Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code.

/pdf/roberta-a-robustly-optimized-bert-pretraining-approach-2rj0bdyhim.pdf

RoBERTa: A Robustly Optimized BERT Pretraining Approach

行銷硏究 : Marketing research

A systematic literature review of blockchain-based applications: Current status, classification and open issues

In the last 16 years, more than 200 research articles were published about research-paper recommender systems. We reviewed these articles and present some descriptive statistics in this paper, as well as a discussion about the major advancements and shortcomings and an overview of the most common recommendation concepts and approaches. We found that more than half of the recommendation approaches applied content-based filtering (55 %). Collaborative filtering was applied by only 18 % of the reviewed approaches, and graph-based recommendations by 16 %. Other recommendation concepts included stereotyping, item-centric recommendations, and hybrid recommendations. The content-based filtering approaches mainly utilized papers that the users had authored, tagged, browsed, or downloaded. TF-IDF was the most frequently applied weighting scheme. In addition to simple terms, n-grams, topics, and citations were utilized to model users' information needs. Our review revealed some shortcomings of the current research. First, it remains unclear which recommendation concepts and approaches are the most promising. For instance, researchers reported different results on the performance of content-based and collaborative filtering. Sometimes content-based filtering performed better than collaborative filtering and sometimes it performed worse. We identified three potential reasons for the ambiguity of the results. (A) Several evaluations had limitations. They were based on strongly pruned datasets, few participants in user studies, or did not use appropriate baselines. (B) Some authors provided little information about their algorithms, which makes it difficult to re-implement the approaches. Consequently, researchers use different implementations of the same recommendations approaches, which might lead to variations in the results. (C) We speculated that minor variations in datasets, algorithms, or user populations inevitably lead to strong variations in the performance of the approaches. Hence, finding the most promising approaches is a challenge. As a second limitation, we noted that many authors neglected to take into account factors other than accuracy, for example overall user satisfaction. In addition, most approaches (81 %) neglected the user-modeling process and did not infer information automatically but let users provide keywords, text snippets, or a single paper as input. Information on runtime was provided for 10 % of the approaches. Finally, few research papers had an impact on research-paper recommender systems in practice. We also identified a lack of authority and long-term research interest in the field: 73 % of the authors published no more than one paper on research-paper recommender systems, and there was little cooperation among different co-author groups. We concluded that several actions could improve the research landscape: developing a common evaluation framework, agreement on the information to include in research papers, a stronger focus on non-accuracy aspects and user modeling, a platform for researchers to exchange information, and an open-source framework that bundles the available recommendation approaches.

/pdf/research-paper-recommender-systems-a-literature-survey-3o8izjcp9m.pdf

Research-paper recommender systems: a literature survey

While blockchain technology is commonly considered potentially disruptive in various regards, there is a lack of understanding where and how blockchain technology is effectively applicable and where it has mentionable practical effects. This issue has given rise to critical voices that judge the technology as over-hyped. Against this backdrop, this study adapts an established research framework to structure the insights of the current body of research on blockchain technology, outline the present research scope as well as disregarded topics, and sketch out multidisciplinary research approaches. The framework differentiates three groups of activities (design and features, measurement and value, management and organization) at four levels of analysis (users and society, intermediaries, platforms, firms and industry). The review shows that research has predominantly focused on technological questions of design and features, while neglecting application, value creation, and governance. In order to foster substantial blockchain research that addresses meaningful questions, this study identifies several avenues for future studies. Given the breadth of open questions, it shows where research can benefit from multidisciplinary collaborations and presents data sources as starting points for empirical investigations.

/pdf/a-blockchain-research-framework-4e30jfchjz.pdf

A Blockchain Research Framework

Over 80 approaches for academic literature recommendation exist today. The approaches were introduced and evaluated in more than 170 research articles, as well as patents, presentations and blogs. We reviewed these approaches and found most evaluations to contain major shortcomings. Of the approaches proposed, 21% were not evaluated. Among the evaluated approaches, 19% were not evaluated against a baseline. Of the user studies performed, 60% had 15 or fewer participants or did not report on the number of participants. Information on runtime and coverage was rarely provided. Due to these and several other shortcomings described in this paper, we conclude that it is currently not possible to determine which recommendation approaches for academic literature are the most promising. However, there is little value in the existence of more than 80 approaches if the best performing approaches are unknown.

/pdf/research-paper-recommender-system-evaluation-a-quantitative-3sktadw0ze.pdf

Research paper recommender system evaluation: a quantitative literature survey

The amount of news published and read online has increased tremendously in recent years, making news data an interesting resource for many research disciplines, such as the social sciences and linguistics. However, large scale collection of news data is cumbersome due to a lack of generic tools for crawling and extracting such data. We present news-please, a generic, multi-language, open-source crawler and extractor for news that works out-of-the-box for a large variety of news websites. Our system allows crawling arbitrary news websites and extracting the major elements of news articles on those websites, i.e., title, lead paragraph, main content, publication date, author, and main image. Compared to existing tools, news-please features full website extraction requiring only the root URL.

news-please : a Generic News Crawler and Extractor

Numerous recommendation approaches are in use today. However, comparing their effectiveness is a challenging task because evaluation results are rarely reproducible. In this article, we examine the challenge of reproducibility in recommender-system research. We conduct experiments using Plista's news recommender system, and Docear's research-paper recommender system. The experiments show that there are large discrepancies in the effectiveness of identical recommendation approaches in only slightly different scenarios, as well as large discrepancies for slightly different approaches in identical scenarios. For example, in one news-recommendation scenario, the performance of a content-based filtering approach was twice as high as the second-best approach, while in another scenario the same content-based filtering approach was the worst performing approach. We found several determinants that may contribute to the large discrepancies observed in recommendation effectiveness. Determinants we examined include user characteristics (gender and age), datasets, weighting schemes, the time at which recommendations were shown, and user-model size. Some of the determinants have interdependencies. For instance, the optimal size of an algorithms' user model depended on users' age. Since minor variations in approaches and scenarios can lead to significant changes in a recommendation approach's performance, ensuring reproducibility of experimental results is difficult. We discuss these findings and conclude that to ensure reproducibility, the recommender-system community needs to (1) survey other research fields and learn from them, (2) find a common understanding of reproducibility, (3) identify and understand the determinants that affect reproducibility, (4) conduct more comprehensive experiments, (5) modernize publication practices, (6) foster the development and use of recommendation frameworks, and (7) establish best-practice guidelines for recommender-systems research.

/pdf/towards-reproducibility-in-recommender-systems-research-21yrirjx6i.pdf

Towards reproducibility in recommender-systems research

This paper evaluates the performance of tools for the extraction of metadata from scientific articles. Accurate metadata extraction is an important task for automating the management of digital libraries. This comparative study is a guide for developers looking to integrate the most suitable and effective metadata extraction tool into their software. We shed light on the strengths and weaknesses of seven tools in common use. In our evaluation using papers from the arXiv collection, GROBID delivered the best results, followed by Mendeley Desktop. SciPlore Xtract, PDFMeat, and SVMHeaderParse also delivered good results depending on the metadata type to be extracted.

/pdf/evaluation-of-header-metadata-extraction-approaches-and-1bggn550fb.pdf

Corinna Breitinger

Papers

Research-paper recommender systems: a literature survey

Research paper recommender system evaluation: a quantitative literature survey

news-please : a Generic News Crawler and Extractor

Towards reproducibility in recommender-systems research

Evaluation of header metadata extraction approaches and tools for scientific PDF documents