Sieper, Anna AbadAmarkhel, OmarDiez, SavinaPetrak, DominicBecker, Michael2021-03-092021-03-092020978-3-88579-750-0https://dl.gi.de/handle/20.500.12116/35781Software developers are often confronted with tasks for which there are widespread solution patterns. Searching for solutions using natural language queries often leads to unsatisfying results. Github, Microsoft Research and Weights & Biases created the CodeSearchNet Challenge to address this problem. Its goal is to develop code search approaches that return the code that best matches a natural language query. In this paper, we investigate two different approaches in this context. First, a Neural Bag-of-Words encoder using TF-IDF weighting and second, a Graph Convolutional Network which includes the call hierarchy in a target method’s representation. In our experiments we were able to improve the Neural Bag-of-Words models, whose results were published in the CodeSearchNet Challenge. Our Neural Bag-of-Words encoder improves the MRR by 4.38% for Python and 4.98% for Java. The Graph Convolutional Network did not improve the results over of the Neural Bag-of-Words model.enSemantic Code SearchGraph Convolutional NetworkNeural Bag-of-WordsCode-SearchNet ChallengeText/Conference Paper1614-3213