Ritesh Sarkhel

Ritesh Sarkhel, Arnab Nandi

Visually rich documents (e.g. leaflets, banners, magazine articles) are physical or digital documents that utilize visual cues to augment their semantics. Information contained in these documents are ad-hoc and often incomplete. Existing works that enable structured querying on these documents do not take this into account. This makes it difficult to contextualize the information retrieved from querying these documents and gather actionable insights from them. We propose Juno -- a cross-modal entity matching framework to address this limitation. It augments heterogeneous documents with supplementary information by matching a text span in the document with semantically similar tuples from an external database. Our main contribution in this is a deep neural network with attention that goes beyond traditional keyword-based matching and finds matching tuples by aligning text spans and relational tuples on a multimodal encoding space without any prior knowledge about the document type or the underlying schema. Exhaustive experiments on multiple real-world datasets show that Juno generalizes to heterogeneous documents with diverse layouts and formats. It outperforms state-of-the-art baselines by more than 6 F1 points with up to 60% less human-labeled samples. Our experiments further show that Juno is a computationally robust framework. We can train it only once, and then adapt it dynamically for multiple resource-constrained environments without sacrificing its downstream performance. This makes it suitable for on-device deployment in various edge-devices. To the best of our knowledge, ours is the first work that investigates the information incompleteness of visually rich documents and proposes a generalizable, performant and computationally robust framework to address it in an end-to-end way.

Ritesh Sarkhel, Binxuan Huang, Colin Lockard et al.

Extracting structured information from HTML documents is a long-studied problem with a broad range of applications, including knowledge base construction, faceted search, and personalized recommendation. Prior works rely on a few human-labeled web pages from each target website or thousands of human-labeled web pages from some seed websites to train a transferable extraction model that generalizes on unseen target websites. Noisy content, low site-level consistency, and lack of inter-annotator agreement make labeling web pages a time-consuming and expensive ordeal. We develop LEAST -- a Label-Efficient Self-Training method for Semi-Structured Web Documents to overcome these limitations. LEAST utilizes a few human-labeled pages to pseudo-annotate a large number of unlabeled web pages from the target vertical. It trains a transferable web-extraction model on both human-labeled and pseudo-labeled samples using self-training. To mitigate error propagation due to noisy training samples, LEAST re-weights each training sample based on its estimated label accuracy and incorporates it in training. To the best of our knowledge, this is the first work to propose end-to-end training for transferable web extraction models utilizing only a few human-labeled pages. Experiments on a large-scale public dataset show that using less than ten human-labeled pages from each seed website for training, a LEAST-trained model outperforms previous state-of-the-art by more than 26 average F1 points on unseen websites, reducing the number of human-labeled pages to achieve similar performance by more than 10x.

Yilun Jin, Zheng Li, Chenwei Zhang et al.

Online shopping is a complex multi-task, few-shot learning problem with a wide and evolving range of entities, relations, and tasks. However, existing models and benchmarks are commonly tailored to specific tasks, falling short of capturing the full complexity of online shopping. Large Language Models (LLMs), with their multi-task and few-shot learning abilities, have the potential to profoundly transform online shopping by alleviating task-specific engineering efforts and by providing users with interactive conversations. Despite the potential, LLMs face unique challenges in online shopping, such as domain-specific concepts, implicit knowledge, and heterogeneous user behaviors. Motivated by the potential and challenges, we propose Shopping MMLU, a diverse multi-task online shopping benchmark derived from real-world Amazon data. Shopping MMLU consists of 57 tasks covering 4 major shopping skills: concept understanding, knowledge reasoning, user behavior alignment, and multi-linguality, and can thus comprehensively evaluate the abilities of LLMs as general shop assistants. With Shopping MMLU, we benchmark over 20 existing LLMs and uncover valuable insights about practices and prospects of building versatile LLM-based shop assistants. Shopping MMLU can be publicly accessed at https://github.com/KL4805/ShoppingMMLU. In addition, with Shopping MMLU, we host a competition in KDD Cup 2024 with over 500 participating teams. The winning solutions and the associated workshop can be accessed at our website https://amazon-kddcup24.github.io/.

Jing Yang Lee, Hamed Bonab, Nasser Zalmout et al.

Large Language Models (LLMs) are increasingly employed in multi-turn conversational tasks, yet their pre-training data predominantly consists of continuous prose, creating a potential mismatch between required capabilities and training paradigms. We introduce a novel approach to address this discrepancy by synthesizing conversational data from existing text corpora. We present a pipeline that transforms a cluster of multiple related documents into an extended multi-turn, multi-topic information-seeking dialogue. Applying our pipeline to Wikipedia articles, we curate DocTalk, a multi-turn pre-training dialogue corpus consisting of over 730k long conversations. We hypothesize that exposure to such synthesized conversational structures during pre-training can enhance the fundamental multi-turn capabilities of LLMs, such as context memory and understanding. Empirically, we show that incorporating DocTalk during pre-training results in up to 40% gain in context memory and understanding, without compromising base performance. DocTalk is available at https://huggingface.co/datasets/AmazonScience/DocTalk.

Animesh Singh, Sandip Saha, Ritesh Sarkhel et al.

Deep neural network-based architectures give promising results in various domains including pattern recognition. Finding the optimal combination of the hyper-parameters of such a large-sized architecture is tedious and requires a large number of laboratory experiments. But, identifying the optimal combination of a hyper-parameter or appropriate kernel size for a given architecture of deep learning is always a challenging and tedious task. Here, we introduced a genetic algorithm-based technique to reduce the efforts of finding the optimal combination of a hyper-parameter (kernel size) of a convolutional neural network-based architecture. The method is evaluated on three popular datasets of different handwritten Bangla characters and digits. The implementation of the proposed methodology can be found in the following link: https://github.com/DeepQn/GA-Based-Kernel-Size.

Zhenhailong Wang, Jiateng Liu, Amin Fazel et al.

Modern conversational agents like ChatGPT and Alexa+ rely on predefined policies specifying metadata, response styles, and tool-usage rules. As these LLM-based systems expand to support diverse business and user queries, such policies, often implemented as in-context prompts, are becoming increasingly complex and lengthy, making faithful adherence difficult and imposing large fixed computational costs. With the rise of multimodal agents, policies that govern visual and multimodal behaviors are critical but remain understudied. Prior prompt-compression work mainly shortens task templates and demonstrations, while existing policy-alignment studies focus only on text-based safety rules. We introduce Multimodal Policy Internalization (MPI), a new task that internalizes reasoning-intensive multimodal policies into model parameters, enabling stronger policy-following without including the policy during inference. MPI poses unique data and algorithmic challenges. We build two datasets spanning synthetic and real-world decision-making and tool-using tasks and propose TriMPI, a three-stage training framework. TriMPI first injects policy knowledge via continual pretraining, then performs supervised finetuning, and finally applies PolicyRollout, a GRPO-style reinforcement learning extension that augments rollouts with policy-aware responses for grounded exploration. TriMPI achieves notable gains in end-to-end accuracy, generalization, and robustness to forgetting. As the first work on multimodal policy internalization, we provide datasets, training recipes, and comprehensive evaluations to foster future research. Project page: https://mikewangwzhl.github.io/TriMPI.

Weizhi Wang, Rongmei Lin, Shiyang Li et al. · amazon-science

The Multimodal Large Language Models (MLLMs) are continually pre-trained on a mixture of image-text caption data and interleaved document data, while the high-quality data filtering towards image-text interleaved document data is under-explored. We propose to train an efficient MLLM as a Unified Mulitmodal Data Quality Classifier to Filter both high-quality image-text caption and interleaved data (UniFilter). To address the challenge of collecting diverse labeled multimodal data, we introduce a semi-synthetic approach that leverages readily available raw images and generates corresponding text across four quality levels. This method enables efficient creation of sample-score pairs for both caption and interleaved document data to train UniFilter. We apply UniFilter to curate high-quality caption data from DataComp caption dataset and interleaved data from the OBELICS image-text interleaved dataset. MLLMs pre-trained on the filtered data demonstrate significantly enhanced capabilities compared to those trained on baseline-filtered data, achieving stronger zero-shot reasoning and in-context learning capabilities. After visual supervised fine-tuning, these UniFilter-induced MLLMs achieve stronger performance on various benchmarks, highlighting the downstream benefits of high-quality multimodal pre-training. We release the synthetic training data used for training UniFilter, the UniFilter model checkpoints, and the high-quality interleaved document subset OBELICS-HQ, curated by UniFilter, to the community for reproduction and further development.

Ritesh Sarkhel, Xiaoqi Ren, Lauro Beltrao Costa et al.

A visually rich document (VRD) utilizes visual features along with linguistic cues to disseminate information. Training a custom extractor that identifies named entities from a document requires a large number of instances of the target document type annotated at textual and visual modalities. This is an expensive bottleneck in enterprise scenarios, where we want to train custom extractors for thousands of different document types in a scalable way. Pre-training an extractor model on unlabeled instances of the target document type, followed by a fine-tuning step on human-labeled instances does not work in these scenarios, as it surpasses the maximum allowable training time allocated for the extractor. We address this scenario by proposing a Noise-Aware Training method or NAT in this paper. Instead of acquiring expensive human-labeled documents, NAT utilizes weakly labeled documents to train an extractor in a scalable way. To avoid degradation in the model's quality due to noisy, weakly labeled samples, NAT estimates the confidence of each training sample and incorporates it as uncertainty measure during training. We train multiple state-of-the-art extractor models using NAT. Experiments on a number of publicly available and in-house datasets show that NAT-trained models are not only robust in performance -- it outperforms a transfer-learning baseline by up to 6% in terms of macro-F1 score, but it is also more label-efficient -- it reduces the amount of human-effort required to obtain comparable performance by up to 73%.

Animesh Singh, Ritesh Sarkhel, Nibaran Das et al.

Finding local invariant patterns in handwrit-ten characters and/or digits for optical character recognition is a difficult task. Variations in writing styles from one person to another make this task challenging. We have proposed a non-explicit feature extraction method using a multi-scale multi-column skip convolutional neural network in this work. Local and global features extracted from different layers of the proposed architecture are combined to derive the final feature descriptor encoding a character or digit image. Our method is evaluated on four publicly available datasets of isolated handwritten Bangla characters and digits. Exhaustive comparative analysis against contemporary methods establishes the efficacy of our proposed approach.

Ritesh Sarkhel, Moniba Keymanesh, Arnab Nandi et al.

Abstractive summarization at controllable lengths is a challenging task in natural language processing. It is even more challenging for domains where limited training data is available or scenarios in which the length of the summary is not known beforehand. At the same time, when it comes to trusting machine-generated summaries, explaining how a summary was constructed in human-understandable terms may be critical. We propose Multi-level Summarizer (MLS), a supervised method to construct abstractive summaries of a text document at controllable lengths. The key enabler of our method is an interpretable multi-headed attention mechanism that computes attention distribution over an input document using an array of timestep independent semantic kernels. Each kernel optimizes a human-interpretable syntactic or semantic property. Exhaustive experiments on two low-resource datasets in the English language show that MLS outperforms strong baselines by up to 14.70% in the METEOR score. Human evaluation of the summaries also suggests that they capture the key concepts of the document at various length-budgets.

Bodhisatwa Mandal, Swarnendu Ghosh, Ritesh Sarkhel et al.

Capsule networks have gained a lot of popularity in short time due to its unique approach to model equivariant class specific properties as capsules from images. However the dynamic routing algorithm comes with a steep computational complexity. In the proposed approach we aim to create scalable versions of the capsule networks that are much faster and provide better accuracy in problems with higher number of classes. By using dynamic routing to extract intermediate features instead of generating output class specific capsules, a large increase in the computational speed has been observed. Moreover, by extracting equivariant feature capsules instead of class specific capsules, the generalization capability of the network has also increased as a result of which there is a boost in accuracy.

Bodhisatwa Mandal, Suvam Dubey, Swarnendu Ghosh et al.

Convolutional neural networks(CNNs) has become one of the primary algorithms for various computer vision tasks. Handwritten character recognition is a typical example of such task that has also attracted attention. CNN architectures such as LeNet and AlexNet have become very prominent over the last two decades however the spatial invariance of the different kernels has been a prominent issue till now. With the introduction of capsule networks, kernels can work together in consensus with one another with the help of dynamic routing, that combines individual opinions of multiple groups of kernels called capsules to employ equivariance among kernels. In the current work, we have implemented capsule network on handwritten Indic digits and character datasets to show its superiority over networks like LeNet. Furthermore, it has also been shown that they can boost the performance of other networks like LeNet and AlexNet.

Aritra Das, Swarnendu Ghosh, Ritesh Sarkhel et al.

Modern deep learning algorithms have triggered various image segmentation approaches. However most of them deal with pixel based segmentation. However, superpixels provide a certain degree of contextual information while reducing computation cost. In our approach, we have performed superpixel level semantic segmentation considering 3 various levels as neighbours for semantic contexts. Furthermore, we have enlisted a number of ensemble approaches like max-voting and weighted-average. We have also used the Dempster-Shafer theory of uncertainty to analyze confusion among various classes. Our method has proved to be superior to a number of different modern approaches on the same dataset.

Ritesh Sarkhel, Amit K Saha, Nibaran Das

Identification of minimum number of local regions of a handwritten character image, containing well-defined discriminating features which are sufficient for a minimal but complete description of the character is a challenging task. A new region selection technique based on the idea of an enhanced Harmony Search methodology has been proposed here. The powerful framework of Harmony Search has been utilized to search the region space and detect only the most informative regions for correctly recognizing the handwritten character. The proposed method has been tested on handwritten samples of Bangla Basic, Compound and mixed (Basic and Compound characters) characters separately with SVM based classifier using a longest run based feature-set obtained from the image subregions formed by a CG based quad-tree partitioning approach. Applying this methodology on the above mentioned three types of datasets, respectively 43.75%, 12.5% and 37.5% gains have been achieved in terms of region reduction and 2.3%, 0.6% and 1.2% gains have been achieved in terms of recognition accuracy. The results show a sizeable reduction in the minimal number of descriptive regions as well a significant increase in recognition accuracy for all the datasets using the proposed technique. Thus the time and cost related to feature extraction is decreased without dampening the corresponding recognition accuracy.