Designing Chemical Feature Analysis for Chemists

Background

Scientists use recommender systems to quickly find the most relevant scientific literature in their field. However, in the chemistry domain, chemists are not only interested in textual relevance, but also want to find and make sense of chemical formulas contained in the literature. For example, within a publication: C9H8O4, C9H8O4, acetylsalicylic acid, and ASA all refer to the same compound: Aspirin! How can we design an interface that highlights such information and lets chemists quickly get an overview of all chemical compound names, chemical formulas, and drug names discussed in a publication?

Goal

• Improve the ease of finding, comparing and understanding chemical named entities (e.g. methanol, acetone, hexane, H20, MG(OH)3, CH2Cl2) in scientific literature.

Tasks

• Research and compare existing toolkits for the automated extraction of chemical information from scientific literature, e.g., ChemDataExtractor (python)
• Build a prototype that visualizes the chemical information present in scientific
  papers to help chemist more quickly find, compare, and make sense of the chemical information present.

Advisor

Corinna Breitinger

▸Slides

AI Solving

Background

The extent of mathematics literature is estimated to be more than 120 million pages. However, the majority of them are not available in a machine-processable format. Hence, valuable mathematical content cannot be used by today’s technologies for developing models capable of doing math-related tasks. There are MathOCR tools such as MathPix, InftyReader, etc. However, they are commercial and not evaluated for their performance. In this project, you will work on an OCR tool capable of converting math equations from PDF or image to machine-processable forms such as LaTeX or MathML. You will work on neural networks capable of detecting math symbols and the structure of math formulae. You can choose either theory or applied track for this project.

Goal

• Develop a MathOCR.

Tasks

• Textual and non-textual elements detection from a PDF or an Image input.
• Math symbol identification.
• Formula structure recognition using computationally inexpensive methods.
• Evaluate MathOCR methods on mathematical formulae of varied complexity.

Advisor

Ankit Satpute

▸Slides

AI Solving

Background

With the digital era and the need to communicate complex information quickly and concise, emojis are used to abbreviate otherwise lengthy textual descriptions. Due to higher resolutions, emojis evolved into rich and detailed graphics for a large variety of use cases. The sheer number of existing graphics motivated emoji riddles, puzzles that try to reverse engineer the intended textual meaning of a set of graphics. Due to the ambiguities of simplified graphics, these emoji riddles often require a certain degree of creativity to solve. Meanwhile, all the information necessary to solve such puzzles are generally accessible to modern algorithms. Hence the pressing question, can an AI be creative and solve emoji riddles?

Goal

• Develop a model that converts a set of emoji graphics into textual descriptions.

Tasks

• Study which models, datasets, metrics can be used in this task.
• Develop a concept model that trains connections between graphics and textual descriptions.
• Train and evaluate the model.

Advisor

Dr. André Greiner-Petter

▸Slides