David Semedo

Diogo Tavares, Pedro Azevedo, David Semedo et al.

Dialogue systems need to deal with the unpredictability of user intents to track dialogue state and the heterogeneity of slots to understand user preferences. In this paper we investigate the hypothesis that solving these challenges as one unified model will allow the transfer of parameter support data across the different tasks. The proposed principled model is based on a Transformer encoder, trained on multiple tasks, and leveraged by a rich input that conditions the model on the target inferences. Conditioning the Transformer encoder on multiple target inferences over the same corpus, i.e., intent and multiple slot types, allows learning richer language interactions than a single-task model would be able to. In fact, experimental results demonstrate that conditioning the model on an increasing number of dialogue inference tasks leads to improved results: on the MultiWOZ dataset, the joint intent and slot detection can be improved by 3.2\% by conditioning on intent, 10.8\% by conditioning on slot and 14.4\% by conditioning on both intent and slots. Moreover, on real conversations with Farfetch costumers, the proposed conditioned BERT can achieve high joint-goal and intent detection performance throughout a dialogue.

Afonso Simplício, Gonçalo Vinagre, Miguel Moura Ramos et al.

Despite rapid progress in open large language models (LLMs), European Portuguese (pt-PT) remains underrepresented in both training data and native evaluation, with machine-translated benchmarks likely missing the variant's linguistic and cultural nuances. We introduce AMALIA, a fully open LLM that prioritizes pt-PT by using more high-quality pt-PT data during both the mid- and post-training stages. To evaluate pt-PT more faithfully, we release a suite of pt-PT benchmarks that includes translated standard tasks and four new datasets targeting pt-PT generation, linguistic competence, and pt-PT/pt-BR bias. Experiments show that AMALIA matches strong baselines on translated benchmarks while substantially improving performance on pt-PT-specific evaluations, supporting the case for targeted training and native benchmarking for European Portuguese.

Frederico Vicente, Rafael Ferreira, David Semedo et al.

Introducing curiosities in a conversation is a way to teach something new to the person in a pleasant and enjoyable way. Enriching dialogues with contextualized curiosities can improve the users' perception of a dialog system and their overall user experience. In this paper, we introduce a set of curated curiosities, targeting dialogues in the cooking and DIY domains. In particular, we use real human-agent conversations collected in the context of the Amazon Alexa TaskBot challenge, a multimodal and multi-turn conversational setting. According to an A/B test with over 1000 conversations, curiosities not only increase user engagement, but provide an average relative rating improvement of 9.7%.

Pedro Domingos, João Pereira, Vasco Lopes et al.

Automated video surveillance with Large Vision-Language Models is limited by their inherent bias towards normality, often failing to detect crimes. While Chain-of-Thought reasoning strategies show significant potential for improving performance in language tasks, the lack of inductive anomaly biases in their reasoning further steers the models towards normal interpretations. To address this, we propose Chain-of-Anomaly-Thoughts (CoAT), a multi-agent reasoning framework that introduces inductive criminal bias in the reasoning process through a final, anomaly-focused classification layer. Our method significantly improves Anomaly Detection, boosting F1-score by 11.8 p.p. on challenging low-resolution footage and Anomaly Classification by 3.78 p.p. in high-resolution videos.

Rafael Ferreira, David Semedo, João Magalhães

Predicting the success of Conversational Task Assistants (CTA) can be critical to understand user behavior and act accordingly. In this paper, we propose TB-Rater, a Transformer model which combines conversational-flow features with user behavior features for predicting user ratings in a CTA scenario. In particular, we use real human-agent conversations and ratings collected in the Alexa TaskBot challenge, a novel multimodal and multi-turn conversational context. Our results show the advantages of modeling both the conversational-flow and behavioral aspects of the conversation in a single model for offline rating prediction. Additionally, an analysis of the CTA-specific behavioral features brings insights into this setting and can be used to bootstrap future systems.

Diogo Glória-Silva, David Semedo, João Magalhães

Guiding users through complex procedural plans is an inherently multimodal task in which having visually illustrated plan steps is crucial to deliver an effective plan guidance. However, existing works on plan-following language models (LMs) often are not capable of multimodal input and output. In this work, we present MM-PlanLLM, the first multimodal LLM designed to assist users in executing instructional tasks by leveraging both textual plans and visual information. Specifically, we bring cross-modality through two key tasks: Conversational Video Moment Retrieval, where the model retrieves relevant step-video segments based on user queries, and Visually-Informed Step Generation, where the model generates the next step in a plan, conditioned on an image of the user's current progress. MM-PlanLLM is trained using a novel multitask-multistage approach, designed to gradually expose the model to multimodal instructional-plans semantic layers, achieving strong performance on both multimodal and textual dialogue in a plan-grounded setting. Furthermore, we show that the model delivers cross-modal temporal and plan-structure representations aligned between textual plan steps and instructional video moments.

Rafael Ferreira, Diogo Tavares, Diogo Silva et al.

In this report, we describe the vision, challenges, and scientific contributions of the Task Wizard team, TWIZ, in the Alexa Prize TaskBot Challenge 2022. Our vision, is to build TWIZ bot as an helpful, multimodal, knowledgeable, and engaging assistant that can guide users towards the successful completion of complex manual tasks. To achieve this, we focus our efforts on three main research questions: (1) Humanly-Shaped Conversations, by providing information in a knowledgeable way; (2) Multimodal Stimulus, making use of various modalities including voice, images, and videos; and (3) Zero-shot Conversational Flows, to improve the robustness of the interaction to unseen scenarios. TWIZ is an assistant capable of supporting a wide range of tasks, with several innovative features such as creative cooking, video navigation through voice, and the robust TWIZ-LLM, a Large Language Model trained for dialoguing about complex manual tasks. Given ratings and feedback provided by users, we observed that TWIZ bot is an effective and robust system, capable of guiding users through tasks while providing several multimodal stimuli.

Inês Vieira, Inês Calvo, Iago Paulo et al.

As Large Language Models (LLMs) expand across multilingual domains, evaluating their performance in under-represented languages becomes increasingly important. European Portuguese (pt-PT) is particularly affected, as existing training data and benchmarks are mainly in Brazilian Portuguese (pt-BR). To address this, we introduce ALBA, a linguistically grounded benchmark designed from the ground up to assess LLM proficiency in linguistic-related tasks in pt-PT across eight linguistic dimensions, including Language Variety, Culture-bound Semantics, Discourse Analysis, Word Plays, Syntax, Morphology, Lexicology, and Phonetics and Phonology. ALBA is manually constructed by language experts and paired with an LLM-as-a-judge framework for scalable evaluation of pt-PT generated language. Experiments on a diverse set of models reveal performance variability across linguistic dimensions, highlighting the need for comprehensive, variety-sensitive benchmarks that support further development of tools in pt-PT.

Rafael Ferreira, David Semedo, João Magalhães

Following complex instructions in conversational assistants can be quite daunting due to the shorter attention and memory spans when compared to reading the same instructions. Hence, when conversational assistants walk users through the steps of complex tasks, there is a need to structure the task into manageable pieces of information of the right length and complexity. In this paper, we tackle the recipes domain and convert reading structured instructions into conversational structured ones. We annotated the structure of instructions according to a conversational scenario, which provided insights into what is expected in this setting. To computationally model the conversational step's characteristics, we tested various Transformer-based architectures, showing that a token-based approach delivers the best results. A further user study showed that users tend to favor steps of manageable complexity and length, and that the proposed methodology can improve the original web-based instructional text. Specifically, 86% of the evaluated tasks were improved from a conversational suitability point of view.

Ricardo Lopes, João Magalhães, David Semedo

Significant strides have been made in natural language tasks, largely attributed to the emergence of powerful large language models (LLMs). These models, pre-trained on extensive and diverse corpora, have become increasingly capable of comprehending the intricacies of language. Despite the abundance of LLMs for many high-resource languages, the availability of such models remains limited for European Portuguese. We introduce GlórIA, a robust European Portuguese decoder LLM. To pre-train GlórIA, we assembled a comprehensive PT-PT text corpus comprising 35 billion tokens from various sources. We present our pre-training methodology, followed by an assessment of the model's effectiveness on multiple downstream tasks. Additionally, to evaluate our models' language modeling capabilities, we introduce CALAME-PT (Context-Aware LAnguage Modeling Evaluation for Portuguese), the first Portuguese zero-shot language-modeling benchmark. Evaluation shows that GlórIA significantly outperforms existing open PT decoder models in language modeling and that it can generate sound, knowledge-rich, and coherent PT-PT text. The model also exhibits strong potential for various downstream tasks.

Diogo Glória-Silva, Rafael Ferreira, Diogo Tavares et al.

Training Large Language Models (LLMs) to follow user instructions has been shown to supply the LLM with ample capacity to converse fluently while being aligned with humans. Yet, it is not completely clear how an LLM can lead a plan-grounded conversation in mixed-initiative settings where instructions flow in both directions of the conversation, i.e. both the LLM and the user provide instructions to one another. In this paper, we tackle a dual goal mixed-initiative conversational setting where the LLM not only grounds the conversation on an arbitrary plan but also seeks to satisfy both a procedural plan and user instructions. The LLM is then responsible for guiding the user through the plan and, at the same time, adapting to new circumstances, answering questions, and activating safety guardrails when needed. We propose a novel LLM that grounds the dialogue on a procedural plan, can take the dialogue initiative, and enforces guardrails on the system's behavior, while also improving the LLM's responses to unexpected user behavior. Experiments in controlled settings and with real users show that the best-performing model, which we call PlanLLM, achieves a 2.1x improvement over a strong baseline. Moreover, experiments also show good generalization to unseen domains.

Diogo Glória-Silva, David Semedo, João Maglhães

We introduce VIGiA, a novel multimodal dialogue model designed to understand and reason over complex, multi-step instructional video action plans. Unlike prior work which focuses mainly on text-only guidance, or treats vision and language in isolation, VIGiA supports grounded, plan-aware dialogue that requires reasoning over visual inputs, instructional plans, and interleaved user interactions. To this end, VIGiA incorporates two key capabilities: (1) multimodal plan reasoning, enabling the model to align uni- and multimodal queries with the current task plan and respond accurately; and (2) plan-based retrieval, allowing it to retrieve relevant plan steps in either textual or visual representations. Experiments were done on a novel dataset with rich Instructional Video Dialogues aligned with Cooking and DIY plans. Our evaluation shows that VIGiA outperforms existing state-of-the-art models on all tasks in a conversational plan guidance setting, reaching over 90\% accuracy on plan-aware VQA.

Rafael Ferreira, David Semedo, João Magalhães

Conversational systems must be robust to user interactions that naturally exhibit diverse conversational traits. Capturing and simulating these diverse traits coherently and efficiently presents a complex challenge. This paper introduces Multi-Trait Adaptive Decoding (mTAD), a method that generates diverse user profiles at decoding-time by sampling from various trait-specific Language Models (LMs). mTAD provides an adaptive and scalable approach to user simulation, enabling the creation of multiple user profiles without the need for additional fine-tuning. By analyzing real-world dialogues from the Conversational Task Assistant (CTA) domain, we identify key conversational traits and developed a framework to generate profile-aware dialogues that enhance conversational diversity. Experimental results validate the effectiveness of our approach in modeling single-traits using specialized LMs, which can capture less common patterns, even in out-of-domain tasks. Furthermore, the results demonstrate that mTAD is a robust and flexible framework for combining diverse user simulators.

Artur Horal, Daniel Pina, Henrique Paz et al.

This paper presents the vision, scientific contributions, and technical details of RedTWIZ: an adaptive and diverse multi-turn red teaming framework, to audit the robustness of Large Language Models (LLMs) in AI-assisted software development. Our work is driven by three major research streams: (1) robust and systematic assessment of LLM conversational jailbreaks; (2) a diverse generative multi-turn attack suite, supporting compositional, realistic and goal-oriented jailbreak conversational strategies; and (3) a hierarchical attack planner, which adaptively plans, serializes, and triggers attacks tailored to specific LLM's vulnerabilities. Together, these contributions form a unified framework -- combining assessment, attack generation, and strategic planning -- to comprehensively evaluate and expose weaknesses in LLMs' robustness. Extensive evaluation is conducted to systematically assess and analyze the performance of the overall system and each component. Experimental results demonstrate that our multi-turn adversarial attack strategies can successfully lead state-of-the-art LLMs to produce unsafe generations, highlighting the pressing need for more research into enhancing LLM's robustness.

Joao Pereira, Vasco Lopes, David Semedo et al.

Large Vision-Language Models (LVLMs) demonstrate remarkable performance in short-video tasks such as video question answering, but struggle in long-video understanding. The linear frame sampling strategy, conventionally used by LVLMs, fails to account for the non-linear distribution of key events in video data, often introducing redundant or irrelevant information in longer contexts while risking the omission of critical events in shorter ones. To address this, we propose SelfReS, a non-linear spatiotemporal self-reflective sampling method that dynamically selects key video fragments based on user prompts. Unlike prior approaches, SelfReS leverages the inherently sparse attention maps of LVLMs to define reflection tokens, enabling relevance-aware token selection without requiring additional training or external modules. Experiments demonstrate that SelfReS can be seamlessly integrated into strong base LVLMs, improving long-video task accuracy and achieving up to 46% faster inference speed within the same GPU memory budget.

Tyler A. Chang, Catherine Arnett, Abdelrahman Eldesokey et al. · uw

To date, there exist almost no culturally-specific evaluation benchmarks for large language models (LLMs) that cover a large number of languages and cultures. In this paper, we present Global PIQA, a participatory commonsense reasoning benchmark for over 100 languages, constructed by hand by 335 researchers from 65 countries around the world. The 116 language varieties in Global PIQA cover five continents, 14 language families, and 23 writing systems. In the non-parallel split of Global PIQA, over 50% of examples reference local foods, customs, traditions, or other culturally-specific elements. We find that state-of-the-art LLMs perform well on Global PIQA in aggregate, but they exhibit weaker performance in lower-resource languages (up to a 37% accuracy gap, despite random chance at 50%). Open models generally perform worse than proprietary models. Global PIQA highlights that in many languages and cultures, everyday knowledge remains an area for improvement, alongside more widely-discussed capabilities such as complex reasoning and expert knowledge. Beyond its uses for LLM evaluation, we hope that Global PIQA provides a glimpse into the wide diversity of cultures in which human language is embedded.

Joao Pereira, Vasco Lopes, David Semedo et al.

The growing demand for surveillance in public spaces presents significant challenges due to the shortage of human resources. Current AI-based video surveillance systems heavily rely on core computer vision models that require extensive finetuning, which is particularly difficult in surveillance settings due to limited datasets and difficult setting (viewpoint, low quality, etc.). In this work, we propose leveraging Large Vision-Language Models (LVLMs), known for their strong zero and few-shot generalization, to tackle video understanding tasks in surveillance. Specifically, we explore VideoLLaMA2, a state-of-the-art LVLM, and an improved token-level sampling method, Self-Reflective Sampling (Self-ReS). Our experiments on the UCF-Crime dataset show that VideoLLaMA2 represents a significant leap in zero-shot performance, with 20% boost over the baseline. Self-ReS additionally increases zero-shot action recognition performance to 44.6%. These results highlight the potential of LVLMs, paired with improved sampling techniques, for advancing surveillance video analysis in diverse scenarios.

Gonçalo Marcelino, David Semedo, André Mourão et al.

Creating a cohesive, high-quality, relevant, media story is a challenge that news media editors face on a daily basis. This challenge is aggravated by the flood of highly relevant information that is constantly pouring onto the newsroom. To assist news media editors in this daunting task, this paper proposes a framework to organize news content into cohesive, high-quality, relevant visual storylines. First, we formalize, in a nonsubjective manner, the concept of visual story transition. Leveraging it, we propose four graph-based methods of storyline creation, aiming for global story cohesiveness. These were created and implemented to take full advantage of existing graph algorithms, ensuring their correctness and good computational performance. They leverage a strong ensemble-based estimator which was trained to predict story transition quality based on both the semantic and visual features present in the pair of images under scrutiny. A user study covered a total of 28 curated stories about sports and cultural events. Experiments showed that (i) visual transitions in storylines can be learned with a quality above 90%, and (ii) the proposed graph methods can produce cohesive storylines with quality in the range of 88% to 96%.

Mariana Leite, Rafael Ferreira, David Semedo et al.

The conversational search paradigm introduces a step change over the traditional search paradigm by allowing users to interact with search agents in a multi-turn and natural fashion. The conversation flows naturally and is usually centered around a target field of knowledge. In this work, we propose a knowledge-driven answer generation approach for open-domain conversational search, where a conversation-wide entities' knowledge graph is used to bias search-answer generation. First, a conversation-specific knowledge graph is extracted from the top passages retrieved with a Transformer-based re-ranker. The entities knowledge-graph is then used to bias a search-answer generator Transformer towards information rich and concise answers. This conversation specific bias is computed by identifying the most relevant passages according to the most salient entities of that particular conversation. Experiments show that the proposed approach successfully exploits entities knowledge along the conversation, and outperforms a set of baselines on the search-answer generation task.

Rafael Ferreira, David Semedo, Joao Magalhaes

The use of conversational assistants to search for information is becoming increasingly more popular among the general public, pushing the research towards more advanced and sophisticated techniques. In the last few years, in particular, the interest in conversational search is increasing, not only because of the generalization of conversational assistants but also because conversational search is a step forward in allowing a more natural interaction with the system. In this work, the focus is on exploring the context present of the conversation via the historical utterances and respective embeddings with the aim of developing a conversational search system that helps people search for information in a natural way. In particular, this system must be able to understand the context where the question is posed, tracking the current state of the conversation and detecting mentions to previous questions and answers. We achieve this by using a context-tracking component based on neural query-rewriting models. Another crucial aspect of the system is to provide the most relevant answers given the question and the conversational history. To achieve this objective, we used a Transformer-based re-ranking method and expanded this architecture to use the conversational context. The results obtained with the system developed showed the advantages of using the context present in the natural language utterances and in the neural embeddings generated throughout the conversation.

Rafael Ferreira, Mariana Leite, David Semedo et al.

Open-domain conversational search assistants aim at answering user questions about open topics in a conversational manner. In this paper we show how the Transformer architecture achieves state-of-the-art results in key IR tasks, leveraging the creation of conversational assistants that engage in open-domain conversational search with single, yet informative, answers. In particular, we propose an open-domain abstractive conversational search agent pipeline to address two major challenges: first, conversation context-aware search and second, abstractive search-answers generation. To address the first challenge, the conversation context is modeled with a query rewriting method that unfolds the context of the conversation up to a specific moment to search for the correct answers. These answers are then passed to a Transformer-based re-ranker to further improve retrieval performance. The second challenge, is tackled with recent Abstractive Transformer architectures to generate a digest of the top most relevant passages. Experiments show that Transformers deliver a solid performance across all tasks in conversational search, outperforming the best TREC CAsT 2019 baseline.

David Semedo, João Magalhães

Cross-modal embeddings, between textual and visual modalities, aim to organise multimodal instances by their semantic correlations. State-of-the-art approaches use maximum-margin methods, based on the hinge-loss, to enforce a constant margin m, to separate projections of multimodal instances from different categories. In this paper, we propose a novel scheduled adaptive maximum-margin (SAM) formulation that infers triplet-specific constraints during training, therefore organising instances by adaptively enforcing inter-category and inter-modality correlations. This is supported by a scheduled adaptive margin function, that is smoothly activated, replacing a static margin by an adaptively inferred one reflecting triplet-specific semantic correlations while accounting for the incremental learning behaviour of neural networks to enforce category cluster formation and enforcement. Experiments on widely used datasets show that our model improved upon state-of-the-art approaches, by achieving a relative improvement of up to ~12.5% over the second best method, thus confirming the effectiveness of our scheduled adaptive margin formulation.

David Semedo, João Magalhães

Understanding the semantic shifts of multimodal information is only possible with models that capture cross-modal interactions over time. Under this paradigm, a new embedding is needed that structures visual-textual interactions according to the temporal dimension, thus, preserving data's original temporal organisation. This paper introduces a novel diachronic cross-modal embedding (DCM), where cross-modal correlations are represented in embedding space, throughout the temporal dimension, preserving semantic similarity at each instant t. To achieve this, we trained a neural cross-modal architecture, under a novel ranking loss strategy, that for each multimodal instance, enforces neighbour instances' temporal alignment, through subspace structuring constraints based on a temporal alignment window. Experimental results show that our DCM embedding successfully organises instances over time. Quantitative experiments, confirm that DCM is able to preserve semantic cross-modal correlations at each instant t while also providing better alignment capabilities. Qualitative experiments unveil new ways to browse multimodal content and hint that multimodal understanding tasks can benefit from this new embedding.

Gonçalo Marcelino, David Semedo, André Mourão et al.

Media editors in the newsroom are constantly pressed to provide a "like-being there" coverage of live events. Social media provides a disorganised collection of images and videos that media professionals need to grasp before publishing their latest news updated. Automated news visual storyline editing with social media content can be very challenging, as it not only entails the task of finding the right content but also making sure that news content evolves coherently over time. To tackle these issues, this paper proposes a benchmark for assessing social media visual storylines. The SocialStories benchmark, comprised by total of 40 curated stories covering sports and cultural events, provides the experimental setup and introduces novel quantitative metrics to perform a rigorous evaluation of visual storytelling with social media data.

David Semedo, João Magalhães

Multimedia information have strong temporal correlations that shape the way modalities co-occur over time. In this paper we study the dynamic nature of multimedia and social-media information, where the temporal dimension emerges as a strong source of evidence for learning the temporal correlations across visual and textual modalities. So far, cross-media retrieval models, explored the correlations between different modalities (e.g. text and image) to learn a common subspace, in which semantically similar instances lie in the same neighbourhood. Building on such knowledge, we propose a novel temporal cross-media neural architecture, that departs from standard cross-media methods, by explicitly accounting for the temporal dimension through temporal subspace learning. The model is softly-constrained with temporal and inter-modality constraints that guide the new subspace learning task by favouring temporal correlations between semantically similar and temporally close instances. Experiments on three distinct datasets show that accounting for time turns out to be important for cross-media retrieval. Namely, the proposed method outperforms a set of baselines on the task of temporal cross-media retrieval, demonstrating its effectiveness for performing temporal subspace learning.

David Semedo, João Magalhães, Flávio Martins

In this paper we address the task of gender classification on picture sharing social media networks such as Instagram and Flickr. We aim to infer the gender of an user given only a small set of the images shared in its profile. We make the assumption that user's images contain a collection of visual elements that implicitly encode discriminative patterns that allow inferring its gender, in a language independent way. This information can then be used in personalisation and recommendation. Our main hypothesis is that semantic visual features are more adequate for discriminating high-level classes. The gender detection task is formalised as: given an user's profile, represented as a bag of images, we want to infer the gender of the user. Social media profiles can be noisy and contain confounding factors, therefore we classify bags of user-profile's images to provide a more robust prediction. Experiments using a dataset from the picture sharing social network Instagram show that the use of multiple images is key to improve detection performance. Moreover, we verify that deep semantic features are more suited for gender detection than low-level image representations. The methods proposed can infer the gender with precision scores higher than 0.825, and the best performing method achieving 0.911 precision.