In Seavus, knowledge is one of the core company values. That is why we started a cooperation with the Royal Institute of Technology (KTH) in Stockholm, where we had a chance to work with their students. The cooperation was fruitful as some of the students were working on their master thesis at Seavus Stockholm AB, under our mentorship and support.
Over the past three years, we have conducted 8 theses with a total of 10 students. Their research area mostly revolved around Machine Learning models.
After finishing the thesis, some of the students were hired in Seavus, while others got jobs in other companies.
For more information, contact us on:
1) News and social media post analysis using NLP techniques
Seavus developed a Media Intelligent Platform to collect and analyze news articles and social media posts related to search terms. This helps customers to analyze the success of media campaigns, determine the popularity of a product, follow trends, etc. The platform core uses different Machine Learning and NLP techniques for the analysis. This project aims to explore, develop and evaluate techniques to improve text classification, grouping and sentiment analysis. The focus will be on matching posts to a search term using text classification and topic modeling. Working together with our AI team, the student will contribute to the product evolution and benefit from the supervision and help of domain experts. Furthermore, he/she will learn how business projects are planned, conducted and which issues can be faced.
2) Data augmentation in Natural Language Processing
Widely used in Computer Vision, data augmentation is relatively unexplored in the natural language domain, which is transforming as quick as never before. This project aims to explore, develop and evaluate techniques for data augmentation in NLP. By sitting next to our AI team, the student will benefit both from domain experts supervision and help, as well as he/she will learn how real projects are conducted and which issues are faced. Depending on the outcome, the solution will be used in the most suitable projects ongoing during next spring.
3) Open Thesis
The AI team works mostly on projects realted to Natural Language Processing and tabular data. However, we love learning always something new, exploring other AI domains and being open to innovaative ideas. Therefore, we encourgage students to propose topics they would like to investigate in their Thesis work.
1) Evaluate and suggest an approach for focused summarization of meeting transcript (code)
Previous researches have shown that unsupervised methods are not achieving business performance in the field of automatic meeting summarization. Also, a generic summary is usually not the best outcome for most of the users. Hence, the scope of this work was to evaluate and propose a method for focused summarization of meeting transcripts, i.e. identifying and summarizing only some specific parts as decisions, action items, problems and so on. Performance has been evaluated both in terms of scientific relevance and business applicability.
2) Generative Adversarial Networks in Natural Language Processing (code)
GANs have shown their capability to generate photorealistic images and to reconstruct 3D models of objects from images. Recently, they have also been applied to generate Natural Language.
The scope of this thesis was to investigate the current GANs state-of-the-art and their applicability to NLP tasks and evaluate whether GANs could be a competitor of currently used methods.
The main purpose of the thesis was to investigate and develop a text generation GAN, and a further application is to use GAN to generate abstractive summaries according to meeting records. The research area of the thesis is in Generative Adversarial Network, Natural Language Processing (NLP) and Reinforcement Learning. The expected outcome should be a GAN that can generate high-quality text and a GAN that can generate an informative and fluent abstractive summary of meeting records.
3) Interface Design for Explainable Deep Models in NLP (code)
There are many different approaches to make people understand how deep models “think”, like generating inference of how a deep model makes its decision or visualizing the contribution made by each neuron to the prediction. We wanted to know how to design the explanation interface to make the explanations appropriate for users.
The scope of this thesis was to investigate proper approaches to integrate explanations generated from explainable deep learning methods into the AI system, and design prototype and explanation interface for both normal users and users with a technology background.
1) A model for a screening tool that will increase understanding of how AI can be applied for increased innovation force
A model for how to conduct a productive dialogue around AI with companies based on AI maturity to provide a high-level framework for identifying the extent to which current technologies, skills, and processes enable the delivery of AI capabilities in different areas. A model that translate AI initiatives in real-world contexts.
The purpose of the thesis was to increase understanding of how different companies work with AI in their product and service development process today and create the prerequisites for a screening tool that will increase understanding of how AI can be applied for increased innovation force. A part of the thesis was to study some companies that have developed or want to develop AI-based offers.
(This thesis was a part of a research study conducted by KTH)
2) Evaluate and suggest an approach for automatic summarization of meeting transcript (code)
Automatic summarization is divided in two main branches: extractive and abstractive. The scope of the thesis is to evaluate methods for generating summaries from meeting transcripts, assuming no-prior knowledge about the meeting content is available. Due to lack of domain specific open source code, the community benefits also from it being released. The work produced new features which improved the extractive summarization performances, as well as conclusions driving future researches.
3) Explainable Deep Learning for Natural Language Process (code)
Deep learning methods get impressive performance in many tasks from Natural Neural Processing (NLP) ﬁeld, but it is still difﬁcult to know what happened inside the neural network. How to monitor that a deep learning model makes the right decision when huge numbers of parameters, many rounds of iteration and optimization from deep learning models are taken into consideration.
The scope of this was to give a general overview of Explainable AI and with focus on explainable deep learning methods to see how it can be used in NLP tasks.
1) Investigate how to implement new features and build a modern conference system
The aim of this study was to investigate how to implement new features and build a modern conference system, using modern technologies and new applications to improve the documentation of meetings and conferences. Speech to text in combination with speaker identification can facilitate documenting meetings and conferences.
2) Performance Evaluation of Bluetooth and Wi-Fi for a smart conference hub
In environments where there is communication between devices, the communication is of high importance. This is so that the interaction between the user and the device must occur without interruptions and delays.
In an environment where there are multiple devices and the user who wants to access them from the same source, it may be problematic to control them. This issue can also occur when a user wants to quickly connect between devices, especially older devices that require a wired connection.
Solution to such problems can be to connect these devices to a common controller, a so-called smart hub. By wirelessly controlling the hub, it is possible to access devices from a single source via an interface that the hub presents to the user.
The scope of the thesis was to determine the appropriate protocol for wireless communication between the central control unit and the user. The choice of communication protocols depends on several factors such as data rate, packet loss and security. To tests and evaluates the wireless communication protocols Bluetooth and Wi-Fi to determine which of these is most suitable for future prototype development of a smart hub, the smart hub can then be used to simplify the management of multiple devices at the same time.