Efficiently retrieving and synthesizing information from large-scale multimodal collections has become a critical challenge. However, existing video retrieval datasets suffer from scope limitations, primarily focusing on matching descriptive but vague queries with small collections of professionally edited, English-centric videos. To address this gap, we introduce MultiVENT 2.0, a large-scale, multilingual event-centric video retrieval benchmark featuring a collection of more than 218,000 news videos and over 3,900 queries targeting specific world events. These queries specifically target information found in the visual content, audio, embedded text, and text metadata of the videos, requiring systems leverage all these sources to succeed at the task. Preliminary results show that state-of-the-art vision-language models struggle significantly with this task, and while alternative approaches show promise, they are still insufficient to adequately address this problem. These findings underscore the need for more robust multimodal retrieval systems, as effective video retrieval is a crucial step towards multimodal content understanding and generation.

Videos inherently contain multiple modalities, including visual events, text overlays, sounds, and speech, all of which are important for retrieval. However, state-of-the-art multimodal language models like VAST and LanguageBind are built on vision-language models (VLMs), and thus overly prioritize visual signals. Retrieval benchmarks further reinforce this bias by focusing on visual queries and neglecting other modalities. We create a search system MMMORRF that extracts text and features from both visual and audio modalities and integrates them with a novel modality-aware weighted reciprocal rank fusion. MMMORRF is both effective and efficient, demonstrating practicality in searching videos based on users' information needs instead of visual descriptive queries. We evaluate MMMORRF on MultiVENT 2.0 and TVR, two multimodal benchmarks designed for more targeted information needs, and find that it improves nDCG@20 by 81% over leading multimodal encoders and 37% over single-modality retrieval.

Everyday news coverage has shifted from traditional broadcasts towards a wide range of presentation formats such as first-hand, unedited video footage. Datasets that reflect the diverse array of multimodal, multilingual news sources available online could be used to teach models to benefit from this shift, but existing news video datasets focus on traditional news broadcasts produced for English-speaking audiences. We address this limitation by constructing MultiVENT, a dataset of multilingual, event-centric videos grounded in text documents across five target languages. MultiVENT includes both news broadcast videos and non-professional event footage, which we use to analyze the state of online news videos and how they can be leveraged to build robust, factually accurate models. Finally, we provide a model for complex, multilingual video retrieval to serve as a baseline for information retrieval using MultiVENT.

Machine translation models have discrete vo cabularies and commonly use subword seg mentation techniques to achieve an 'open vo cabulary.'This approach relies on consis tent and correct underlying unicode sequences, and makes models susceptible to degrada tion from common types of noise and vari ation.Motivated by the robustness of hu man language processing, we propose the use of visual text representations, which dispense with a finite set of text embeddings in favor of continuous vocabularies created by process ing visually rendered text with sliding win dows.We show that models using visual text representations approach or match per formance of traditional text models on small and larger datasets.More importantly, mod els with visual embeddings demonstrate sig nificant robustness to varied types of noise, achieving e.g., 25.9 BLEU on a character per muted German-English task where subword models degrade to 1.9.

Named entity recognition (NER) identifies spans of text that contain names. Many researchers have reported the results of NER on text created through optical character recognition (OCR) over the past two decades. Unfortunately, the test collections that support this research are annotated with named entities after optical character recognition (OCR) has been run. This means that the collection must be re-annotated if the OCR output changes. Instead by tying annotations to character locations on the page, a collection can be built that supports OCR and NER research without requiring re-annotation when either improves. This means that named entities are annotated on the transcribed text. The transcribed text is all that is needed to evaluate the performance of OCR. For NER evaluation, the tagged OCR output is aligned to the transcriptions the aligned files, creating modified files of each, which are scored. This paper presents a methodology for building such a test collection and releases a collection of Chinese OCR-NER data constructed using the methodology. The paper provides performance baselines for current OCR and NER systems applied to this new collection.

Hybrid deep neural network hidden Markov models (DNN-HMM) have achieved impressive results on large vocabulary continuous speech recognition (LVCSR) tasks. However, the recent approaches using DNN-HMM models are not explored much for text recognition. Inspired by the current work in automatic speech recognition (ASR) and machine translation, we present an open vocabulary sub-word text recognition system. The sub-word lexicon and sub-word language model (LM) helps in overcoming the challenge of recognizing out of vocabulary (OOV) words, and a time delay neural network (TDNN) and convolution neural network (CNN) based DNN-HMM optical model (OM) efficiently models the sequence dependency in the line image. We present results on 12 datasets with training data varying from 6k lines to 600k lines. The system is built for 8 languages, i.e., English, French, Arabic, Chinese, Farsi, Tamil, Russian, and Korean. We report competitive results on several commonly used handwritten and printed text datasets.

Synthetic data generation for optical character recognition (OCR) promises unlimited training data at zero annotation cost. With enough fonts and seed text, we should be able to generate data to train a model that approaches or exceeds the performance with real annotated data. Unfortunately, this is not always the reality. Unconstrained image settings, such as internet memes, scanned web pages, or newspapers, present diverse scripts, fonts, layouts, and complex backgrounds, which cause models trained with synthetic data to break down. In this work, we investigate the synthetic image generation problem on a large multilingual set of unconstrained document images. Our work presents a comprehensive evaluation of the impact of synthetic data attributes on model performance. The results provide a recipe for synthetic data generation that will help guide future research.

In this paper, we apply several modifications to script identification, several of which inspired by techniques from the similar audio task of spoken language recognition. Specifically, we alter the architecture of a convolutional network with global average pooling to include variance pooling as well, we utilize score calibration of the output scores of the network, and we utilize prior distribution estimation to condition the calibrated scores. We show that these methods are effective in script identification, with the use of priors showing especially promising improvements. Furthermore, in the domain of script identification, several additional extensions of distribution estimation are available which consider the distribution within each image, and we demonstrate much larger improvements when employing these extensions. Finally, we also show that an embedding-plus-classifier approach performs similarly to the full network, and so its potential for increased flexibility may be beneficial for future consideration. With all modifications, overall accuracy on the ICDAR 2017 validation dataset increases from 89.7% to 93.6%.

Known Item Search (KIS) is a specialized task of the general multimedia search problem. It describes the scenario where a user has previously seen a video and wants to find it again in a large collection using a text description. While there exists only one correct answer to a query (or topic), the goal is to return a ranked list of videos most likely to satisfy the request. This search problem includes content from speech, visual, and meta-data, and it is not clear how the individual modalities should be combined in the final result. Reranking models have been shown to be effective in problems such as image search, but the single ground truth video for a topic presents a challenge for building a model. In this paper, we propose a semi-supervised rank learning approach to the multimedia problem. We use a large training set of topics and ground truth videos to learn a pairwise ranking model based on gradient boosted regression trees. We define a learning feature space that consists of features derived from topics, videos, and topic-video dependent results. To overcome the KIS class imbalance problem, a set of pseudo positive training examples are identified from each of the multimedia modalities. This semi-supervised approach uses a ground truth video to select similar videos in each of the individual modalities. We then model the similarities as a graph and use a K-Step Markov approach to estimate the importance of nodes in the graph relative to the truth root node.

Reverse nearest neighbor queries are useful in identifying objects that are of significant influence or importance. Existing methods either rely on pre-computation of nearest neighbor distances, do not scale well with high dimensionality, or do not produce exact solutions. In this work we motivate and investigate the problem of reverse nearest neighbor search on high dimensional, multimedia data. We propose exact and approximate algorithms that do not require pre-computation of nearest neighbor distances, and can potentially prune off most of the search space. We demonstrate the utility of reverse nearest neighbor search by showing how it can help improve the classification accuracy.

We introduce regression databases (REDB) to formalize and automate probabilistic querying using sparse learning sets. The REDB data model involves observation data, learning set data, views definitions, and a regression model instance. The observation data is a collection of relational tuples over a set of attributes; the learning data set involves a subset of observation tuples, augmented with learned attributes, which are modeled as random variables; the views are expressed as linear combinations of observation and learned attributes; and the regression model involves functions that map observation tuples to probability distributions of the random variables, which are learned dynamically from the learning data set. The REDB query language extends relational algebra project-select queries with conditions on probabilities of first-order logical expressions, which in turn involve linear combinations of learned attributes and views, and arithmetic comparison operators. Such capability relies on the underlying regression model for the learned attributes. We show that REDB queries are computable by developing conceptual evaluation algorithms and by proving their correctness and termination