Christoph Tillmann

Dr. Christoph Tillmann is a Senior Research Staff Member in the AI Models Science department at IBM Research - Yorktown Heights, and a member of the team that releases the IBM Granite models. For most of his career, he has worked on statistical machine translation (SMT) and multilingual natural language processing (NLP).

Dr. Tillmann received a Diplom (master's-level degree) in Computer Science and a Ph.D. from RWTH Aachen University, Germany. During his doctoral studies and in the following years, his research focused on search algorithms for SMT, in particular dynamic programming (DP) based algorithms. These algorithms model translation as a sequential, mostly monotonic process that generates the target sentence from left to right, with restricted word re-ordering. This sequential view of translation came many years before the current sequence-based models.

He then proposed a sequential, block-based model for phrase-based SMT, in which translation proceeds as a sequence of phrase pairs ("blocks") with limited re-ordering between neighboring blocks. One version of the model was trained as a multilingual language model over block sequences. Another version was trained fully discriminatively at the sentence level, which resulted in the first successful experiment with a linear model that relies entirely on binary feature functions. This model was trained end-to-end, similar to how current neural network-based models are trained.

In addition, Dr. Tillmann has worked on beam search algorithms for extracting parallel sentence pairs from comparable data. These algorithms score candidate pairs directly at the sentence level, and the search handles a trillion candidate pairs or more. This approach anticipated more recent methods that use embeddings to find parallel sentence pairs.

More recently, he worked on the detection of hate, abuse, and profanity (HAP) to remove harmful content from text. His current work includes supervised fine-tuning (SFT), reinforcement learning (RL), and synthetic data generation. Dr. Tillmann has also been a frequent program committee member for major NLP conferences.

Personal

• He is a native speaker of German. He has excellent reading proficiency in Russian, English, and French, and excellent speaking proficiency in English and Russian. His Russian is close to native speaker level (some days).
• In 1986, Dr. Tillmann won third place in a nationwide Russian language studies competition for high-school students, Stuttgart, Germany.
• In 1991-1992, he was an exchange student at Leningradskii Gosudarstvennyi Universitet (LGU), St. Petersburg, Russian Federation, with a grant from the German Academic Exchange Service (DAAD).
• During his studies in Computer Science from 1991 to 1995, he received a scholarship from the Cusanuswerk, the Catholic Church's academic scholarship foundation in Germany.
• He likes reading, in particular French-language and Russian-language works of fiction (mostly). He also reads history, philosophy, and biographies. He enjoys photography and solving chess puzzles. In the last few years, he has started running 5K and 10K races.