Thesis work: Graph neural networks

Graph neural networks on dependency parse trees of text
Thesis project: 30 credits, starting in January 2021

Would you like to extract knowledge from text using the latest and greatest techniques from deep learning and NLP?

At AstraZeneca, we turn ideas into life changing medicines and strive to continuously meet the unmet needs of patients worldwide. Working here means being entrepreneurial, thinking big and working together to make the impossible a reality. If you are swift to action, confident to lead, willing to collaborate, and curious about what science can do, then you’re our kind of person.

In R&D IT and specifically in the AI Engineering Lab, we create scalable AI capabilities for our research colleagues, with the constant aim of delivering better science, faster.

In relationship extraction we are looking to deduce if there exists a relationship between two given entities in a text. We would like the model to find these signals based on syntactic properties of the text, rather than knowledge of the entities in question to avoid overfitting and "cheating". Transformer models, while they are very powerful, like to cheat and find spurious words or patterns in the data set that correlate with the label, usually exploiting bias in the data selection process or annotation process. While these models do capture syntactic structure, evidence suggest that they ignore this when trained on a downstream task.

A way to force a model to focus on syntax is to do machine learning on the dependency tree of the text. Given a sentence, we would extract the dependency tree with an off-the-shelf model, and use this graph structure as the basis of our model. A graph convolution neural network shares many properties with transformer models, and we expect that working on the tree directly will force the model to take syntax into account to a much higher degree than a plain transformer model. We would evaluate our approach on our current evaluation sets, to see if the desired properties have been established.

Randstad Life Sciences is cooperating with AstraZeneca in this recruitment process. We only accept applications through Randstad’s website.

Deadline for application: 2020-11-05, selection and interviews will be ongoing. The position may be filled before the last day of application, therefore, apply as soon as possible.

For more information: Kerstin Karlsson kerstin.karlsson@randstad.se or Linnea Öster linnea.oster@randstad.se

Our Gothenburg site is one of AstraZeneca's three strategic science centers. We thrive in a multinational environment working cross-functionally across the globe with AstraZeneca colleagues as well as academic and industry partners. Our way of life is to foster a working environment that nurtures, collaboration, openness and innovation. Therefore, we have created space for meetings, socializing and relaxation, where spontaneous meetings can give birth to new innovations. The unexpected ideas or thoughts that can come from a chat over something as simple as a cup of coffee or a stroll on our “walk and talk” meeting trail.

AstraZeneca is an equal opportunity employer. AstraZeneca will consider all qualified applicants for employment without discrimination on grounds of disability, sex or sexual orientation, pregnancy or maternity leave status, race or national or ethnic origin, age, religion or belief, gender identity or re-assignment, marital or civil partnership status, protected veteran status (if applicable) or any other characteristic protected by law. AstraZeneca only employs individuals with the right to work in the country/ies where the role is advertised.