-
Document Understanding Dataset and Evaluation (DUDE)
Authors:
Jordy Van Landeghem,
Rubén Tito,
Łukasz Borchmann,
Michał Pietruszka,
Paweł Józiak,
Rafał Powalski,
Dawid Jurkiewicz,
Mickaël Coustaty,
Bertrand Ackaert,
Ernest Valveny,
Matthew Blaschko,
Sien Moens,
Tomasz Stanisławek
Abstract:
We call on the Document AI (DocAI) community to reevaluate current methodologies and embrace the challenge of creating more practically-oriented benchmarks. Document Understanding Dataset and Evaluation (DUDE) seeks to remediate the halted research progress in understanding visually-rich documents (VRDs). We present a new dataset with novelties related to types of questions, answers, and document…
▽ More
We call on the Document AI (DocAI) community to reevaluate current methodologies and embrace the challenge of creating more practically-oriented benchmarks. Document Understanding Dataset and Evaluation (DUDE) seeks to remediate the halted research progress in understanding visually-rich documents (VRDs). We present a new dataset with novelties related to types of questions, answers, and document layouts based on multi-industry, multi-domain, and multi-page VRDs of various origins, and dates. Moreover, we are pushing the boundaries of current methods by creating multi-task and multi-domain evaluation setups that more accurately simulate real-world situations where powerful generalization and adaptation under low-resource settings are desired. DUDE aims to set a new standard as a more practical, long-standing benchmark for the community, and we hope that it will lead to future extensions and contributions that address real-world challenges. Finally, our work illustrates the importance of finding more efficient ways to model language, images, and layout in DocAI.
△ Less
Submitted 11 September, 2023; v1 submitted 15 May, 2023;
originally announced May 2023.
-
STable: Table Generation Framework for Encoder-Decoder Models
Authors:
Michał Pietruszka,
Michał Turski,
Łukasz Borchmann,
Tomasz Dwojak,
Gabriela Pałka,
Karolina Szyndler,
Dawid Jurkiewicz,
Łukasz Garncarek
Abstract:
The output structure of database-like tables, consisting of values structured in horizontal rows and vertical columns identifiable by name, can cover a wide range of NLP tasks. Following this constatation, we propose a framework for text-to-table neural models applicable to problems such as extraction of line items, joint entity and relation extraction, or knowledge base population. The permutatio…
▽ More
The output structure of database-like tables, consisting of values structured in horizontal rows and vertical columns identifiable by name, can cover a wide range of NLP tasks. Following this constatation, we propose a framework for text-to-table neural models applicable to problems such as extraction of line items, joint entity and relation extraction, or knowledge base population. The permutation-based decoder of our proposal is a generalized sequential method that comprehends information from all cells in the table. The training maximizes the expected log-likelihood for a table's content across all random permutations of the factorization order. During the content inference, we exploit the model's ability to generate cells in any order by searching over possible orderings to maximize the model's confidence and avoid substantial error accumulation, which other sequential models are prone to. Experiments demonstrate a high practical value of the framework, which establishes state-of-the-art results on several challenging datasets, outperforming previous solutions by up to 15%.
△ Less
Submitted 12 October, 2022; v1 submitted 8 June, 2022;
originally announced June 2022.
-
Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer
Authors:
Rafał Powalski,
Łukasz Borchmann,
Dawid Jurkiewicz,
Tomasz Dwojak,
Michał Pietruszka,
Gabriela Pałka
Abstract:
We address the challenging problem of Natural Language Comprehension beyond plain-text documents by introducing the TILT neural network architecture which simultaneously learns layout information, visual features, and textual semantics. Contrary to previous approaches, we rely on a decoder capable of unifying a variety of problems involving natural language. The layout is represented as an attenti…
▽ More
We address the challenging problem of Natural Language Comprehension beyond plain-text documents by introducing the TILT neural network architecture which simultaneously learns layout information, visual features, and textual semantics. Contrary to previous approaches, we rely on a decoder capable of unifying a variety of problems involving natural language. The layout is represented as an attention bias and complemented with contextualized visual information, while the core of our model is a pretrained encoder-decoder Transformer. Our novel approach achieves state-of-the-art results in extracting information from documents and answering questions which demand layout understanding (DocVQA, CORD, SROIE). At the same time, we simplify the process by employing an end-to-end model.
△ Less
Submitted 12 July, 2021; v1 submitted 18 February, 2021;
originally announced February 2021.
-
From Dataset Recycling to Multi-Property Extraction and Beyond
Authors:
Tomasz Dwojak,
Michał Pietruszka,
Łukasz Borchmann,
Jakub Chłędowski,
Filip Graliński
Abstract:
This paper investigates various Transformer architectures on the WikiReading Information Extraction and Machine Reading Comprehension dataset. The proposed dual-source model outperforms the current state-of-the-art by a large margin. Next, we introduce WikiReading Recycled-a newly developed public dataset and the task of multiple property extraction. It uses the same data as WikiReading but does n…
▽ More
This paper investigates various Transformer architectures on the WikiReading Information Extraction and Machine Reading Comprehension dataset. The proposed dual-source model outperforms the current state-of-the-art by a large margin. Next, we introduce WikiReading Recycled-a newly developed public dataset and the task of multiple property extraction. It uses the same data as WikiReading but does not inherit its predecessor's identified disadvantages. In addition, we provide a human-annotated test set with diagnostic subsets for a detailed analysis of model performance.
△ Less
Submitted 6 November, 2020;
originally announced November 2020.
-
Successive Halving Top-k Operator
Authors:
Michał Pietruszka,
Łukasz Borchmann,
Filip Graliński
Abstract:
We propose a differentiable successive halving method of relaxing the top-k operator, rendering gradient-based optimization possible. The need to perform softmax iteratively on the entire vector of scores is avoided by using a tournament-style selection. As a result, a much better approximation of top-k with lower computational cost is achieved compared to the previous approach.
We propose a differentiable successive halving method of relaxing the top-k operator, rendering gradient-based optimization possible. The need to perform softmax iteratively on the entire vector of scores is avoided by using a tournament-style selection. As a result, a much better approximation of top-k with lower computational cost is achieved compared to the previous approach.
△ Less
Submitted 8 October, 2020;
originally announced October 2020.
-
Sparsifying Transformer Models with Trainable Representation Pooling
Authors:
Michał Pietruszka,
Łukasz Borchmann,
Łukasz Garncarek
Abstract:
We propose a novel method to sparsify attention in the Transformer model by learning to select the most-informative token representations during the training process, thus focusing on the task-specific parts of an input. A reduction of quadratic time and memory complexity to sublinear was achieved due to a robust trainable top-$k$ operator. Our experiments on a challenging long document summarizat…
▽ More
We propose a novel method to sparsify attention in the Transformer model by learning to select the most-informative token representations during the training process, thus focusing on the task-specific parts of an input. A reduction of quadratic time and memory complexity to sublinear was achieved due to a robust trainable top-$k$ operator. Our experiments on a challenging long document summarization task show that even our simple baseline performs comparably to the current SOTA, and with trainable pooling, we can retain its top quality, while being $1.8\times$ faster during training, $4.5\times$ faster during inference, and up to $13\times$ more computationally efficient in the decoder.
△ Less
Submitted 7 March, 2022; v1 submitted 10 September, 2020;
originally announced September 2020.
-
On the Multi-Property Extraction and Beyond
Authors:
Tomasz Dwojak,
Michał Pietruszka,
Łukasz Borchmann,
Filip Graliński,
Jakub Chłędowski
Abstract:
In this paper, we investigate the Dual-source Transformer architecture on the WikiReading information extraction and machine reading comprehension dataset. The proposed model outperforms the current state-of-the-art by a large margin. Next, we introduce WikiReading Recycled - a newly developed public dataset, supporting the task of multiple property extraction. It keeps the spirit of the original…
▽ More
In this paper, we investigate the Dual-source Transformer architecture on the WikiReading information extraction and machine reading comprehension dataset. The proposed model outperforms the current state-of-the-art by a large margin. Next, we introduce WikiReading Recycled - a newly developed public dataset, supporting the task of multiple property extraction. It keeps the spirit of the original WikiReading but does not inherit the identified disadvantages of its predecessor.
△ Less
Submitted 15 June, 2020;
originally announced June 2020.
-
Modified Bohm's theory for abstruse measurements: application to layer depth profiling by Auger spectroscopy
Authors:
Edward Rówiński,
Mariusz Pietruszka
Abstract:
Modified Bohm formalism is applied to solve a problem of abstruse layer depth profiles measured by the Auger electron spectroscopy technique in real physical systems, i.e., the desorbed carbon/passive layer on NiTi substrate and the adsorbed oxygen/surface of NiTi alloy. It is shown that abstruse layer profiles may be converted to real layer structures using the modified Bohm theory, where the qua…
▽ More
Modified Bohm formalism is applied to solve a problem of abstruse layer depth profiles measured by the Auger electron spectroscopy technique in real physical systems, i.e., the desorbed carbon/passive layer on NiTi substrate and the adsorbed oxygen/surface of NiTi alloy. It is shown that abstruse layer profiles may be converted to real layer structures using the modified Bohm theory, where the quantum potential is due to an Auger electron effect. It is also pointed out that the stationary probability density predicts multilayer structures of abstruse depth profiles caused by carbon desorption and oxygen adsorption processes. The criterion for a kind of break between the physical and unphysical multilayer systems was found. We have concluded that the physics is also characterized by the abstruse measurement and the modified Bohm formalism
△ Less
Submitted 18 February, 2019; v1 submitted 17 October, 2016;
originally announced October 2016.
-
Effective diffusion rates and cross-correlation analysis of "acid growth" data
Authors:
Mariusz Pietruszka,
Aleksandra Haduch-Sendecka
Abstract:
We investigated the growth-temperature relationship in plants using a quantitative perspective of a recently derived growth functional. We showed that auxin-induced growth is achieved by the diffusion rate, which is almost constant or slowly ascending in temperature while the diffusion rate of fusicoccin-induced growth increases monotonically with temperature for the entire temperature range, thou…
▽ More
We investigated the growth-temperature relationship in plants using a quantitative perspective of a recently derived growth functional. We showed that auxin-induced growth is achieved by the diffusion rate, which is almost constant or slowly ascending in temperature while the diffusion rate of fusicoccin-induced growth increases monotonically with temperature for the entire temperature range, though for some concentrations of IAA "super-diffusion" takes place for unperturbed growth. Furthermore, three kind of experiments were compared: for abraded coleoptiles, coleoptile segments and intact growing seedlings. From cross-correlation analysis it was found that the timing of IAA and FC-induced proton secretion and growth matches well. Unambiguous results, concerning fundamental conditions of the acid growth hypothesis, were obtained by cross- and auto-correlation analysis: (1) For abraded coleoptiles, because of the lowering of the cuticule potential barrier, auxin-induced cell wall pH decreases simultaneously with the change in growth rate; no advancement or retardation of pH (proton efflux rate) or growth rate takes place (2) Exogenous protons are able to substitute for auxin causing wall loosening and growth (3) Although the underlying molecular mechanisms vastly differ, a potent stimulator of proton secretion, the fungal toxin FC, promotes growth similar to auxin, however of much elevated intensity; as for auxin - no advancement or retardation takes place.
△ Less
Submitted 3 August, 2015;
originally announced August 2015.
-
Frequency landscape of tip-growing plants
Authors:
Mariusz Pietruszka,
Aleksandra Haduch-Sendecka
Abstract:
It has been interesting that nearly all of the ion activities that have been analysed thus far have exhibited oscillations that are tightly coupled to growth. Here, we present discrete Fourier transform (DFT) spectra with a finite sampling of tip-growing cells and organs that were obtained from voltage measurements of the elongating coleoptiles of maize in situ. The electromotive force (EMF) oscil…
▽ More
It has been interesting that nearly all of the ion activities that have been analysed thus far have exhibited oscillations that are tightly coupled to growth. Here, we present discrete Fourier transform (DFT) spectra with a finite sampling of tip-growing cells and organs that were obtained from voltage measurements of the elongating coleoptiles of maize in situ. The electromotive force (EMF) oscillations (~ 0.1 μV) were measured in a simple but highly sensitive RL circuit, in which the solenoid was initially placed at the tip of the specimen and then was moved thus changing its position in relation to growth (EMF can be measured first at the tip, then at the sub-apical part and finally at the shank). The influx- and efflux-induced oscillations of Ca$^{2+}$, along with H$^{+}$, K$^{+}$ and Cl$^{-}$ were densely sampled (preserving the Nyquist theorem in order to 'grasp the structure' of the pulse), the logarithmic amplitude of pulse spectrum was calculated, and the detected frequencies, which displayed a periodic sequence of pulses, were compared with the literature data. A band of life vital individual pulses was obtained in a single run of the experiment, which not only allowed the fundamental frequencies (and intensities of the processes) to be determined but also permitted the phase relations of the various transport processes in the plasma membrane and tonoplast to be established. A discrete frequency spectrum (like the hydrogen spectrum in quantum physics) was achieved for a growing plant for the first time, while all of the metabolic and enzymatic functions of the life cell cycle were preserved using this totally non-invasive treatment.
△ Less
Submitted 28 July, 2015; v1 submitted 1 June, 2015;
originally announced June 2015.
-
pH/$T$ duality - wall properties and time evolution of plant cells
Authors:
Mariusz A. Pietruszka
Abstract:
We examined the pH/$T$ (or $μ$/$T$) duality of acidic pH and temperature ($T$) for the growth of grass shoots in order to determine the equation of state (EoS) for living plants. By considering non-meristematic growth as a dynamic series of 'state transitions' (STs) in the extending primary wall, we identified the critical (read: optimum) exponents for this phenomenon, which exhibit a singular beh…
▽ More
We examined the pH/$T$ (or $μ$/$T$) duality of acidic pH and temperature ($T$) for the growth of grass shoots in order to determine the equation of state (EoS) for living plants. By considering non-meristematic growth as a dynamic series of 'state transitions' (STs) in the extending primary wall, we identified the critical (read: optimum) exponents for this phenomenon, which exhibit a singular behaviour at a critical temperature, critical pH and critical chemical potential ($μ$) in the form of four power laws: $F(τ)_π\propto|τ|^{β-1}$, $F(π)_τ\propto|π|^{1-α}$, $G(τ)_μ\propto|τ|^{-2-α+2β}$ and $G(μ)_τ\propto|μ|^{2-α}$. The power-law exponents $α$ and $β$ are numbers, which are independent of pH (or $μ$) and $T$ that are known as critical exponents, while $π$ and $τ$ represent a reduced pH and reduced temperature, respectively. Various 'scaling' predictions were obtained - a convexity relation $α+ β\ge 2$ for practical pH-based analysis and a $β\equiv 2$ identity in microscopic representation. In the presented scenario, the magnitude that is decisive is the chemical potential of H$^+$ ions (protons), enforcing subsequent STs and growth. The EoS span areas of the biological, physical, chemical and Earth sciences cross the borders with the language (adapted formalism) of phase transitions.
△ Less
Submitted 27 July, 2017; v1 submitted 2 May, 2015;
originally announced May 2015.
-
Frustration-induced inherent instability and growth oscillations in pollen tubes
Authors:
Mariusz Pietruszka
Abstract:
In a seed plant a pollen tube is a vessel that transports male gamete cells to an ovule to achieve fertilization. It consists of one elongated cell, which exhibits growth oscillations, until it bursts completing its function. Up till now, the mechanism behind the periodic character of the growth has not been fully understood. An attempt to understand these oscillations lead us to an attractive sce…
▽ More
In a seed plant a pollen tube is a vessel that transports male gamete cells to an ovule to achieve fertilization. It consists of one elongated cell, which exhibits growth oscillations, until it bursts completing its function. Up till now, the mechanism behind the periodic character of the growth has not been fully understood. An attempt to understand these oscillations lead us to an attractive scenario: We show that the mechanism of pressure-induced symmetry frustration occuring in the wall at the perimeter of cylindrical and approximately hemispherical parts of a growing pollen cell, together with the addition of cell wall material, suffices to release and sustain mechanical self-oscillations and cell extension in pollen tubes. At the transition zone where symmetry frustration occurs and one cannot distinguish either of the involved symmetries, a kind of 'entangled state' appears where either single or both symmetry(ies) can be realized by the system. We anticipate that testifiable predictions made by the model may deliver, after calibration, a new tool to estimate turgor pressure from oscillation period of the growing cell. Since the mechanical principles apply to all turgor regulated walled cells including those of plant, fungal and bacterial origin, the relevance of this work is not limited to the case of the pollen tube.
△ Less
Submitted 31 December, 2012; v1 submitted 6 November, 2012;
originally announced November 2012.
