The field of multimodal document understanding has produced a suite of models that have achieved stellar performance across several tasks, even coming close to human performance on certain benchmarks. Nevertheless, the application of these models to real-world enterprise datasets remains constrained by a number of limitations. In this position paper, we discuss these limitations in the context of three key aspects of research: dataset curation, model development, and evaluation on downstream tasks. By analyzing 14 datasets and 7 SotA models, we identify major gaps in their utility in the context of a real-world scenario. We demonstrate how each limitation impedes the widespread use of SotA models in enterprise settings, and present a set of research challenges that are motivated by these limitations. Lastly, we propose a research agenda that is aimed at driving the field towards higher impact in enterprise applications.

The rise of large language models (LLMs) for visually rich document understanding (VRDU) has kindled a need for prompt-response, document-based datasets. As annotating new datasets from scratch is labor-intensive, the existing literature has generated prompt-response datasets from available resources using simple templates. For the case of key information extraction (KIE), one of the most common VRDU tasks, past work has typically employed the template “What is the value for the key?”. However, given the variety of questions encountered in the wild, simple and uniform templates are insufficient for creating robust models in research and industrial contexts. In this work, we present K2Q, a diverse collection of five datasets converted from KIE to a prompt-response format using a plethora of bespoke templates. The questions in K2Q can span multiple entities and be extractive or boolean. We empirically compare the performance of seven baseline generative models on K2Q with zero-shot prompting. We further compare three of these models when training on K2Q versus training on simpler templates to motivate the need of our work. We find that creating diverse and intricate KIE questions enhances the performance and robustness of VRDU models. We hope this work encourages future studies on data quality for generative model training.

Forms constitute a large portion of layout-rich documents that convey information through key-value pairs. Form understanding involves two main tasks, namely, the identification of keys and values (a.k.a Key Information Extraction or KIE) and the association of keys to corresponding values (a.k.a. Relation Extraction or RE). State of the art models for form understanding often rely on training paradigms that yield poorly calibrated output probabilities and low performance on RE. In this paper, we present AliGATr, a graph-based model that uses a generative objective to represent complex grid-like layouts that are often found in forms. Using a grid-based graph topology, our model learns to generate the layout of each page token by token in a data efficient manner. Despite using 30% fewer parameters than the smallest SotA, AliGATr performs on par with or better than SotA models on the KIE and RE tasks against four datasets. We also show that AliGATr’s output probabilities are better calibrated and do not exhibit the over-confident distributions of other SotA models.

Visually Rich Form Understanding (VRFU) poses a complex research problemdue to the documents’ highly structured nature and yet highly variable style and content. Current annotation schemes decompose form understanding and omit key hierarchical structure, making development and evaluation of end-to-end models difficult. In this paper, we propose a novel F1 metric to evaluate form parsers and describe a new content-agnostic, tree-based annotation scheme for VRFU: TreeForm. We provide methods to convert previous annotation schemes into TreeForm structures and evaluate TreeForm predictions using a modified version of the normalized tree-edit distance. We present initial baselines for our end-to-end performance metric and the TreeForm edit distance, averaged over the FUNSD and XFUND datasets, of 61.5 and 26.4 respectively. We hope that TreeForm encourages deeper research in annotating, modeling, and evaluating the complexities of form-like documents.

Enterprise documents such as forms, receipts, reports, and other such records, often carry rich semantics at the intersection of textual and spatial modalities. The visual cues offered by their complex layouts play a crucial role in comprehending these documents effectively. In this paper, we present DocLLM, a lightweight extension to traditional large language models (LLMs) for reasoning over visual documents, taking into account both textual semantics and spatial layout. Our model differs from existing multimodal LLMs by avoiding expensive image encoders and focuses exclusively on bounding box information to incorporate the spatial layout structure. Specifically, the cross-alignment between text and spatial modalities is captured by decomposing the attention mechanism in classical transformers to a set of disentangled matrices. Furthermore, we devise a pre-training objective that learns to infill text segments. This approach allows us to address irregular layouts and heterogeneous content frequently encountered in visual documents. The pre-trained model is fine-tuned using a large-scale instruction dataset, covering four core document intelligence tasks. We demonstrate that our solution outperforms SotA LLMs on 14 out of 16 datasets across all tasks, and generalizes well to 4 out of 5 previously unseen datasets.

In quantitative question answering, compositional generalization is one of the main challenges of state of the art models, especially when longer sequences of reasoning steps are required. In this paper we propose CounterComp, a method that uses counterfactual scenarios to generate samples with compositional contrast. Instead of a data augmentation approach, CounterComp is based on metric learning, which allows for direct sampling from the training set and circumvents the need for additional human labels. Our proposed auxiliary metric learning loss improves the performance of three state of the art models on four recently released datasets. We also show how the approach can improve OOD performance on unseen domains, as well as unseen compositions. Lastly, we demonstrate how the method can lead to better compositional attention patterns during training.

Quantitative reasoning is an important aspect of question answering, especially when numeric and verbal cues interact to indicate sophisticated, multi-step programs. In this paper, we demonstrate how modeling the compositional nature of quantitative text can enhance the performance and robustness of QA models, allowing them to capture arithmetic logic that is expressed verbally. Borrowing from the literature on semantic parsing, we propose a method that encourages the QA models to adjust their attention patterns and capture input/output alignments that are meaningful to the reasoning task. We show how this strategy improves program accuracy and renders the models more robust against overfitting as the number of reasoning steps grows. Our approach is designed as a standalone module which can be prepended to many existing models and trained in an end-to-end fashion without the need for additional supervisory signal. As part of this exercise, we also create a unified dataset building on four previously released numerical QA datasets over tabular data.

This paper describes the approach we used for SemEval-2017 Task 4: Sentiment Analysis in Twitter. Topic-based (target-dependent) sentiment analysis has become attractive and been used in some applications recently, but it is still a challenging research task. In our approach, we take the left and right context of a target into consideration when generating polarity classification features. We use two types of word embeddings in our classifiers: the general word embeddings learned from 200 million tweets, and sentiment-specific word embeddings learned from 10 million tweets using distance supervision. We also incorporate a text feature model in our algorithm. This model produces features based on text negation, tf.idf weighting scheme, and a Rocchio text classification method. We participated in four subtasks (B, C, D & E for English), all of which are about topic-based message polarity classification. Our team is ranked #6 in subtask B, #3 by MAEu and #9 by MAEm in subtask C, #3 using RAE and #6 using KLD in subtask D, and #3 in subtask E.

This paper describes the approach we used for SemEval-2017 Task 5: Fine-Grained Sentiment Analysis on Financial Microblogs. We use three types of word embeddings in our algorithm: word embeddings learned from 200 million tweets, sentiment-specific word embeddings learned from 10 million tweets using distance supervision, and word embeddings learned from 20 million StockTwits messages. In our approach, we also take the left and right context of the target company into consideration when generating polarity prediction features. All the features generated from different word embeddings and contexts are integrated together to train our algorithm