Francesco Barbieri

Jose Camacho-Collados, Kiamehr Rezaee, Talayeh Riahi et al. · deepmind

In this paper we present TweetNLP, an integrated platform for Natural Language Processing (NLP) in social media. TweetNLP supports a diverse set of NLP tasks, including generic focus areas such as sentiment analysis and named entity recognition, as well as social media-specific tasks such as emoji prediction and offensive language identification. Task-specific systems are powered by reasonably-sized Transformer-based language models specialized on social media text (in particular, Twitter) which can be run without the need for dedicated hardware or cloud services. The main contributions of TweetNLP are: (1) an integrated Python library for a modern toolkit supporting social media analysis using our various task-specific models adapted to the social domain; (2) an interactive online demo for codeless experimentation using our models; and (3) a tutorial covering a wide variety of typical social media applications.

Asahi Ushio, Leonardo Neves, Vitor Silva et al.

Recent progress in language model pre-training has led to important improvements in Named Entity Recognition (NER). Nonetheless, this progress has been mainly tested in well-formatted documents such as news, Wikipedia, or scientific articles. In social media the landscape is different, in which it adds another layer of complexity due to its noisy and dynamic nature. In this paper, we focus on NER in Twitter, one of the largest social media platforms, and construct a new NER dataset, TweetNER7, which contains seven entity types annotated over 11,382 tweets from September 2019 to August 2021. The dataset was constructed by carefully distributing the tweets over time and taking representative trends as a basis. Along with the dataset, we provide a set of language model baselines and perform an analysis on the language model performance on the task, especially analyzing the impact of different time periods. In particular, we focus on three important temporal aspects in our analysis: short-term degradation of NER models over time, strategies to fine-tune a language model over different periods, and self-labeling as an alternative to lack of recently-labeled data. TweetNER7 is released publicly (https://huggingface.co/datasets/tner/tweetner7) along with the models fine-tuned on it.

Jiaxin Pei, Vítor Silva, Maarten Bos et al. · stanford

We propose MINT, a new Multilingual INTimacy analysis dataset covering 13,372 tweets in 10 languages including English, French, Spanish, Italian, Portuguese, Korean, Dutch, Chinese, Hindi, and Arabic. We benchmarked a list of popular multilingual pre-trained language models. The dataset is released along with the SemEval 2023 Task 9: Multilingual Tweet Intimacy Analysis (https://sites.google.com/umich.edu/semeval-2023-tweet-intimacy).

Dimosthenis Antypas, Asahi Ushio, Francesco Barbieri et al. · stanford

Despite its relevance, the maturity of NLP for social media pales in comparison with general-purpose models, metrics and benchmarks. This fragmented landscape makes it hard for the community to know, for instance, given a task, which is the best performing model and how it compares with others. To alleviate this issue, we introduce a unified benchmark for NLP evaluation in social media, SuperTweetEval, which includes a heterogeneous set of tasks and datasets combined, adapted and constructed from scratch. We benchmarked the performance of a wide range of models on SuperTweetEval and our results suggest that, despite the recent advances in language modelling, social media remains challenging.

Zhihan Zhang, Dong-Ho Lee, Yuwei Fang et al.

Instruction tuning has remarkably advanced large language models (LLMs) in understanding and responding to diverse human instructions. Despite the success in high-resource languages, its application in lower-resource ones faces challenges due to the imbalanced foundational abilities of LLMs across different languages, stemming from the uneven language distribution in their pre-training data. To tackle this issue, we propose pivot language guided generation (PLUG), an approach that utilizes a high-resource language, primarily English, as the pivot to enhance instruction tuning in lower-resource languages. It trains the model to first process instructions in the pivot language, and then produce responses in the target language. To evaluate our approach, we introduce a benchmark, X-AlpacaEval, of instructions in 4 languages (Chinese, Korean, Italian, and Spanish), each annotated by professional translators. Our approach demonstrates a significant improvement in the instruction-following abilities of LLMs by 29% on average, compared to directly responding in the target language alone. Further experiments validate the versatility of our approach by employing alternative pivot languages beyond English to assist languages where LLMs exhibit lower proficiency. Our code and data are available at https://github.com/ytyz1307zzh/PLUG.

Ligong Han, Jian Ren, Hsin-Ying Lee et al.

Most methods for conditional video synthesis use a single modality as the condition. This comes with major limitations. For example, it is problematic for a model conditioned on an image to generate a specific motion trajectory desired by the user since there is no means to provide motion information. Conversely, language information can describe the desired motion, while not precisely defining the content of the video. This work presents a multimodal video generation framework that benefits from text and images provided jointly or separately. We leverage the recent progress in quantized representations for videos and apply a bidirectional transformer with multiple modalities as inputs to predict a discrete video representation. To improve video quality and consistency, we propose a new video token trained with self-learning and an improved mask-prediction algorithm for sampling video tokens. We introduce text augmentation to improve the robustness of the textual representation and diversity of generated videos. Our framework can incorporate various visual modalities, such as segmentation masks, drawings, and partially occluded images. It can generate much longer sequences than the one used for training. In addition, our model can extract visual information as suggested by the text prompt, e.g., "an object in image one is moving northeast", and generate corresponding videos. We run evaluations on three public datasets and a newly collected dataset labeled with facial attributes, achieving state-of-the-art generation results on all four.

Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados et al.

Social media platforms host discussions about a wide variety of topics that arise everyday. Making sense of all the content and organising it into categories is an arduous task. A common way to deal with this issue is relying on topic modeling, but topics discovered using this technique are difficult to interpret and can differ from corpus to corpus. In this paper, we present a new task based on tweet topic classification and release two associated datasets. Given a wide range of topics covering the most important discussion points in social media, we provide training and testing data from recent time periods that can be used to evaluate tweet classification models. Moreover, we perform a quantitative evaluation and analysis of current general- and domain-specific language models on the task, which provide more insights on the challenges and nature of the task.

Daniel Loureiro, Aminette D'Souza, Areej Nasser Muhajab et al.

Language evolves over time, and word meaning changes accordingly. This is especially true in social media, since its dynamic nature leads to faster semantic shifts, making it challenging for NLP models to deal with new content and trends. However, the number of datasets and models that specifically address the dynamic nature of these social platforms is scarce. To bridge this gap, we present TempoWiC, a new benchmark especially aimed at accelerating research in social media-based meaning shift. Our results show that TempoWiC is a challenging benchmark, even for recently-released language models specialized in social media.

Daniel Loureiro, Kiamehr Rezaee, Talayeh Riahi et al.

This paper introduces a large collection of time series data derived from Twitter, postprocessed using word embedding techniques, as well as specialized fine-tuned language models. This data comprises the past five years and captures changes in n-gram frequency, similarity, sentiment and topic distribution. The interface built on top of this data enables temporal analysis for detecting and characterizing shifts in meaning, including complementary information to trending metrics, such as sentiment and topic association over time. We release an online demo for easy experimentation, and we share code and the underlying aggregated data for future work. In this paper, we also discuss three case studies unlocked thanks to our platform, showcasing its potential for temporal linguistic analysis.

Shashwat Goel, Rishi Hazra, Dulhan Jayalath et al.

AI co-scientists are emerging as a tool to assist human researchers in achieving their research goals. A crucial feature of these AI co-scientists is the ability to generate a research plan given a set of aims and constraints. The plan may be used by researchers for brainstorming, or may even be implemented after further refinement. However, language models currently struggle to generate research plans that follow all constraints and implicit requirements. In this work, we study how to leverage the vast corpus of existing research papers to train language models that generate better research plans. We build a scalable, diverse training corpus by automatically extracting research goals and goal-specific grading rubrics from papers across several domains. We then train models for research plan generation via reinforcement learning with self-grading. A frozen copy of the initial policy acts as the grader during training, with the rubrics creating a generator-verifier gap that enables improvements without external human supervision. To validate this approach, we conduct a study with human experts for machine learning research goals, spanning 225 hours. The experts prefer plans generated by our finetuned Qwen3-30B-A3B model over the initial model for 70% of research goals, and approve 84% of the automatically extracted goal-specific grading rubrics. To assess generality, we also extend our approach to research goals from medical papers, and new arXiv preprints, evaluating with a jury of frontier models. Our finetuning yields 12-22% relative improvements and significant cross-domain generalization, proving effective even in problem settings like medical research where execution feedback is infeasible. Together, these findings demonstrate the potential of a scalable, automated training recipe as a step towards improving general AI co-scientists.

Heinrich Peters, Yozen Liu, Francesco Barbieri et al.

The success of online social platforms hinges on their ability to predict and understand user behavior at scale. Here, we present data suggesting that context-aware modeling approaches may offer a holistic yet lightweight and potentially privacy-preserving representation of user engagement on online social platforms. Leveraging deep LSTM neural networks to analyze more than 100 million Snapchat sessions from almost 80.000 users, we demonstrate that patterns of active and passive use are predictable from past behavior (R2=0.345) and that the integration of context features substantially improves predictive performance compared to the behavioral baseline model (R2=0.522). Features related to smartphone connectivity status, location, temporal context, and weather were found to capture non-redundant variance in user engagement relative to features derived from histories of in-app behaviors. Further, we show that a large proportion of variance can be accounted for with minimal behavioral histories if momentary context is considered (R2=0.442). These results indicate the potential of context-aware approaches for making models more efficient and privacy-preserving by reducing the need for long data histories. Finally, we employ model explainability techniques to glean preliminary insights into the underlying behavioral mechanisms. Our findings are consistent with the notion of context-contingent, habit-driven patterns of active and passive use, underscoring the value of contextualized representations of user behavior for predicting user engagement on social platforms.

Dimosthenis Antypas, Indira Sen, Carla Perez-Almendros et al.

The detection of sensitive content in large datasets is crucial for ensuring that shared and analysed data is free from harmful material. However, current moderation tools, such as external APIs, suffer from limitations in customisation, accuracy across diverse sensitive categories, and privacy concerns. Additionally, existing datasets and open-source models focus predominantly on toxic language, leaving gaps in detecting other sensitive categories such as substance abuse or self-harm. In this paper, we put forward a unified dataset tailored for social media content moderation across six sensitive categories: conflictual language, profanity, sexually explicit material, drug-related content, self-harm, and spam. By collecting and annotating data with consistent retrieval strategies and guidelines, we address the shortcomings of previous focalised research. Our analysis demonstrates that fine-tuning large language models (LLMs) on this novel dataset yields significant improvements in detection performance compared to open off-the-shelf models such as LLaMA, and even proprietary OpenAI models, which underperform by 10-15% overall. This limitation is even more pronounced on popular moderation APIs, which cannot be easily tailored to specific sensitive content categories, among others.

Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov et al.

Existing works on long-term open-domain dialogues focus on evaluating model responses within contexts spanning no more than five chat sessions. Despite advancements in long-context large language models (LLMs) and retrieval augmented generation (RAG) techniques, their efficacy in very long-term dialogues remains unexplored. To address this research gap, we introduce a machine-human pipeline to generate high-quality, very long-term dialogues by leveraging LLM-based agent architectures and grounding their dialogues on personas and temporal event graphs. Moreover, we equip each agent with the capability of sharing and reacting to images. The generated conversations are verified and edited by human annotators for long-range consistency and grounding to the event graphs. Using this pipeline, we collect LoCoMo, a dataset of very long-term conversations, each encompassing 300 turns and 9K tokens on avg., over up to 35 sessions. Based on LoCoMo, we present a comprehensive evaluation benchmark to measure long-term memory in models, encompassing question answering, event summarization, and multi-modal dialogue generation tasks. Our experimental results indicate that LLMs exhibit challenges in understanding lengthy conversations and comprehending long-range temporal and causal dynamics within dialogues. Employing strategies like long-context LLMs or RAG can offer improvements but these models still substantially lag behind human performance.

Dong-Ho Lee, Adyasha Maharana, Jay Pujara et al.

Long-term, open-domain dialogue capabilities are essential for chatbots aiming to recall past interactions and demonstrate emotional intelligence (EI). Yet, most existing research relies on synthetic, LLM-generated data, leaving open questions about real-world conversational patterns. To address this gap, we introduce REALTALK, a 21-day corpus of authentic messaging app dialogues, providing a direct benchmark against genuine human interactions. We first conduct a dataset analysis, focusing on EI attributes and persona consistency to understand the unique challenges posed by real-world dialogues. By comparing with LLM-generated conversations, we highlight key differences, including diverse emotional expressions and variations in persona stability that synthetic dialogues often fail to capture. Building on these insights, we introduce two benchmark tasks: (1) persona simulation where a model continues a conversation on behalf of a specific user given prior dialogue context; and (2) memory probing where a model answers targeted questions requiring long-term memory of past interactions. Our findings reveal that models struggle to simulate a user solely from dialogue history, while fine-tuning on specific user chats improves persona emulation. Additionally, existing models face significant challenges in recalling and leveraging long-term context within real-world conversations.

Zhihan Zhou, Qixiang Fang, Leonardo Neves et al.

User embeddings play a crucial role in user engagement forecasting and personalized services. Recent advances in sequence modeling have sparked interest in learning user embeddings from behavioral data. Yet behavior-based user embedding learning faces the unique challenge of dynamic user modeling. As users continuously interact with the apps, user embeddings should be periodically updated to account for users' recent and long-term behavior patterns. Existing methods highly rely on stateless sequence models that lack memory of historical behavior. They have to either discard historical data and use only the most recent data or reprocess the old and new data jointly. Both cases incur substantial computational overhead. To address this limitation, we introduce User Stateful Embedding (USE). USE generates user embeddings and reflects users' evolving behaviors without the need for exhaustive reprocessing by storing previous model states and revisiting them in the future. Furthermore, we introduce a novel training objective named future W-behavior prediction to transcend the limitations of next-token prediction by forecasting a broader horizon of upcoming user behaviors. By combining it with the Same User Prediction, a contrastive learning-based objective that predicts whether different segments of behavior sequences belong to the same user, we further improve the embeddings' distinctiveness and representativeness. We conducted experiments on 8 downstream tasks using Snapchat users' behavioral logs in both static (i.e., fixed user behavior sequences) and dynamic (i.e., periodically updated user behavior sequences) settings. We demonstrate USE's superior performance over established baselines. The results underscore USE's effectiveness and efficiency in integrating historical and recent user behavior sequences into user embeddings in dynamic user modeling.

William F. Shen, Xinchi Qiu, Chenxi Whitehouse et al.

Recently, rubrics have been used to guide LLM judges in capturing subjective, nuanced, multi-dimensional human preferences, and have been extended from evaluation to reward signals for reinforcement fine-tuning (RFT). However, rubric generation remains hard to control: rubrics often lack coverage, conflate dimensions, misalign preference direction, and contain redundant or highly correlated criteria, degrading judge accuracy and producing suboptimal rewards during RFT. We propose RRD, a principled framework for rubric refinement built on a recursive decompose-filter cycle. RRD decomposes coarse rubrics into fine-grained, discriminative criteria, expanding coverage while sharpening separation between responses. A complementary filtering mechanism removes misaligned and redundant rubrics, and a correlation-aware weighting scheme prevents over-representing highly correlated criteria, yielding rubric sets that are informative, comprehensive, and non-redundant. Empirically, RRD delivers large, consistent gains across both evaluation and training: it improves preference-judgment accuracy on JudgeBench and PPE for both GPT-4o and Llama3.1-405B judges, achieving top performance in all settings with up to +17.7 points on JudgeBench. When used as the reward source for RFT on WildChat, it yields substantially stronger and more stable learning signals, boosting reward by up to 160% (Qwen3-4B) and 60% (Llama3.1-8B) versus 10-20% for prior rubric baselines, with gains that transfer to HealthBench-Hard and BiGGen Bench. Overall, RRD establishes recursive rubric refinement as a scalable and interpretable foundation for LLM judging and reward modeling in open-ended domains.

Agostina Calabrese, Leonardo Neves, Neil Shah et al.

Content moderators play a key role in keeping the conversation on social media healthy. While the high volume of content they need to judge represents a bottleneck to the moderation pipeline, no studies have explored how models could support them to make faster decisions. There is, by now, a vast body of research into detecting hate speech, sometimes explicitly motivated by a desire to help improve content moderation, but published research using real content moderators is scarce. In this work we investigate the effect of explanations on the speed of real-world moderators. Our experiments show that while generic explanations do not affect their speed and are often ignored, structured explanations lower moderators' decision making time by 7.4%.

Daniel Loureiro, Francesco Barbieri, Leonardo Neves et al.

Despite its importance, the time variable has been largely neglected in the NLP and language model literature. In this paper, we present TimeLMs, a set of language models specialized on diachronic Twitter data. We show that a continual learning strategy contributes to enhancing Twitter-based language models' capacity to deal with future and out-of-distribution tweets, while making them competitive with standardized and more monolithic benchmarks. We also perform a number of qualitative analyses showing how they cope with trends and peaks in activity involving specific named entities or concept drift.

Francesco Barbieri, Luis Espinosa Anke, Jose Camacho-Collados

Language models are ubiquitous in current NLP, and their multilingual capacity has recently attracted considerable attention. However, current analyses have almost exclusively focused on (multilingual variants of) standard benchmarks, and have relied on clean pre-training and task-specific corpora as multilingual signals. In this paper, we introduce XLM-T, a model to train and evaluate multilingual language models in Twitter. In this paper we provide: (1) a new strong multilingual baseline consisting of an XLM-R (Conneau et al. 2020) model pre-trained on millions of tweets in over thirty languages, alongside starter code to subsequently fine-tune on a target task; and (2) a set of unified sentiment analysis Twitter datasets in eight different languages and a XLM-T model fine-tuned on them.

Brihi Joshi, Neil Shah, Francesco Barbieri et al.

Contextual embeddings derived from transformer-based neural language models have shown state-of-the-art performance for various tasks such as question answering, sentiment analysis, and textual similarity in recent years. Extensive work shows how accurately such models can represent abstract, semantic information present in text. In this expository work, we explore a tangent direction and analyze such models' performance on tasks that require a more granular level of representation. We focus on the problem of textual similarity from two perspectives: matching documents on a granular level (requiring embeddings to capture fine-grained attributes in the text), and an abstract level (requiring embeddings to capture overall textual semantics). We empirically demonstrate, across two datasets from different domains, that despite high performance in abstract document matching as expected, contextual embeddings are consistently (and at times, vastly) outperformed by simple baselines like TF-IDF for more granular tasks. We then propose a simple but effective method to incorporate TF-IDF into models that use contextual embeddings, achieving relative improvements of up to 36% on granular tasks.

Xisen Jin, Francesco Barbieri, Brendan Kennedy et al.

Fine-tuned language models have been shown to exhibit biases against protected groups in a host of modeling tasks such as text classification and coreference resolution. Previous works focus on detecting these biases, reducing bias in data representations, and using auxiliary training objectives to mitigate bias during fine-tuning. Although these techniques achieve bias reduction for the task and domain at hand, the effects of bias mitigation may not directly transfer to new tasks, requiring additional data collection and customized annotation of sensitive attributes, and re-evaluation of appropriate fairness metrics. We explore the feasibility and benefits of upstream bias mitigation (UBM) for reducing bias on downstream tasks, by first applying bias mitigation to an upstream model through fine-tuning and subsequently using it for downstream fine-tuning. We find, in extensive experiments across hate speech detection, toxicity detection, occupation prediction, and coreference resolution tasks over various bias factors, that the effects of UBM are indeed transferable to new downstream tasks or domains via fine-tuning, creating less biased downstream models than directly fine-tuning on the downstream task or transferring from a vanilla upstream model. Though challenges remain, we show that UBM promises more efficient and accessible bias mitigation in LM fine-tuning.

Francesco Barbieri, Jose Camacho-Collados, Leonardo Neves et al.

The experimental landscape in natural language processing for social media is too fragmented. Each year, new shared tasks and datasets are proposed, ranging from classics like sentiment analysis to irony detection or emoji prediction. Therefore, it is unclear what the current state of the art is, as there is no standardized evaluation protocol, neither a strong set of baselines trained on such domain-specific data. In this paper, we propose a new evaluation framework (TweetEval) consisting of seven heterogeneous Twitter-specific classification tasks. We also provide a strong set of baselines as starting point, and compare different language modeling pre-training strategies. Our initial experiments show the effectiveness of starting off with existing pre-trained generic language models, and continue training them on Twitter corpora.

Jose Camacho-Collados, Yerai Doval, Eugenio Martínez-Cámara et al.

Cross-lingual embeddings represent the meaning of words from different languages in the same vector space. Recent work has shown that it is possible to construct such representations by aligning independently learned monolingual embedding spaces, and that accurate alignments can be obtained even without external bilingual data. In this paper we explore a research direction that has been surprisingly neglected in the literature: leveraging noisy user-generated text to learn cross-lingual embeddings particularly tailored towards social media applications. While the noisiness and informal nature of the social media genre poses additional challenges to cross-lingual embedding methods, we find that it also provides key opportunities due to the abundance of code-switching and the existence of a shared vocabulary of emoji and named entities. Our contribution consists of a very simple post-processing step that exploits these phenomena to significantly improve the performance of state-of-the-art alignment methods.

Francesco Barbieri, Luis Marujo, Pradeep Karuturi et al.

The frequent use of Emojis on social media platforms has created a new form of multimodal social interaction. Developing methods for the study and representation of emoji semantics helps to improve future multimodal communication systems. In this paper, we explore the usage and semantics of emojis over time. We compare emoji embeddings trained on a corpus of different seasons and show that some emojis are used differently depending on the time of the year. Moreover, we propose a method to take into account the time information for emoji prediction systems, outperforming state-of-the-art systems. We show that, using the time information, the accuracy of some emojis can be significantly improved.

Francesco Barbieri, Miguel Ballesteros, Francesco Ronzano et al.

Emojis are small images that are commonly included in social media text messages. The combination of visual and textual content in the same message builds up a modern way of communication, that automatic systems are not used to deal with. In this paper we extend recent advances in emoji prediction by putting forward a multimodal approach that is able to predict emojis in Instagram posts. Instagram posts are composed of pictures together with texts which sometimes include emojis. We show that these emojis can be predicted by using the text, but also using the picture. Our main finding is that incorporating the two synergistic modalities, in a combined model, improves accuracy in an emoji prediction task. This result demonstrates that these two modalities (text and images) encode different information on the use of emojis and therefore can complement each other.

Sergio Oramas, Oriol Nieto, Francesco Barbieri et al.

Music genres allow to categorize musical items that share common characteristics. Although these categories are not mutually exclusive, most related research is traditionally focused on classifying tracks into a single class. Furthermore, these categories (e.g., Pop, Rock) tend to be too broad for certain applications. In this work we aim to expand this task by categorizing musical items into multiple and fine-grained labels, using three different data modalities: audio, text, and images. To this end we present MuMu, a new dataset of more than 31k albums classified into 250 genre classes. For every album we have collected the cover image, text reviews, and audio tracks. Additionally, we propose an approach for multi-label genre classification based on the combination of feature embeddings learned with state-of-the-art deep learning methodologies. Experiments show major differences between modalities, which not only introduce new baselines for multi-label genre classification, but also suggest that combining them yields improved results.

Francesco Barbieri, Miguel Ballesteros, Horacio Saggion

Emojis are ideograms which are naturally combined with plain text to visually complement or condense the meaning of a message. Despite being widely used in social media, their underlying semantics have received little attention from a Natural Language Processing standpoint. In this paper, we investigate the relation between words and emojis, studying the novel task of predicting which emojis are evoked by text-based tweet messages. We train several models based on Long Short-Term Memory networks (LSTMs) in this task. Our experimental results show that our neural model outperforms two baselines as well as humans solving the same task, suggesting that computational models are able to better capture the underlying semantics of emojis.