Retrieval Research Research Articles

Text-Enhanced Graph Attention Hashing for Cross-Modal Retrieval

Deep hashing technology, known for its low-cost storage and rapid retrieval, has become a focal point in cross-modal retrieval research as multimodal data continue to grow. However, existing supervised methods often overlook noisy labels and multiscale features in different modal datasets, leading to higher information entropy in the generated hash codes and features, which reduces retrieval performance. The variation in text annotation information across datasets further increases the information entropy during text feature extraction, resulting in suboptimal outcomes. Consequently, reducing the information entropy in text feature extraction, supplementing text feature information, and enhancing the retrieval efficiency of large-scale media data are critical challenges in cross-modal retrieval research. To tackle these, this paper introduces the Text-Enhanced Graph Attention Hashing for Cross-Modal Retrieval (TEGAH) framework. TEGAH incorporates a deep text feature extraction network and a multiscale label region fusion network to minimize information entropy and optimize feature extraction. Additionally, a Graph-Attention-based modal feature fusion network is designed to efficiently integrate multimodal information, enhance the affinity of the network for different modes, and retain more semantic information. Extensive experiments on three multilabel datasets demonstrate that the TEGAH framework significantly outperforms state-of-the-art cross-modal hashing methods.

Improving coniferous forests leaf area index estimation by filling the occluded point cloud from airborne laser scanning

Accurate estimation of Leaf Area Index (LAI) is pivotal for understanding vegetation dynamics and ecosystem productivity. Light Detection and Ranging (LiDAR), with its capability to directly capture the three-dimensional (3D) structure of vegetation, offers a promising method for LAI estimation. However, dense leaves, especially coniferous forest, can significantly occlude the penetration of laser beams during Airborne Laser Scanning (ALS) as they traverse the canopy from above. This occlusion results in incomplete point cloud, leading to underestimation of LAI due to the omission of some leaves. Considering the characteristic of leaf clumping growth in coniferous vegetation, a method based on voxelizing point cloud and modeling correlations among neighborhood voxels was developed in this study to fill the occluded point cloud at plot level. The proposed method was validated by applying it to Loblolly pine (Pinus taeda L.) plots. Validation against field-measured LAI demonstrates that the method proposed for filling the occluded point cloud markedly enhances LAI retrieval accuracy, achieving an R2 of 0.7069 and an RMSE of 0.4760. In contrast, LAI retrieval using the original canopy point cloud yielded an R2 of 0.2835 and an RMSE of 1.5883. This advancement highlights a shift from focusing on enhancing retrieval algorithms to addressing data completeness and reliability, offering new perspectives for LAI retrieval research.

Boosting cross-modal retrieval in remote sensing via a novel unified attention network

With the rapid advent and abundance of remote sensing data in different modalities, cross-modal retrieval tasks have gained importance in the research community. Cross-modal retrieval belongs to the research paradigm in which the query is of one modality and the retrieved output is of the other modality. In this paper, the remote sensing (RS) data modalities considered are the earth observation optical data (aerial photos) and the corresponding hand-drawn sketches. The main challenge of the cross-modal retrieval research objective for optical remote sensing images and the corresponding sketches is the distribution gap between the shared embedding space of the modalities. Prior attempts to resolve this issue have not yielded satisfactory outcomes regarding accurately retrieving cross-modal sketch-image RS data. The state-of-the-art architectures used conventional convolutional architectures, which focused on local pixel-wise information about the modalities to be retrieved. This limits the interaction between the sketch texture and the corresponding image, making these models susceptible to overfitting datasets with particular scenarios. To circumvent this limitation, we suggest establishing multi-modal correspondence using a novel architecture of the combined self and cross-attention algorithms, SPCA-Net to minimize the modality gap by employing attention mechanisms for the query and other modalities. Efficient cross-modal retrieval is achieved through the suggested attention architecture, which empirically emphasizes the global information of the relevant query modality and bridges the domain gap through a unique pairwise cross-attention network. In addition to the novel architecture, this paper introduces a unique loss function, label-specific supervised contrastive loss, tailored to the intricacies of the task and to enhance the discriminative power of the learned embeddings. Extensive evaluations are conducted on two sketch-image remote sensing datasets, Earth-on-Canvas and RSketch. Under the same experimental conditions, the performance metrics of our proposed model beat the state-of-the-art architectures by significant margins of 16.7%, 18.9%, 33.7%, and 40.9% correspondingly.

Surveying More Than Two Decades of Music Information Retrieval Research on Playlists

In this paper, we present an extensive survey of music information retrieval (MIR) research into music playlists. Our survey spans more than 20 years, and includes around 300 papers about playlists, with over 70 supporting sources. It is the first survey that is self-contained in the sense that it combines all the different MIR research into playlists. It embraces topics such as algorithms for automatic generation, for automatic continuation, for assisting with manual generation, for tagging and for captioning. It looks at manually constructed playlists, both those that are constructed for and by individuals, and those constructed in collaboration with others. It covers ground-breaking research into enhancing playlists by cross-fading consecutive songs and by interleaving consecutive songs with speech, similar to what happens on a radio show. Most significantly, it is the first survey that can fully incorporate the paradigm shift that has taken place in the way people consume recorded music: the shift from physical media to music streaming. This has wrought profound changes in the size of music collections available to listeners and thus the algorithms that support the construction, curation and presentation of playlists and the methods adopted by users when they also construct, curate and listen to playlists.

SKYPER: Legal case retrieval via skeleton-aware hypergraph embedding in the hyperbolic space

Legal Case Retrieval (LCR) plays a significant role in ensuring judicial justice in various legal systems and has recently received increasing attention in Information Retrieval (IR) research. Existing LCR methods typically encode documents into low-dimensional vectors by using either bag of keywords or pretrained language models. However, such approaches generally overlook the interactions among numerous cases and the hierarchy skeleton between event and event type. In this paper, we propose to construct an event-aware knowledge hypergraph for representing legal cases. And we put forward a SKeleton-aware hYPERgraph representation framework named SKYPER, to learn the case embeddings. Concretely, a global skeleton aggregates coarse-grained type information and roughly locates a case. A local skeleton learns the fine-grained information and precisely locates a case. And SKYPER also uses hyperbolic space strengthens the hierarchy between event and event type and avoids vocabulary mismatch. At the retrieval stage, the SKYPER learned case embedding is leveraged for further case similarity matching with a few annotated similarity ranking data. Extensive experiments demonstrate that SKYPER outperforms existing models on the public LCR datasets LeCard and COLIEE20-T1 in both supervised and unsupervised settings. Especially, the retrieval efficiency is improved by a large margin.

The Chosen Ummah By Muhammad 'Ali As-Shabuni (Interpretation in Q.S. Al-Baqarah 143 and Q.S. Ali ‘Imran 110)

Muslims today, in various aspects of life, are experiencing conditions or circumstances that are increasingly lagging behind Western nations. One of the causes is the division among Muslims themselves, which resulted in the emergence of several groups of thought that pose a danger and threat to the unity of the people. Religious moderation is present in society to prevent the emergence of extreme thinking and can create a living environment that is in harmony and peace. The book of Shafwah at-Tafasir by Muhammad ‘Ali as-Shabuni is used in this article because it uses light language and is easy to understand for readers. as-Shabuni is a scholar who pays attention to the condition of humankind. This article explores as-Shabuni's thoughts on the verses of religious moderation in this interpretation about The Chosen Ummah. The author uses a thematic method in describing the interpretation of as-Shabuni contained in the book Shafwah at-Tafasir. This article is qualitative research and uses library research in data retrieval to include information related to the discussion. The result of this study is that the interpretation as-Shabuni on Q.S. al-Baqarah: 143 is Allah made Muslims the best and chosen people, being in the Middle, not exaggerating and diminishing an act. They are witnesses to the day of judgment. In Q.S. Ali Imran:110, the chosen Ummah means a Ummah that benefits other Ummah, and this is intended for Muslims. In this case, the chosen people referred to by as-Shabuni are the best people (Muslims/Muhammad's Ummah), who do not add or subtract actions that result in being hated by God, people who are helpful because they command to do good and prevent the wicked and believe in Allah SWT, these actions are forms of religious moderation actions

Chinese Title Generation for Short Videos: Dataset, Metric and Algorithm.

Previous work for video captioning aims to objectively describe the video content but the captions lack human interest and attractiveness, limiting its practical application scenarios. The intention of video title generation (video titling) is to produce attractive titles, but there is a lack of benchmarks. This work offers CREATE, the first large-scale Chinese shoRt vidEo retrievAl and Title gEneration dataset, to assist research and applications in video titling, video captioning, and video retrieval in Chinese. CREATE comprises a high-quality labeled 210 K dataset and two web-scale 3 M and 10 M pre-training datasets, covering 51 categories, 50K+ tags, 537K+ manually annotated titles and captions, and 10M+ short videos with original video information. This work presents ACTEr, a unique Attractiveness-Consensus-based Title Evaluation, to objectively evaluate the quality of video title generation. This metric measures the semantic correlation between the candidate (model-generated title) and references (manual-labeled titles) and introduces attractive consensus weights to assess the attractiveness and relevance of the video title. Accordingly, this work proposes a novel multi-modal ALignment WIth Generation model, ALWIG, as one strong baseline to aid future model development. With the help of a tag-driven video-text alignment module and a GPT-based generation module, this model achieves video titling, captioning, and retrieval simultaneously. We believe that the release of the CREATE dataset, ACTEr metric, and ALWIG model will encourage in-depth research on the analysis and creation of Chinese short videos.

New Algorithms and Methodologies for Building Information Retrieval Collections

Information retrieval systems play a crucial role in addressing users' information needs by aiding their exploration of vast collections of information. This thesis is framed in a critical information retrieval research aspect: evaluation. In particular, we propose new approaches for creating annotated test collections. Such collections are essential for evaluating retrieval systems' effectiveness in controlled experiments. Reflecting real-world conditions accurately in these test collections is pivotal for progress in the field. We aim to introduce innovative techniques for efficiently assembling reliable test collections, facilitating broader research and development in information retrieval. The thesis first proposes a new method for building new pooled test collections without requiring costly evaluation campaigns [Otero et al., 2021b]. This approach simplifies and economizes the process of building new benchmarks. Then, we introduce a novel adjudication method for determining which pooled documents warrant human judgment, aiming to reduce the need for extensive expert assessments [Otero et al., 2023a]. This method is both cost-effective and efficient. Additionally, the thesis presents a fresh perspective on evaluating adjudication methods, emphasizing statistical significance, an aspect often overlooked in previous document adjudication research [Otero et al., 2023b]. As a demonstration of the methods explored in this thesis, we applied them to develop a new test collection whose construction process we describe here as an example of the use of reduced-budget methods [Otero et al., 2021a, 2020]. In summary, this thesis integrates established information retrieval knowledge with new methodologies to create annotated collections that are both cost-effective and reliable. This fusion is crucial for advancing the development of more effective retrieval systems. Awarded by : Universidade da Coruña, A Coruña, Spain on 5 April 2024. Supervised by : Javier Parapar, Álvaro Barreiro. Available at : http://hdl.handle.net/2183/36378.

Towards bridging the semantic gap between image and text: An empirical approach

Vlogs, Recordings, news, sport coverages are huge sources of multimodal information that do not just limit to text but extend to audio, images and videos. Applications such as summary generation, image/video captioning, multimodal sentiment analysis, cross modal retrieval requires Computer Vision along with Natural Language Processing techniques to extract relevant information. Information from different modalities must be leveraged in order to extract quality content. Hence, reducing the gap between different modalities is of utmost importance. Image to text conversion is an emerging field and employs the use of encoder decoder architecture. Deep CNNs extract the feature of images and sequence to sequence models are used to generate text description. This paper is a contribution towards the growing body of research in multimodal information retrieval. In order to generate the textual description of images, we have performed 5 experiments using the benchmark Flickr8k dataset. In these experiments we have utilized different architectures - simple sequence to sequence model, attention mechanism, transformer-based architecture to name a few. The results have been evaluated using BLEAU score. Results show that the best descriptions are attained by making use of transformer architecture. We have also compared our results with the pretrained visual model vit-gpt2 that incorporates visual transformer.

NhanesA: achieving transparency and reproducibility in NHANES research.

The National Health and Nutrition Examination Survey provides comprehensive data on demographics, sociology, health and nutrition. Conducted in 2-year cycles since 1999, most of its data are publicly accessible, making it pivotal for research areas like studying social determinants of health or tracking trends in health metrics such as obesity or diabetes. Assembling the data and analyzing it presents a number of technical and analytic challenges. This paper introduces the nhanesA R package, which is designed to assist researchers in data retrieval and analysis and to enable the sharing and extension of prior research efforts. We believe that fostering community-driven activity in data reproducibility and sharing of analytic methods will greatly benefit the scientific community and propel scientific advancements. Database URL: https://github.com/cjendres1/nhanes.

Precipitation Estimation Using FY-4B/AGRI Satellite Data Based on Random Forest

Precipitation is the basic component of the Earth’s water cycle. Obtaining high-resolution and high-precision precipitation data is of great significance. This paper establishes a precipitation retrieval model based on a random forest classification and regression model during the day and at night with FY-4B/AGRI Level1 data on China from July to August 2022. To evaluate the retrieval effect of the model, the GPM IMERG product is used as a reference, and the retrieval results are compared against those of the FY-4B/AGRI operational precipitation product. In addition, the retrieval results are analyzed according to different underlying surfaces. The results showed that compared with the FY-4B/AGRI operational precipitation product, the retrieval model can better identify precipitation and capture precipitation areas of light rain, moderate rain, heavy rain and torrential rain. Among them, the probability of detection (POD) of the day model increased from 0.328 to 0.680, and the equitable threat score (ETS) increased from 0.252 to 0.432. The POD of the night model increased from 0.337 to 0.639, and the ETS score increased from 0.239 to 0.369. Meanwhile, the precipitation estimation accuracy of the day model increased by 38.98% and that of the night model increased by 40.85%. Our results also showed that due to the surface uniformity of the ocean, the model can identify precipitation better on the ocean than on the land. Our findings also indicated that for the different underlying surfaces of the land, there is no significant difference in each evaluation index of the model. This is a strong argument for the universal applicability of the model. Notably, the results showed that, especially for more vegetated areas and areas covered by water, the model is capable of estimating precipitation. In conclusion, the precipitation retrieval model that is proposed herein can better determine precipitation regions and estimate precipitation intensities compared with the FY-4B/AGRI operational precipitation product. It can provide some reference value for future precipitation retrieval research on FY-4B/AGRI.

Semantic Image Retrieval Analysis Based on Deep Learning and Singular Value Decomposition

The exponential growth in the total quantity of digital images has necessitated the development of systems that are capable of retrieving these images. Content-based image retrieval is a technique used to get images from a database. The user provides a query image, and the system retrieves those photos from the database that are most similar to the query image. The image retrieval problem pertains to the task of locating digital photographs inside extensive datasets. Image retrieval researchers are transitioning from the use of keywords to the utilization of low-level characteristics and semantic features. The push for semantic features arises from the issue of subjective and time-consuming keywords, as well as the limitation of low-level characteristics in capturing high-level concepts that users have in mind. The main goal of this study is to examine how convolutional neural networks can be used to acquire advanced visual features. These high-level feature descriptors have the potential to be the most effective compared to the handcrafted feature descriptors in terms of image representation, which would result in improved image retrieval performance. The (CBIR-VGGSVD) model is an ideal solution for content-based image retrieval that is based on the VGG-16 algorithm and uses the Singular Value Decomposition (SVD) technique. The suggested model incorporates the VGG-16 model for the purpose of extracting features from both the query images and the images kept in the database. Afterwards, the dimensionality of the features retrieved from the VGG-16 model is reduced using SVD. Then, we compare the query photographs to the dataset images using the cosine metric to see how similar they are. When all is said and done, images that share a high degree of similarity will be successfully extracted from the dataset. A validation of the retrieval performance of the CBIR-VGGSVD model is performed using the Corel-1K dataset. When the VGG-16 standard model is the sole one used, the implementation will produce an average precision of 0.864. On the other hand, when the CBIR-VGGSVD model is utilized, this average precision is revealed to be (0.948). The findings of the retrieval ensured that the CBIR-VGGSVD model provided an improvement in performance on the test pictures that were utilized, surpassing the performance of the most recent approaches.

Multimodal learning with only image data: A deep unsupervised model for street view image retrieval by fusing visual and scene text features of images

AbstractAs one of the classic tasks in information retrieval, the core of image retrieval is to identify the images sharing similar features with a query image, aiming to enable users to find the required information from a large number of images conveniently. Street view image retrieval, in particular, finds extensive applications in many fields, such as improvements to navigation and mapping services, formulation of urban development planning scheme, and analysis of historical evolution of buildings. However, the intricate foreground and background details in street view images, coupled with a lack of attribute annotations, render it among the most challenging issues in practical applications. Current image retrieval research mainly uses the visual model that is completely dependent on the image visual features, and the multimodal learning model that necessitates additional data sources (e.g., annotated text). Yet, creating annotated datasets is expensive, and street view images, which contain a large amount of scene texts themselves, are often unannotated. Therefore, this paper proposes a deep unsupervised learning algorithm that combines visual and text features from image data for improving the accuracy of street view image retrieval. Specifically, we employ text detection algorithms to identify scene text, utilize the Pyramidal Histogram of Characters encoding predictor model to extract text information from images, deploy deep convolutional neural networks for visual feature extraction, and incorporate a contrastive learning module for image retrieval. Upon testing across three street view image datasets, the results demonstrate that our model holds certain advantages over the state‐of‐the‐art multimodal models pre‐trained on extensive datasets, characterized by fewer parameters and lower floating point operations. Code and data are available at https://github.com/nwuSY/svtRetrieval.

A multilayer network diffusion-based model for reviewer recommendation

With the rapid growth of manuscript submissions, finding eligible reviewers for every submission has become a heavy task. Recommender systems are powerful tools developed in computer science and information science to deal with this problem. However, most existing approaches resort to text mining techniques to match manuscripts with potential reviewers, which require high-quality textual information to perform well. In this paper, we propose a reviewer recommendation algorithm based on a network diffusion process on a scholar–paper multilayer network, with no requirement for textual information. The network incorporates the relationship of scholar–paper pairs, the collaboration among scholars, and the bibliographic coupling among papers. Experimental results show that our proposed algorithm outperforms other state-of-the-art recommendation methods that use graph random walk and matrix factorization and methods that use machine learning and natural language processing, with improvements of over 7.62% in recall, 5.66% in hit rate, and 47.53% in ranking score. Our work sheds light on the effectiveness of multilayer network diffusion-based methods in the reviewer recommendation problem, which will help to facilitate the peer-review process and promote information retrieval research in other practical scenes.

Report on the 46th ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2023): Reflections from the Program Co-Chairs

The ACM SIGIR Conference on Research and Development in Information Retrieval has been experiencing significant growth over the past few years. In 2023, SIGIR received a total of 822 full paper submissions and 1,787 papers/proposals. This has brought several challenges to the conference program organization. This article reports some of the main aspects of the work conducted by the program committee (PC) co-chairs for SIGIR 2023, and discusses emergent issues of the review process and program design. We expect that this document serve as a reference for future PC co-chars and as a starting point for further discussions towards improving the review process and overall conference experience. Date : 23--27 July 2023. Website : https://sigir.org/sigir2023/.

Report on the 1st Workshop on Implicit Author Characterization from Texts for Search and Retrieval (IACT 2023) at SIGIR 2023

The first edition of the International Workshop on Implicit Author Characterization from Texts for Search and Retrieval (IACT'23) was held on July 27 th , 2023, in conjunction with the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) in Taipei, Taiwan. To support both online and in-person participation, the workshop was held as a hybrid event. Online participation was allowed using the EventX platform. During the course of the day, about 15 attendees had the opportunity to follow up and discuss the recent advances in topics related to identifying and extracting implicit information about authors (e.g., human or AI) from texts and using it in IR tasks. The workshop program included two invited keynotes, four paper presentations, and the report of the ColiE competition which was held as a part of the workshop1. The proceedings of the workshop are available online at https://ceur-ws.org/Vol-3477/. Date : 27 July 2023. Website : https://en.sce.ac.il/news/iact23.

Report on the 1st Workshop on Generative Information Retrieval (Gen-IR 2023) at SIGIR 2023

The first edition of the workshop on Generative Information Retrieval (Gen-IR 2023) took place in July 2023 in a hybrid fashion, co-located with the ACM SIGIR Conference 2023 in Taipei (SIGIR 2023). The aim was to bring information retrieval researchers together around the topic of generative AI that gathered attention in 2022 and 2023 with large language models and diffusion models. Given the novelty of the topic, the workshop was focused around multi-sided discussions, namely panels and poster sessions of the accepted proceedings papers. Two main research outcomes are the proceedings of the workshop1 and the potential research directions discussed in this report. Date : 27 July 2023. Website : https://coda.io/@sigir/gen-ir.

Fintech Key-Phrase: A New Chinese Financial High-Tech Dataset Accelerating Expression-Level Information Retrieval

Expression-level information extraction is a challenging task in natural language processing (NLP), which aims to retrieve crucial semantic information from linguistic documents. However, there is a lack of up-to-date data resources for accelerating expression-level information extraction, particularly in the Chinese financial high technology field. To address this gap, we introduce Fintech Key-Phrase: a human-annotated key-phrase dataset for the Chinese financial high technology domain. This dataset comprises over 12K paragraphs along with annotated domain-specific key-phrases. We extract the publicly released reports, Chinese management’s discussion and analysis (CMD&A), from the renowned Chinese research data services platform (CNRDS) and then filter the reports related to high technology. The high technology key-phrases are annotated following pre-defined philosophy guidelines to ensure annotation quality. In order to better understand the limitations and challenges in the purposed dataset, we conducted comprehensive noise evaluation experiments for the Fintech Key-Phrase, including annotation consistency assessment and absolute annotation quality evaluation. To demonstrate the usefulness of our released Fintech Key-Phrase in retrieving valuable information in the Chinese financial high technology field, we evaluate its significance using several superior information retrieval systems as representative baselines and report corresponding performance statistics. Additionally, we further applied ChatGPT to the text augmentation approach of the Fintech Key-Phrase dataset. Extensive comparative experiments demonstrate that the augmented Fintech Key-Phrase dataset significantly improved the coverage and accuracy of extracting key phrases in the finance and high-tech domains. We believe that this dataset can facilitate scientific research and exploration in the Chinese financial high technology field. We have made the Fintech Key-Phrase dataset and the experimental code of the adopted baselines accessible on Github: https://github.com/albert-jin/Fintech-Key-Phrase. To encourage newcomers to participate in the financial high-tech domain information retrieval research, we have developed a series of tools, including an open website1and corresponding real-time information retrieval APIs.2

The effects of emotional stimuli on Word retrieval in people with aphasia

PurposePrior studies have shown that people with aphasia (PWA) have demonstrated superior language performance for emotional compared to nonemotional stimuli on a range of tasks, including auditory comprehension, verbal pragmatics, repetition, reading, and writing. However, studies on word retrieval, specifically, have suggested a possible interference effect of emotion on naming. The purpose of this study was to examine the effect of the emotional valence of stimuli on word retrieval in a series of naming tasks in PWA. MethodThirteen PWA and 13 neurotypical controls participated in four single-word naming tasks, including 1) object picture naming, 2) action picture naming, 3) category-member generation, and 4) verb generation. Each task included three valence sets of positively-, negatively-, and neutrally-rated pictures or words, which were obtained from the standardized International Affective Picture System (Lang et al., 2008) and the Affective Norms for Emotional Words (Bradley and Lang, 1999) databases. Accuracy and reaction time (RT) were measured and compared across groups, tasks, and valence sets. ResultsEmotional stimuli, especially negative stimuli, resulted in worse naming performance, as measured by accuracy and RT, compared to nonemotional stimuli in PWA and neurotypical controls. This effect was relatively robust across the four naming tasks. In most cases, negative stimuli resulted in lower accuracy and slower RT than positive stimuli. ConclusionsThese findings suggest that stimulus valence may interfere with word retrieval for PWA and neurotypical adults and that this effect is robust across different types of naming tasks that vary by word class (nouns versus verbs) and stimulus type (pictures versus words). Negative stimuli resulted in worse naming performance than positive stimuli. These results suggest that emotionality of stimuli is an important variable to consider in word retrieval research.

High-Quality and Reproducible Automatic Drum Transcription from Crowdsourced Data

Within the broad problem known as automatic music transcription, we considered the specific task of automatic drum transcription (ADT). This is a complex task that has recently shown significant advances thanks to deep learning (DL) techniques. Most notably, massive amounts of labeled data obtained from crowds of annotators have made it possible to implement large-scale supervised learning architectures for ADT. In this study, we explored the untapped potential of these new datasets by addressing three key points: First, we reviewed recent trends in DL architectures and focused on two techniques, self-attention mechanisms and tatum-synchronous convolutions. Then, to mitigate the noise and bias that are inherent in crowdsourced data, we extended the training data with additional annotations. Finally, to quantify the potential of the data, we compared many training scenarios by combining up to six different datasets, including zero-shot evaluations. Our findings revealed that crowdsourced datasets outperform previously utilized datasets, and regardless of the DL architecture employed, they are sufficient in size and quality to train accurate models. By fully exploiting this data source, our models produced high-quality drum transcriptions, achieving state-of-the-art results. Thanks to this accuracy, our work can be more successfully used by musicians (e.g., to learn new musical pieces by reading, or to convert their performances to MIDI) and researchers in music information retrieval (e.g., to retrieve information from the notes instead of audio, such as the rhythm or structure of a piece).