adopt two approaches for developing COVID-19 vaccine misinformation detection models: domain-specific pre-training and text augmentation using a large language model. Our best misinformation detection models demonstrate improvements ranging from 2.7 to 15.9 percentage points in macro F1-score compared to the baseline models. Additionally, we apply our misinformation detection models in a large-scale study of 19 million unlabeled tweets from the three countries between 2020 and 2022, showcasing the practical application of our dataset and models for detecting and analyzing vaccine misinformation in multiple countries and languages. Our analysis indicates that percentage changes in the number of new COVID-19 cases are positively associated with COVID-19 vaccine misinformation rates in a staggered manner for Brazil and Indonesia, and there are significant positive associations between the misinformation rates across the three countries.

Lucky Susanto, Ryandito Diandaru, Adila Krisnadhi et al.

Neural machine translation (NMT) for low-resource local languages in Indonesia faces significant challenges, including the need for a representative benchmark and limited data availability. This work addresses these challenges by comprehensively analyzing training NMT systems for four low-resource local languages in Indonesia: Javanese, Sundanese, Minangkabau, and Balinese. Our study encompasses various training approaches, paradigms, data sizes, and a preliminary study into using large language models for synthetic low-resource languages parallel data generation. We reveal specific trends and insights into practical strategies for low-resource language translation. Our research demonstrates that despite limited computational resources and textual data, several of our NMT systems achieve competitive performances, rivaling the translation quality of zero-shot gpt-3.5-turbo. These findings significantly advance NMT for low-resource languages, offering valuable guidance for researchers in similar contexts.

Hieu Le, Taufiq Daryanto, Fabian Zhafransyah et al.

This paper focuses on a referring expression generation (REG) task in which the aim is to pick out an object in a complex visual scene. One common theoretical approach to this problem is to model the task as a two-agent cooperative scheme in which a `speaker' agent would generate the expression that best describes a targeted area and a `listener' agent would identify the target. Several recent REG systems have used deep learning approaches to represent the speaker/listener agents. The Rational Speech Act framework (RSA), a Bayesian approach to pragmatics that can predict human linguistic behavior quite accurately, has been shown to generate high quality and explainable expressions on toy datasets involving simple visual scenes. Its application to large scale problems, however, remains largely unexplored. This paper applies a combination of the probabilistic RSA framework and deep learning approaches to larger datasets involving complex visual scenes in a multi-step process with the aim of generating better-explained expressions. We carry out experiments on the RefCOCO and RefCOCO+ datasets and compare our approach with other end-to-end deep learning approaches as well as a variation of RSA to highlight our key contribution. Experimental results show that while achieving lower accuracy than SOTA deep learning methods, our approach outperforms similar RSA approach in human comprehension and has an advantage over end-to-end deep learning under limited data scenario. Lastly, we provide a detailed analysis on the expression generation process with concrete examples, thus providing a systematic view on error types and deficiencies in the generation process and identifying possible areas for future improvements.

Tahsina Hashem, Weiqing Wang, Derry Tanti Wijaya et al.

While large multimodal models (LMMs) have obtained strong performance on many multimodal tasks, they may still hallucinate while generating text. Their performance on detecting salient features from visual data is also unclear. In this paper, we develop a framework to generate faithful and salient text from mixed-modal data, which includes images and structured data ( represented in knowledge graphs or tables). Specifically, we train a small vision critic model to identify hallucinated and non-salient features from the image modality. The critic model also generates a list of salient image features. This information is used in the post editing step to improve the generation quality. Experiments on two datasets show that our framework improves LMMs' generation quality on both faithfulness and saliency, outperforming recent techniques aimed at reducing hallucination.

Isidora Chara Tourni, Derry Wijaya

With the advent of the Transformer architecture, Neural Machine Translation (NMT) results have shown great improvement lately. However, results in low-resource conditions still lag behind in both bilingual and multilingual setups, due to the limited amount of available monolingual and/or parallel data; hence, the need for methods addressing data scarcity in an efficient, and explainable way, is eminent. We propose an explainability-based training approach for NMT, applied in Unsupervised and Supervised model training, for translation of three languages of varying resources, French, Gujarati, Kazakh, to and from English. Our results show our method can be promising, particularly when training in low-resource conditions, outperforming simple training baselines; though the improvement is marginal, it sets the ground for further exploration of the approach and the parameters, and its extension to other languages.

Zilu Tang, Derry Wijaya

Incorporating tagging into neural machine translation (NMT) systems has shown promising results in helping translate rare words such as named entities (NE). However, translating NE in low-resource setting remains a challenge. In this work, we investigate the effect of using tags and NE hypernyms from knowledge graphs (KGs) in parallel corpus in different levels of resource conditions. We find the tag-and-copy mechanism (tag the NEs in the source sentence and copy them to the target sentence) improves translation in high-resource settings only. Introducing copying also results in polarizing effects in translating different parts-of-speech (POS). Interestingly, we find that copy accuracy for hypernyms is consistently higher than that of entities. As a way of avoiding "hard" copying and utilizing hypernym in bootstrapping rare entities, we introduced a "soft" tagging mechanism and found consistent improvement in high and low-resource settings.

David Anugraha, Zilu Tang, Lester James V. Miranda et al. · cambridge

Reward models are essential for aligning language model outputs with human preferences, yet existing approaches often lack both controllability and interpretability. These models are typically optimized for narrow objectives, limiting their generalizability to broader downstream tasks. Moreover, their scalar outputs are difficult to interpret without contextual reasoning. To address these limitations, we introduce $\shortmethodname$, a novel reward modeling framework that is rubric-agnostic, generalizable across evaluation dimensions, and provides interpretable, reasoned score assignments. $\shortmethodname$ enables more transparent and flexible evaluation of language models, supporting robust alignment with diverse human values and use cases. Our models, data, and code are available as open source at https://github.com/rubricreward/r3.

Genta Indra Winata, David Anugraha, Emmy Liu et al. · amazon-science

High-quality datasets are fundamental to training and evaluating machine learning models, yet their creation-especially with accurate human annotations-remains a significant challenge. Many dataset paper submissions lack originality, diversity, or rigorous quality control, and these shortcomings are often overlooked during peer review. Submissions also frequently omit essential details about dataset construction and properties. While existing tools such as datasheets aim to promote transparency, they are largely descriptive and do not provide standardized, measurable methods for evaluating data quality. Similarly, metadata requirements at conferences promote accountability but are inconsistently enforced. To address these limitations, this position paper advocates for the integration of systematic, rubric-based evaluation metrics into the dataset review process-particularly as submission volumes continue to grow. We also explore scalable, cost-effective methods for synthetic data generation, including dedicated tools and LLM-as-a-judge approaches, to support more efficient evaluation. As a call to action, we introduce DataRubrics, a structured framework for assessing the quality of both human- and model-generated datasets. Leveraging recent advances in LLM-based evaluation, DataRubrics offers a reproducible, scalable, and actionable solution for dataset quality assessment, enabling both authors and reviewers to uphold higher standards in data-centric research. We also release code to support reproducibility of LLM-based evaluations at https://github.com/datarubrics/datarubrics.

David Anugraha, Shou-Yi Hung, Zilu Tang et al.

Evaluation using Large Language Model (LLM) judges has been widely adopted in English and shown to be effective for automatic evaluation. However, their performance does not generalize well to non-English settings, and it remains unclear what constitutes effective multilingual training for such judges. In this paper, we introduce mR3, a massively multilingual, rubric-agnostic reward reasoning model trained on 72 languages, achieving the broadest language coverage in reward modeling to date. We present a comprehensive study of data and curriculum selection for training to identify effective strategies and data sources for building high-quality reward models, including the integration of target-language reasoning datasets. Our approach attains state-of-the-art performance on multilingual reward model benchmarks, surpassing much larger models (i.e., GPT-OSS-120B) while being up to 9x smaller, and its effectiveness is further confirmed through extensive ablation studies. Our models, data, and code are available as open source at https://github.com/rubricreward/mr3.

Alex Jones, Derry Tanti Wijaya

The explosion of user-generated content (UGC)--e.g. social media posts, comments, and reviews--has motivated the development of NLP applications tailored to these types of informal texts. Prevalent among these applications have been sentiment analysis and machine translation (MT). Grounded in the observation that UGC features highly idiomatic, sentiment-charged language, we propose a decoder-side approach that incorporates automatic sentiment scoring into the MT candidate selection process. We train separate English and Spanish sentiment classifiers, then, using n-best candidates generated by a baseline MT model with beam search, select the candidate that minimizes the absolute difference between the sentiment score of the source sentence and that of the translation, and perform a human evaluation to assess the produced translations. Unlike previous work, we select this minimally divergent translation by considering the sentiment scores of the source sentence and translation on a continuous interval, rather than using e.g. binary classification, allowing for more fine-grained selection of translation candidates. The results of human evaluations show that, in comparison to the open-source MT baseline model on top of which our sentiment-based pipeline is built, our pipeline produces more accurate translations of colloquial, sentiment-heavy source texts.

Alyssa Smith, David Assefa Tofu, Mona Jalal et al.

When journalists cover a news story, they can cover the story from multiple angles or perspectives. A news article written about COVID-19 for example, might focus on personal preventative actions such as mask-wearing, while another might focus on COVID-19's impact on the economy. These perspectives are called "frames," which when used may influence public perception and opinion of the issue. We introduce a Web-based system for analyzing and classifying frames in text documents. Our goal is to make effective tools for automatic frame discovery and labeling based on topic modeling and deep learning widely accessible to researchers from a diverse array of disciplines. To this end, we provide both state-of-the-art pre-trained frame classification models on various issues as well as a user-friendly pipeline for training novel classification models on user-provided corpora. Researchers can submit their documents and obtain frames of the documents. The degree of user involvement is flexible: they can run models that have been pre-trained on select issues; submit labeled documents and train a new model for frame classification; or submit unlabeled documents and obtain potential frames of the documents. The code making up our system is also open-sourced and well-documented, making the system transparent and expandable. The system is available on-line at http://www.openframing.org and via our GitHub page https://github.com/davidatbu/openFraming .

Genta Indra Winata, Frederikus Hudi, Patrick Amadeus Irawan et al.

Vision Language Models (VLMs) often struggle with culture-specific knowledge, particularly in languages other than English and in underrepresented cultural contexts. To evaluate their understanding of such knowledge, we introduce WorldCuisines, a massive-scale benchmark for multilingual and multicultural, visually grounded language understanding. This benchmark includes a visual question answering (VQA) dataset with text-image pairs across 30 languages and dialects, spanning 9 language families and featuring over 1 million data points, making it the largest multicultural VQA benchmark to date. It includes tasks for identifying dish names and their origins. We provide evaluation datasets in two sizes (12k and 60k instances) alongside a training dataset (1 million instances). Our findings show that while VLMs perform better with correct location context, they struggle with adversarial contexts and predicting specific regional cuisines and languages. To support future research, we release a knowledge base with annotated food entries and images along with the VQA data.

David Anugraha, Garry Kuwanto, Lucky Susanto et al.

We present MetaMetrics-MT, an innovative metric designed to evaluate machine translation (MT) tasks by aligning closely with human preferences through Bayesian optimization with Gaussian Processes. MetaMetrics-MT enhances existing MT metrics by optimizing their correlation with human judgments. Our experiments on the WMT24 metric shared task dataset demonstrate that MetaMetrics-MT outperforms all existing baselines, setting a new benchmark for state-of-the-art performance in the reference-based setting. Furthermore, it achieves comparable results to leading metrics in the reference-free setting, offering greater efficiency.

Mohammad Rifqi Farhansyah, Iwan Darmawan, Adryan Kusumawardhana et al.

The Javanese language features a complex system of honorifics that vary according to the social status of the speaker, listener, and referent. Despite its cultural and linguistic significance, there has been limited progress in developing a comprehensive corpus to capture these variations for natural language processing (NLP) tasks. In this paper, we present Unggah-Ungguh, a carefully curated dataset designed to encapsulate the nuances of Unggah-Ungguh Basa, the Javanese speech etiquette framework that dictates the choice of words and phrases based on social hierarchy and context. Using Unggah-Ungguh, we assess the ability of language models (LMs) to process various levels of Javanese honorifics through classification and machine translation tasks. To further evaluate cross-lingual LMs, we conduct machine translation experiments between Javanese (at specific honorific levels) and Indonesian. Additionally, we explore whether LMs can generate contextually appropriate Javanese honorifics in conversation tasks, where the honorific usage should align with the social role and contextual cues. Our findings indicate that current LMs struggle with most honorific levels, exhibitinga bias toward certain honorific tiers.

Garry Kuwanto, Chaitanya Agarwal, Genta Indra Winata et al.

Code-switching, the phenomenon of alternating between two or more languages in a single conversation, presents unique challenges for Natural Language Processing (NLP). Most existing research focuses on either syntactic constraints or neural generation, with few efforts to integrate linguistic theory with large language models (LLMs) for generating natural code-switched text. In this paper, we introduce EZSwitch, a novel framework that combines Equivalence Constraint Theory (ECT) with LLMs to produce linguistically valid and fluent code-switched text. We evaluate our method using both human judgments and automatic metrics, demonstrating a significant improvement in the quality of generated code-switching sentences compared to baseline LLMs. To address the lack of suitable evaluation metrics, we conduct a comprehensive correlation study of various automatic metrics against human scores, revealing that current metrics often fail to capture the nuanced fluency of code-switched text. Additionally, we create CSPref, a human preference dataset based on human ratings and analyze model performance across ``hard`` and ``easy`` examples. Our findings indicate that incorporating linguistic constraints into LLMs leads to more robust and human-aligned generation, paving the way for scalable code-switching text generation across diverse language pairs.

Ryandito Diandaru, Lucky Susanto, Zilu Tang et al.

Large Language Models (LLMs) demonstrate strong machine translation capabilities on languages they are trained on. However, the impact of factors beyond training data size on translation performance remains a topic of debate, especially concerning languages not directly encountered during training. Our study delves into Llama2's translation capabilities. By modeling a linear relationship between linguistic feature distances and machine translation scores, we ask ourselves if there are potentially better central languages for LLMs other than English. Our experiments show that the 7B Llama2 model yields above 10 BLEU when translating into all languages it has seen, which rarely happens for languages it has not seen. Most translation improvements into unseen languages come from scaling up the model size rather than instruction tuning or increasing shot count. Furthermore, our correlation analysis reveals that syntactic similarity is not the only linguistic factor that strongly correlates with machine translation scores. Interestingly, we discovered that under specific circumstances, some languages (e.g. Swedish, Catalan), despite having significantly less training data, exhibit comparable correlation levels to English. These insights challenge the prevailing landscape of LLMs, suggesting that models centered around languages other than English could provide a more efficient foundation for multilingual applications.

Muhammad Farid Adilazuarda, Musa Izzanardi Wijanarko, Lucky Susanto et al.

Indonesia is rich in languages and scripts. However, most NLP progress has been made using romanized text. In this paper, we present NusaAksara, a novel public benchmark for Indonesian languages that includes their original scripts. Our benchmark covers both text and image modalities and encompasses diverse tasks such as image segmentation, OCR, transliteration, translation, and language identification. Our data is constructed by human experts through rigorous steps. NusaAksara covers 8 scripts across 7 languages, including low-resource languages not commonly seen in NLP benchmarks. Although unsupported by Unicode, the Lampung script is included in this dataset. We benchmark our data across several models, from LLMs and VLMs such as GPT-4o, Llama 3.2, and Aya 23 to task-specific systems such as PP-OCR and LangID, and show that most NLP technologies cannot handle Indonesia's local scripts, with many achieving near-zero performance.

Zilu Tang, Afra Feyza Akyürek, Ekin Akyürek et al. · mit

A prominent issue in aligning language models (LMs) to personalized preferences is underspecification -- the lack of information from users about their preferences. A popular trend of injecting such specification is adding a prefix (e.g. prior relevant conversations) to the current user's conversation to steer preference distribution. Most methods passively model personal preferences with prior example preferences pairs. We ask whether models benefit from actively inferring preference descriptions, and address this question by creating a synthetic personalized alignment dataset based on famous people with known public preferences. We then test how effective finetuned 1-8B size models are at inferring and aligning to personal preferences. Results show that higher-quality active prefixes lead to better generalization, more contextually faithful models, and less systematic biases across different protected attributes. All our results suggest active alignment can lead to a more controllable and efficient path for personalized alignment.

Lucky Susanto, Musa Wijanarko, Prasetia Pratama et al.

Polarization is defined as divisive opinions held by two or more groups on substantive issues. As the world's third-largest democracy, Indonesia faces growing concerns about the interplay between political polarization and online toxicity, which is often directed at vulnerable minority groups. Despite the importance of this issue, previous NLP research has not fully explored the relationship between toxicity and polarization. To bridge this gap, we present a novel multi-label Indonesian dataset that incorporates toxicity, polarization, and annotator demographic information. Benchmarking this dataset using BERT-base models and large language models (LLMs) shows that polarization information enhances toxicity classification, and vice versa. Furthermore, providing demographic information significantly improves the performance of polarization classification.

Sarah Bonna, Yu-Cheng Huang, Ekaterina Novozhilova et al.

Ethical intervention prompting has emerged as a tool to counter demographic biases of text-to-image generative AI models. Existing solutions either require to retrain the model or struggle to generate images that reflect desired distributions on gender and race. We propose an inference-time process called DebiasPI for Debiasing-by-Prompt-Iteration that provides prompt intervention by enabling the user to control the distributions of individuals' demographic attributes in image generation. DebiasPI keeps track of which attributes have been generated either by probing the internal state of the model or by using external attribute classifiers. Its control loop guides the text-to-image model to select not yet sufficiently represented attributes, With DebiasPI, we were able to create images with equal representations of race and gender that visualize challenging concepts of news headlines. We also experimented with the attributes age, body type, profession, and skin tone, and measured how attributes change when our intervention prompt targets the distribution of an unrelated attribute type. We found, for example, if the text-to-image model is asked to balance racial representation, gender representation improves but the skin tone becomes less diverse. Attempts to cover a wide range of skin colors with various intervention prompts showed that the model struggles to generate the palest skin tones. We conducted various ablation studies, in which we removed DebiasPI's attribute control, that reveal the model's propensity to generate young, male characters. It sometimes visualized career success by generating two-panel images with a pre-success dark-skinned person becoming light-skinned with success, or switching gender from pre-success female to post-success male, thus further motivating ethical intervention prompting with DebiasPI.

Mohammad Rifqi Farhansyah, Muhammad Zuhdi Fikri Johari, Afinzaki Amiral et al.

Indonesia is one of the most diverse countries linguistically. However, despite this linguistic diversity, Indonesian languages remain underrepresented in Natural Language Processing (NLP) research and technologies. In the past two years, several efforts have been conducted to construct NLP resources for Indonesian languages. However, most of these efforts have been focused on creating manual resources thus difficult to scale to more languages. Although many Indonesian languages do not have a web presence, locally there are resources that document these languages well in printed forms such as books, magazines, and newspapers. Digitizing these existing resources will enable scaling of Indonesian language resource construction to many more languages. In this paper, we propose an alternative method of creating datasets by digitizing documents, which have not previously been used to build digital language resources in Indonesia. DriveThru is a platform for extracting document content utilizing Optical Character Recognition (OCR) techniques in its system to provide language resource building with less manual effort and cost. This paper also studies the utility of current state-of-the-art LLM for post-OCR correction to show the capability of increasing the character accuracy rate (CAR) and word accuracy rate (WAR) compared to off-the-shelf OCR.

Lucky Susanto, Musa Izzanardi Wijanarko, Prasetia Anugrah Pratama et al.

Hate speech poses a significant threat to social harmony. Over the past two years, Indonesia has seen a ten-fold increase in the online hate speech ratio, underscoring the urgent need for effective detection mechanisms. However, progress is hindered by the limited availability of labeled data for Indonesian texts. The condition is even worse for marginalized minorities, such as Shia, LGBTQ, and other ethnic minorities because hate speech is underreported and less understood by detection tools. Furthermore, the lack of accommodation for subjectivity in current datasets compounds this issue. To address this, we introduce IndoToxic2024, a comprehensive Indonesian hate speech and toxicity classification dataset. Comprising 43,692 entries annotated by 19 diverse individuals, the dataset focuses on texts targeting vulnerable groups in Indonesia, specifically during the hottest political event in the country: the presidential election. We establish baselines for seven binary classification tasks, achieving a macro-F1 score of 0.78 with a BERT model (IndoBERTweet) fine-tuned for hate speech classification. Furthermore, we demonstrate how incorporating demographic information can enhance the zero-shot performance of the large language model, gpt-3.5-turbo. However, we also caution that an overemphasis on demographic information can negatively impact the fine-tuned model performance due to data fragmentation.

Isidora Chara Tourni, Lei Guo, Hengchang Hu et al.

News media structure their reporting of events or issues using certain perspectives. When describing an incident involving gun violence, for example, some journalists may focus on mental health or gun regulation, while others may emphasize the discussion of gun rights. Such perspectives are called \say{frames} in communication research. We study, for the first time, the value of combining lead images and their contextual information with text to identify the frame of a given news article. We observe that using multiple modes of information(article- and image-derived features) improves prediction of news frames over any single mode of information when the images are relevant to the frames of the headlines. We also observe that frame image relevance is related to the ease of conveying frames via images, which we call frame concreteness. Additionally, we release the first multimodal news framing dataset related to gun violence in the U.S., curated and annotated by communication researchers. The dataset will allow researchers to further examine the use of multiple information modalities for studying media framing.

Derry Wijaya, Brendan Callahan, John Hewitt et al.

Bilingual Lexicon Induction is the task of learning word translations without bilingual parallel corpora. We model this task as a matrix completion problem, and present an effective and extendable framework for completing the matrix. This method harnesses diverse bilingual and monolingual signals, each of which may be incomplete or noisy. Our model achieves state-of-the-art performance for both high and low resource languages.

Afra Feyza Akyürek, Ekin Akyürek, Leshem Choshen et al.

While language models (LMs) can sometimes generate factually correct text and estimate truth values of individual claims, these generally do not reflect a globally coherent, manipulable model of the world. As a consequence, current LMs also generate incorrect or nonsensical content, and are difficult to edit and bring up to date. We present a method called Deductive Closure Training (DCT) that uses LMs themselves to identify implications of (and contradictions within) the text that they generate, yielding an efficient self-supervised procedure for improving LM factuality. Given a collection of seed documents, DCT prompts LMs to generate additional text implied by these documents, reason globally about the correctness of this generated text, and finally fine-tune on text inferred to be correct. Given seed documents from a trusted source, DCT provides a tool for supervised model updating; if seed documents are sampled from the LM itself, DCT enables fully unsupervised fine-tuning for improved coherence and accuracy. Across the CREAK, MQUaKE, and Reversal Curse datasets, supervised DCT improves LM fact verification and text generation accuracy by 3-26%; on CREAK fully unsupervised DCT improves verification accuracy by 12%. These results show that LMs' reasoning capabilities during inference can be leveraged during training to improve their reliability.

Isidora Chara Tourni, Derry Wijaya

Unsupervised Neural Machine Translation (UNMT) focuses on improving NMT results under the assumption there is no human translated parallel data, yet little work has been done so far in highlighting its advantages compared to supervised methods and analyzing its output in aspects other than translation accuracy. We focus on three very diverse languages, French, Gujarati, and Kazakh, and train bilingual NMT models, to and from English, with various levels of supervision, in high- and low- resource setups, measure quality of the NMT output and compare the generated sequences' word order and semantic similarity to source and reference sentences. We also use Layer-wise Relevance Propagation to evaluate the source and target sentences' contribution to the result, expanding the findings of previous works to the UNMT paradigm.

Afra Feyza Akyürek, Ekin Akyürek, Aman Madaan et al.

Despite their unprecedented success, even the largest language models make mistakes. Similar to how humans learn and improve using feedback, previous work proposed providing language models with natural language feedback to guide them in repairing their outputs. Because human-generated critiques are expensive to obtain, researchers have devised learned critique generators in lieu of human critics while assuming one can train downstream models to utilize generated feedback. However, this approach does not apply to black-box or limited access models such as ChatGPT, as they cannot be fine-tuned. Moreover, in the era of large general-purpose language agents, fine-tuning is neither computationally nor spatially efficient as it results in multiple copies of the network. In this work, we introduce RL4F (Reinforcement Learning for Feedback), a multi-agent collaborative framework where the critique generator is trained to maximize end-task performance of GPT-3, a fixed model more than 200 times its size. RL4F produces critiques that help GPT-3 revise its outputs. We study three datasets for action planning, summarization and alphabetization and show relative improvements up to 10% in multiple text similarity metrics over other learned, retrieval-augmented or prompting-based critique generators.

Wenda Qin, Teruhisa Misu, Derry Wijaya

Vision-and-Language Navigation (VLN) is a challenging task in the field of artificial intelligence. Although massive progress has been made in this task over the past few years attributed to breakthroughs in deep vision and language models, it remains tough to build VLN models that can generalize as well as humans. In this paper, we provide a new perspective to improve VLN models. Based on our discovery that snapshots of the same VLN model behave significantly differently even when their success rates are relatively the same, we propose a snapshot-based ensemble solution that leverages predictions among multiple snapshots. Constructed on the snapshots of the existing state-of-the-art (SOTA) model $\circlearrowright$BERT and our past-action-aware modification, our proposed ensemble achieves the new SOTA performance in the R2R dataset challenge in Navigation Error (NE) and Success weighted by Path Length (SPL).

Afra Feyza Akyürek, Ekin Akyürek, Derry Tanti Wijaya et al.

Few-shot class incremental learning -- the problem of updating a trained classifier to discriminate among an expanded set of classes with limited labeled data -- is a key challenge for machine learning systems deployed in non-stationary environments. Existing approaches to the problem rely on complex model architectures and training procedures that are difficult to tune and re-use. In this paper, we present an extremely simple approach that enables the use of ordinary logistic regression classifiers for few-shot incremental learning. The key to this approach is a new family of subspace regularization schemes that encourage weight vectors for new classes to lie close to the subspace spanned by the weights of existing classes. When combined with pretrained convolutional feature extractors, logistic regression models trained with subspace regularization outperform specialized, state-of-the-art approaches to few-shot incremental image classification by up to 22% on the miniImageNet dataset. Because of its simplicity, subspace regularization can be straightforwardly extended to incorporate additional background information about the new classes (including class names and descriptions specified in natural language); these further improve accuracy by up to 2%. Our results show that simple geometric regularization of class representations offers an effective tool for continual learning.

Mohammad Sadegh Rasooli, Chris Callison-Burch, Derry Tanti Wijaya

We present a simple but effective approach for leveraging Wikipedia for neural machine translation as well as cross-lingual tasks of image captioning and dependency parsing without using any direct supervision from external parallel data or supervised models in the target language. We show that first sentences and titles of linked Wikipedia pages, as well as cross-lingual image captions, are strong signals for a seed parallel data to extract bilingual dictionaries and cross-lingual word embeddings for mining parallel text from Wikipedia. Our final model achieves high BLEU scores that are close to or sometimes higher than strong supervised baselines in low-resource languages; e.g. supervised BLEU of 4.0 versus 12.1 from our model in English-to-Kazakh. Moreover, we tailor our wikily supervised translation models to unsupervised image captioning, and cross-lingual dependency parser transfer. In image captioning, we train a multi-tasking machine translation and image captioning pipeline for Arabic and English from which the Arabic training data is a translated version of the English captioning data, using our wikily-supervised translation models. Our captioning results on Arabic are slightly better than that of its supervised model. In dependency parsing, we translate a large amount of monolingual text, and use it as artificial training data in an annotation projection framework. We show that our model outperforms recent work on cross-lingual transfer of dependency parsers.

Garry Kuwanto, Afra Feyza Akyürek, Isidora Chara Tourni et al.

We conduct an empirical study of neural machine translation (NMT) for truly low-resource languages, and propose a training curriculum fit for cases when both parallel training data and compute resource are lacking, reflecting the reality of most of the world's languages and the researchers working on these languages. Previously, unsupervised NMT, which employs back-translation (BT) and auto-encoding (AE) tasks has been shown barren for low-resource languages. We demonstrate that leveraging comparable data and code-switching as weak supervision, combined with BT and AE objectives, result in remarkable improvements for low-resource languages even when using only modest compute resources. The training curriculum proposed in this work achieves BLEU scores that improve over supervised NMT trained on the same backbone architecture by +12.2 BLEU for English to Gujarati and +3.7 BLEU for English to Kazakh, showcasing the potential of weakly-supervised NMT for the low-resource languages. When trained on supervised data, our training curriculum achieves a new state-of-the-art result on the Somali dataset (BLEU of 29.3 for Somali to English). We also observe that adding more time and GPUs to training can further improve performance, which underscores the importance of reporting compute resource usage in MT research.

Alex Jones, Derry Tanti Wijaya

Obtaining high-quality parallel corpora is of paramount importance for training NMT systems. However, as many language pairs lack adequate gold-standard training data, a popular approach has been to mine so-called "pseudo-parallel" sentences from paired documents in two languages. In this paper, we outline some problems with current methods, propose computationally economical solutions to those problems, and demonstrate success with novel methods on the Tatoeba similarity search benchmark and on a downstream task, namely NMT. We uncover the effect of resource-related factors (i.e. how much monolingual/bilingual data is available for a given language) on the optimal choice of bitext mining approach, and echo problems with the oft-used BUCC dataset that have been observed by others. We make the code and data used for our experiments publicly available.

Andrea Burns, Donghyun Kim, Derry Wijaya et al.

Current multilingual vision-language models either require a large number of additional parameters for each supported language, or suffer performance degradation as languages are added. In this paper, we propose a Scalable Multilingual Aligned Language Representation (SMALR) that supports many languages with few model parameters without sacrificing downstream task performance. SMALR learns a fixed size language-agnostic representation for most words in a multilingual vocabulary, keeping language-specific features for just a few. We use a masked cross-language modeling loss to align features with context from other languages. Additionally, we propose a cross-lingual consistency module that ensures predictions made for a query and its machine translation are comparable. The effectiveness of SMALR is demonstrated with ten diverse languages, over twice the number supported in vision-language tasks to date. We evaluate on multilingual image-sentence retrieval and outperform prior work by 3-4% with less than 1/5th the training parameters compared to other word embedding methods.

Siyi Liu, Ziang Leng, Derry Wijaya

Mirroring is the behavior in which one person subconsciously imitates the gesture, speech pattern, or attitude of another. In conversations, mirroring often signals the speakers enjoyment and engagement in their communication. In chatbots, methods have been proposed to add personas to the chatbots and to train them to speak or to shift their dialogue style to that of the personas. However, they often require a large dataset consisting of dialogues of the target personalities to train. In this work, we explore a method that can learn to mirror the speaking styles of a person incrementally. Our method extracts ngrams that capture a persons speaking styles and uses the ngrams to create patterns for transforming sentences to the persons speaking styles. Our experiments show that our method is able to capture patterns of speaking style that can be used to transform regular sentences into sentences with the target style.

Joao Sedoc, Derry Wijaya, Masoud Rouhizadeh et al.

Hand-built verb clusters such as the widely used Levin classes (Levin, 1993) have proved useful, but have limited coverage. Verb classes automatically induced from corpus data such as those from VerbKB (Wijaya, 2016), on the other hand, can give clusters with much larger coverage, and can be adapted to specific corpora such as Twitter. We present a method for clustering the outputs of VerbKB: verbs with their multiple argument types, e.g. "marry(person, person)", "feel(person, emotion)." We make use of a novel low-dimensional embedding of verbs and their arguments to produce high quality clusters in which the same verb can be in different clusters depending on its argument type. The resulting verb clusters do a better job than hand-built clusters of predicting sarcasm, sentiment, and locus of control in tweets.