We exhibited at the 25th Annual Meeting of the Natural Language Processing Society (NLP2019)
About the whole tournamentI felt that the culture of open source publishing the programs used in research papers has become widespread, and as a result, the base of research has expanded significantly.There were many studies in which English-speaking research results were localized into Japanese and applied and tuned to the domains that each laboratory is good at.
About research presentation"P7-9 Consistent job article classification by multi-task learning and generation of job phrase and article headline" From the text input of job information, we learn three types of tasks, classifying job types, generating job type phrases, and generating catchphrases, using the same model to improve the consistency of each output.Improving consistency between multiple outputs is an important concept, and it is also related to ensuring the explanation of artificial intelligence, such as multi-task learning of "recommendation + generation of recommendation reason".In commercial deep learning, it is expected that there will be more and more situations where multitask learning is considered from a non-accuracy aspect. "D1-4 Correspondence between automatically constructed Japanese frame and English FrameNet" "F2-3 Simultaneous learning of ambiguity resolution and dependency labeling" We are trying to improve the data for machine learning in Japanese by successfully linking English-speaking language resources and Japanese language resources.As the latest models become easier to use and the importance of data increases, I would like to pay attention to basic research that will lead to the qualitative and quantitative enhancement of the Japanese corpus. "Large-scale semi-supervised learning with a mixed network of E5-1 Expert and Imitator" Recently, the number of cases where deep learning is used for natural language is increasing, but it is often not possible to prepare a large amount of labeled data in natural language.Therefore, a method called semi-supervised learning is drawing attention.Semi-supervised learning is a method of learning by automatically tagging a small amount of labeled data to something with high confidence in unlabeled data.In this research, we focused on unlabeled data that had not been focused on until now, and aimed to improve the generalization performance of machine learning devices by increasing the amount of unlabeled data.If unlabeled data improves the generalization performance of machine learning devices, machine learning will be easier to apply even in fields where it is difficult to prepare labeled data such as texts in the future. FRONTEO's KIBIT has already been learning efficiently with a small amount of training data, but I felt that incorporating these technologies would greatly expand the opportunities for us to play an active role. At the FRONTEO booth, we exchanged information on technologies and initiatives that we were interested in with researchers, students, and companies participating in the conference, and also gave us opinions on how important natural language processing technology would spread to the world. We exchanged and deepened exchanges.In addition to delving into the research field, I heard words such as internship and job hunting in a seasonal manner, indicating that there is a high level of interest in research and implementation opportunities that actually use data owned by companies. FRONTEO will continue to support the Society and carry out activities with the aim of further developing natural language processing and utilizing it in society. [speech_bubble type = ”std” subtype = ”L1 ″ icon =” kibiro.jpg ”name =” ”] Thank you to everyone who came to the booth. [/ speech_bubble]
Articles in the same category
FRONTEO CEO Morimoto gave a lecture at Sophia University's university-wide common course "Tsukuru II (Career Development II)"
[Webinar held on 10/26] Email monitoring using AI-Discovering risks from daily communication-
[10/19 webinar] Don't miss the signs of an accident! -Anyone, easy, and on-site AI x danger prediction-
[Webinar held in October] Basics of thinking about antitrust law violation risk that can arise due to new social conditions
[Upcoming webinar in October] Restarted crackdown! Get ready for government enforcement of Antitrust laws and the FCPA under the Biden administration
[Webinar held on 10/12] The latest method to inherit the "craftsmanship" that is lost at the manufacturing site
[Held on October 10th!13th Economic Security Study Group] Geopolitical Risk — Risk Management Upstream / Downstream of Supply Chain
[Webinar held on 9/28] Email monitoring using AI-Discovering risks from daily communication-
[September Webinar] AI and Technology Assisted Review: Best Practices & Cutting-Edge Issues in Litigations and Investigations
[Webinar held on 9/21] Data analysis by AI that learned the "judgment criteria" of experts-DX cases of compliance, personnel, and sales-