Ridho Reinanda

Vivek Gupta, Pranshu Kandoi, Mahek Bhavesh Vora et al.

Semi-structured data, such as Infobox tables, often include temporal information about entities, either implicitly or explicitly. Can current NLP systems reason about such information in semi-structured tables? To tackle this question, we introduce the task of temporal question answering on semi-structured tables. We present a dataset, TempTabQA, which comprises 11,454 question-answer pairs extracted from 1,208 Wikipedia Infobox tables spanning more than 90 distinct domains. Using this dataset, we evaluate several state-of-the-art models for temporal reasoning. We observe that even the top-performing LLMs lag behind human performance by more than 13.5 F1 points. Given these results, our dataset has the potential to serve as a challenging benchmark to improve the temporal reasoning capabilities of NLP models.

Mariya Hendriksen, Shuo Zhang, Ridho Reinanda et al.

Image-Text Retrieval (ITR) systems are central to multimodal information access, with Vision-Language Models (VLMs) showing strong performance on standard benchmarks. However, these benchmarks predominantly rely on coarse-grained annotations, limiting their ability to reveal how models perform under real-world conditions, where query granularity varies. Motivated by this gap, we examine how dataset granularity and query perturbations affect retrieval performance and robustness across four architecturally diverse VLMs (ALIGN, AltCLIP, CLIP, and GroupViT). Using both standard benchmarks (MS-COCO, Flickr30k) and their fine-grained variants, we show that richer captions consistently enhance retrieval, especially in text-to-image tasks, where we observe an average improvement of 16.23%, compared to 6.44% in image-to-text. To assess robustness, we introduce a taxonomy of perturbations and conduct extensive experiments, revealing that while perturbations typically degrade performance, they can also unexpectedly improve retrieval, exposing nuanced model behaviors. Notably, word order emerges as a critical factor -- contradicting prior assumptions of model insensitivity to it. Our results highlight variation in model robustness and a dataset-dependent relationship between caption granularity and perturbation sensitivity and emphasize the necessity of evaluating models on datasets of varying granularity.

Nikos Voskarides, Edgar Meij, Ridho Reinanda et al.

Knowledge graphs (KGs) model facts about the world, they consist of nodes (entities such as companies and people) that are connected by edges (relations such as founderOf). Facts encoded in KGs are frequently used by search applications to augment result pages. When presenting a KG fact to the user, providing other facts that are pertinent to that main fact can enrich the user experience and support exploratory information needs. KG fact contextualization is the task of augmenting a given KG fact with additional and useful KG facts. The task is challenging because of the large size of KGs, discovering other relevant facts even in a small neighborhood of the given fact results in an enormous amount of candidates. We introduce a neural fact contextualization method (NFCM) to address the KG fact contextualization task. NFCM first generates a set of candidate facts in the neighborhood of a given fact and then ranks the candidate facts using a supervised learning to rank model. The ranking model combines features that we automatically learn from data and that represent the query-candidate facts with a set of hand-crafted features we devised or adjusted for this task. In order to obtain the annotations required to train the learning to rank model at scale, we generate training data automatically using distant supervision on a large entity-tagged text corpus. We show that ranking functions learned on this data are effective at contextualizing KG facts. Evaluation using human assessors shows that it significantly outperforms several competitive baselines.

Ridho Reinanda, Edgar Meij, Maarten de Rijke

Filtering relevant documents with respect to entities is an essential task in the context of knowledge base construction and maintenance. It entails processing a time-ordered stream of documents that might be relevant to an entity in order to select only those that contain vital information. State-of-the-art approaches to document filtering for popular entities are entity-dependent: they rely on and are also trained on the specifics of differentiating features for each specific entity. Moreover, these approaches tend to use so-called extrinsic information such as Wikipedia page views and related entities which is typically only available only for popular head entities. Entity-dependent approaches based on such signals are therefore ill-suited as filtering methods for long-tail entities. In this paper we propose a document filtering method for long-tail entities that is entity-independent and thus also generalizes to unseen or rarely seen entities. It is based on intrinsic features, i.e., features that are derived from the documents in which the entities are mentioned. We propose a set of features that capture informativeness, entity-saliency, and timeliness. In particular, we introduce features based on entity aspect similarities, relation patterns, and temporal expressions and combine these with standard features for document filtering. Experiments following the TREC KBA 2014 setup on a publicly available dataset show that our model is able to improve the filtering performance for long-tail entities over several baselines. Results of applying the model to unseen entities are promising, indicating that the model is able to learn the general characteristics of a vital document. The overall performance across all entities---i.e., not just long-tail entities---improves upon the state-of-the-art without depending on any entity-specific training data.

V. A. Traag, R. Reinanda, G. van Klinken

Social networks have been of much interest in recent years. We here focus on a network structure derived from co-occurrences of people in traditional newspaper media. We find three clear deviations from what can be expected in a random graph. First, the average degree in the empirical network is much lower than expected, and the average weight of a link much higher than expected. Secondly, high degree nodes attract disproportionately much weight. Thirdly, relatively much of the weight seems to concentrate between high degree nodes. We believe this can be explained by the fact that most people tend to co-occur repeatedly with the same people. We create a model that replicates these observations qualitatively based on two self-reinforcing processes: (1) more frequently occurring persons are more likely to occur again; and (2) if two people co-occur frequently, they are more likely to co-occur again. This suggest that the media tends to focus on people that are already in the news, and that they reinforce existing co-occurrences.