We present TerraMind, the first any-to-any generative, multimodal foundation model for Earth observation (EO). Unlike other multimodal models, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind leverages fine-grained representations to capture critical spatial nuances. We pretrained TerraMind on nine geospatial modalities of a global, large-scale dataset. In this paper, we demonstrate that (i) TerraMind's dual-scale early fusion approach unlocks a range of zero-shot and few-shot applications for Earth observation, (ii) TerraMind introduces "Thinking-in-Modalities" (TiM) -- the capability of generating additional artificial data during finetuning and inference to improve the model output -- and (iii) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks for EO like PANGAEA. The pretraining dataset, the model weights, and our code are open-sourced under a permissive license.

SDForger is a flexible and efficient framework for generating high-quality multivariate time series using LLMs. Leveraging a compact data representation, SDForger provides synthetic time series generation from a few samples and low-computation fine-tuning of any autoregressive LLM. Specifically, the framework transforms univariate and multivariate signals into tabular embeddings, which are then encoded into text and used to fine-tune the LLM. At inference, new textual embeddings are sampled and decoded into synthetic time series that retain the original data's statistical properties and temporal dynamics. Across a diverse range of datasets, SDForger outperforms existing generative models in many scenarios, both in similarity-based evaluations and downstream forecasting tasks. By enabling textual conditioning in the generation process, SDForger paves the way for multimodal modeling and the streamlined integration of time series with textual information. The model is open-sourced at https://github.com/IBM/fms-dgt/tree/main/fms_dgt/public/databuilders/time_series.

Companies invest a substantial amount of time and resources in ensuring the compliance to the existing regulations or in the form of fines when compliance cannot be proven in auditing procedures. The topic is not only relevant, but also highly complex, given the frequency of changes and amendments, the complexity of the cases and the difficulty of the juristic language. This paper aims at applying advanced extractive summarization to democratize the understanding of regulations, so that non-jurists can decide which regulations deserve further follow-up. To achieve that, we first create a corpus named EUR-LexSum EUR-LexSum containing 4595 curated European regulatory documents and their corresponding summaries. We then fine-tune transformer-based models which, applied to this corpus, yield a superior performance (in terms of ROUGE metrics) compared to a traditional extractive summarization baseline. Our experiments reveal that even with limited amounts of data such transformer-based models are effective in the field of legal document summarization.

It has never been easier to get access to content. Rather, we are facing an ever increasing overload which undermines the ability to identify high-quality content relevant to the user. Automatic summarization techniques have been developed to distil down the content to the key points and shorten therewith the time required to grasp the essence and judge about the relevance of the document. Summarization is not a deterministic task and depends very much on the writing style of the person creating the summary. In this work we present a method to, given a set of human-created summaries for a corpus, establishes which automatic extractive summarization technique preserves best the style of the human summary writer. To prove our approach, we use a corpus of 1000 articles by Science Daily with the corresponding human-written summaries and benchmark 3 extractive summarization techniques (BERT-based, keyword-scoring-based and a Luhn summarizer), indicating the best style-preserving method and discussing the results.

Polarity detection plays a pivotal role in the modern cognitive research field. Common approaches to compute the polarity of a given word rely on experimental dictionaries providing always the same value, no matter where the word is used and lacking therefore adaptivity to particular contexts. In a previous article, we proposed a method supported by fuzzy linguistic modelling to quantify this contextual bias and to enable the bias-aware sentiment analysis. In this work, we implement the bias contextualization based on a word embeddings technique to capture a larger portion of the contextual bias. To show how our approach works, we measure the bias of common concepts in two different domains and discuss the results compared with our previous attempt based on document contextualization.

Recommender systems predict and recommend items that could be from the user's interest. One of the main problems from recommender systems is the cold start problem, which happens when new users or items are registered into the system. In order to deal with this problem in the literature we can find many proposal. In many cases the solutions require users to provide explicit information which requires effort on their part. For that reason and given the great boom of social networks, we focus on systems that make use of external data sources. Then, in this paper will be presented an approach in which using social media data will classify the users and create some predictions by using classification trees. Therefore the users will not need to explicitly provide any data other than the source of their social media, helping in this way to alleviate the cold start problem, because the system always has information available to establish user profiles.

Recommender systems are key enablers to provide personalization and to make systems be adapted to users' needs. Both, users and content or commercial providers benefit from these techniques. While user profiling is required during the recommendation process, it can also introduce additional threats to the user's privacy. New regulations come into place to palliate the misuse of personal information from companies and public institutions. However, there are no clear rules defined for recommender systems. We find in the literature different proposals to privacy-preserving recommender system, but none of them tackle the compliance with the General Data Protection Regulation (GDPR). In this work we suggest a set of guidelines to assess and implement GDPR compliant recommender systems. Recommender providers shall follow our guidelines to make sure that their systems are not only privacy-preserving, but also GDPR compliant.

Wikipedia is the best-developed attempt so far to gather all human knowledge in one place yet there's no approach to exhaustively address all different types of information requirements. In this work we propose an information discovery, retrieval and recommendation system supported by an ensemble of techniques to cover all informational needs in a holistic and dynamic way. We first introduce the typology of informational needs and then present the system building blocks, explaining the ensemble of supporting techniques and which particular information requirement type benefits from each one. To prove how our system works, we discuss the results of the Machine Learning semantic field.