A Trainable Tokenizer, solution for multilingual texts and compound expression tokenization

AbstractTokenization is one of the initial steps done for almost any text processing task. It is not particularly recognized as a challenging task for English monolingual systems but it rapidly increases in complexity for systems that apply it for different languages. This article proposes a supervised learning approach to perform the tokenization task. The method presented in this article is based on character transitions representation, a representation that allows compound expressions to be recognized as a single token. Compound tokens are identified independent of the character that creates the expression. The method automatically learns tokenization rules from a pre-tokenized corpus. The results obtained using the trainable system show that for Romanian and English a statistical significant improvement is obtained over a baseline system that tokenizes texts on every non-alphanumeric character.