Tables are a popular and efficient means of presenting structured\ninformation. They are used extensively in various kinds of documents including\nweb pages. Tables display information as a two-dimensional matrix, the\nsemantics of which is conveyed by a mixture of structure (rows, columns),\nheaders, caption, and content. Recent research has started to consider tables\nas first class objects, not just as an addendum to texts, yielding interesting\nresults for problems like table matching, table completion, or value\nimputation. All of these problems inherently rely on an accurate measure for\nthe semantic similarity of two tables. We present TabSim, a novel method to\ncompute table similarity scores using deep neural networks. Conceptually,\nTabSim represents a table as a learned concatenation of embeddings of its\ncaption, its content, and its structure. Given two tables in this\nrepresentation, a Siamese neural network is trained to compute a score\ncorrelating with the tables' semantic similarity. To train and evaluate our\nmethod, we created a gold standard corpus consisting of 1500 table pairs\nextracted from biomedical articles and manually scored regarding their degree\nof similarity, and adopted two other corpora originally developed for a\ndifferent yet similar task. Our evaluation shows that TabSim outperforms other\ntable similarity measures on average by app. 7% pp F1-score in a binary\nsimilarity classification setting and by app. 1.5% pp in a ranking scenario.\n