Scribd Logo
Upload

Welcome to Scribd!

  • 31 views

    Conclusion and Future Work

    Uploaded by

    ANANTH UPADHYA
    The document discusses using deep learning techniques to identify sections of scientific publications such as the title, author, abstract, and body text by analyzing PDF documents. The system was able to reliably identify body text and reject other sections with 94.32% accuracy. Future work involves improving accuracy, identifying specific section types, and using the output to extract text from PDFs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Conclusion and Future Work

    Uploaded by

    ANANTH UPADHYA
    31 views1 page
    The document discusses using deep learning techniques to identify sections of scientific publications such as the title, author, abstract, and body text by analyzing PDF documents. The system was able to reliably identify body text and reject other sections with 94.32% accuracy. Future work involves improving accuracy, identifying specific section types, and using the output to extract text from PDFs.

    Original Description:

    A Deep learning approach in analyzing pdfs

    Original Title

    1460210-4

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    The document discusses using deep learning techniques to identify sections of scientific publications such as the title, author, abstract, and body text by analyzing PDF documents. The system was able to reliably identify body text and reject other sections with 94.32% accuracy. Future work involves improving accuracy, identifying specific section types, and using the output to extract text from PDFs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    31 views1 page

    Conclusion and Future Work

    Uploaded by

    ANANTH UPADHYA
    The document discusses using deep learning techniques to identify sections of scientific publications such as the title, author, abstract, and body text by analyzing PDF documents. The system was able to reliably identify body text and reject other sections with 94.32% accuracy. Future work involves improving accuracy, identifying specific section types, and using the output to extract text from PDFs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 1

    ples of an input provided to the network, examples of net- Lafferty, J. D., McCallum, A., and Pereira, F. C. N. (2001).

    work output, and ground truth target output. These results Conditional random fields: Probabilistic models for seg-
    demonstrate impressive results with such a small dataset. menting and labeling sequence data. In ICML ’01 Pro-
    In particular, the network is able to reject header and footer ceedings of the Eighteenth International Conference on
    text extremely reliably. The network rejects most abstracts, Machine Learning, pages 282–289.
    figure captions and references, confusing only some where LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-
    the text formatting is extremely similar to typical paragraph ing. nature, 521(7553):436.
    text. The per pixel classification accuracy on the validation
    Lipinski, M., Yao, K., Breitinger, C., Beel, J., and Gipp,
    set was 94.32%, compared to a baseline of classifying each
    B. (2013). Evaluation of header metadata extraction ap-
    pixel as “not paragraph” which would provide 79.67% ac-
    proaches and tools for scientific PDF documents. In Pro-
    curacy.
    ceedings of the 13th ACM/IEEE-CS joint conference on
    Digital libraries - JCDL ’13. ACM Press.
    6. Conclusion and Future Work
    Lopez, P. (2009). GROBID: Combining Automatic Bib-
    In this paper we demonstrated that deep learning-based im- liographic Data Recognition and Term Extraction for
    age analysis can be used to identify sections of scientific Scholarship Publications. In International Conference
    publications. Given the results from our current experi- on Theory and Practice of Digital Libraries, pages 473–
    ments, we feel that deep learning can be successfully used 474. Springer.
    to enhance current PDF extraction methods, and based on
    Mao, S., Rosenfeld, A., and Kanungo, T. (2003). Docu-
    our findings we plan to continue collecting data in order
    ment structure analysis algorithms: a literature survey.
    to further increase our networks results, as we feel many
    In Tapas Kanungo, et al., editors, Document Recognition
    of the misclassified portions of text are due to insufficient
    and Retrieval X. SPIE, jan.
    training data that does not currently characterize features
    such as reference sections and abstracts sufficiently. Peng, F. and McCallum, A. (2004). Accurate informa-
    Our current results show that a deep learning network can tion extraction from research papers using conditional
    successfully distinguish and learn the difference between random fields. In HLT-NAACL 2004: Human Language
    the body text and other portions of a PDF document. The Technology Conference of the North America Chapter of
    next step is to extend the approach to identifying each type the Association for Computational Linguistics, Proceed-
    of text (title, author, abstract, body text, etc.) rather than ings of the Main Conference, pages 329–336.
    simply body text versus other. Additionally, we plan to in- Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:
    crease the accuracy of our network by adding more data Convolutional networks for biomedical image segmen-
    and to create an extraction tool that leverages the output tation. In International Conference on Medical image
    of the deep learning network to extract text. While we are computing and computer-assisted intervention, pages
    currently evaluating accuracy based on a per pixel count of 234–241. Springer.
    estimated versus redacted image, an improved test of accu- Siegel, N., Lourie, N., Power, R., and Ammar, W. (2018).
    racy would be to leverage such an extraction tool to identify Extracting Scientific Figures with Distantly Supervised
    the per character accuracy of this text extraction approach. Neural Networks. In To appear in ACM/IEEE Joint
    Conference on Digital Libraries in 2018 (JCDL 2018).
    7. Bibliographical References ACM/IEEE.
    Beel, J., Langer, S., Genzmehr, M., and Müller, C. (2013). Singh, M., Barua, B., Palod, P., Garg, M., Satapathy, S.,
    Docear’s PDF Inspector: Title Extraction from PDF Bushi, S., Ayush, K., Rohith, K. S., Gamidi, T., Goyal, P.,
    Files. In Proceedings of the 13th ACM/IEEE-CS joint and Mukherjee, A. (2016). OCR++: A Robust Frame-
    conference on Digital libraries, pages 443–444. ACM. work For Information Extraction from Scholarly Arti-
    Constantin, A., Pettifer, S., and Voronkov, A. (2013). Pdfx: cles. International Conference on Computational Lin-
    fully-automated pdf-to-xml conversion of scientific liter- guistics (COLING), pages 3390–3400.
    ature. In Proceedings of the 2013 ACM symposium on Tkaczyk, D., Szostek, P., Dendek, P. J., Fedoryszak, M.,
    Document engineering, pages 177–180. ACM. and Bolikowski, L. (2014). Cermine – Automatic Ex-
    Councill, I. G., Giles, C. L., and Kan, M.-Y. (2008). traction of Metadata and References from Scientific Lit-
    ParsCit: an Open-source CRF Reference String Parsing erature. In Document Analysis Systems (DAS), 2014 11th
    Package. In LREC, volume 8, pages 661–667. IAPR International Workshop on, pages 217–221. IEEE.

    You might also like