From Statistical Term Extraction to Hybrid Machine Translation

Tapta: A user-driven translation system for patent documents based on domain-aware Statistical Machine Translation Bruno Pouliquen Christophe Mazenc Aldo Iorio World Intellectual Property Organization 34, chemin des Colombettes CH-1211 Geneva 20 Bruno.Pouliquen@wipo.int, Christophe.Mazenc@wipo.int, Aldo.Iorio@wipo.int Abstract This paper presents a study conducted in the course of implementing a project in the World Intellectual Property Organiza- tion (WIPO) on assisted translation of pat- ent abstracts and titles from English to French. The tool (called ‘Tapta’) is trained on an extensive corpus of manually trans- lated patents. These patents are classified, each class belonging to one of the 32 pre- defined domains. The trained Statistical Machine Translation (SMT) tool uses this additional information to propose more accurate translations according to the con- text. The performance of the SMT system was shown to be above the current state of the art, but, in order to produce an accept- able translation, a human has to supervise the process. Therefore, a graphical user in- terface was built in which the translator drives the automatic translation process. A significant experiment with human opera- tors was conducted within WIPO, the out- put was judged to be successful and a pro- ject to use Tapta in production is now un- der discussion. 1 Introduction This paper describes a tool called Tapta1 aimed at facilitating the work of translators in the World Intellectual Property Organization (WIPO) which is a specialized agency of the United Nations2. One of the activities of WIPO is to administer the Patent Cooperation Treaty (PCT) which offers patent applicants an advantageous route for ob- taining patent protection in multiple countries. For 1 Translation Assistant for Patent Titles and Abstracts 2 “WIPO is dedicated to developing a balanced and accessible international intellectual property (IP) system, which rewards creativity, stimulates innovation and contributes to economic development while safeguarding the public interest” http://www.wipo.int/about-wipo publication purposes, WIPO translates the title and the abstract of each received patent document (called a PCT application) from its filing language into both the English and French languages. The number of published documents per year has been steadily increasing in the last two decades (from 16,000 in 1990, 80,000 in 2000 to 150,000 in 2010). Originally only performed by WIPO trans- lators, a large share of these translations (more than 80%) is now outsourced to private companies and we are looking at the latest achievements in Computer Aided Translation (CAT) tools and Machine Translation (MT) to optimize costs and improve quality. High quality of translation is re- quired, WIPO conducts a quality check. PCT sta- tistics are published in WIPO (2010). We have an extensive parallel corpus of manu- ally translated patent documents collected over time, especially for the language pair English- French (1.8 million documents). A test was launched: a statistical machine translation (SMT) system was trained using our corpus and a tool to help translation of patent titles and abstract from English to French provided. WIPO’s expectations in term of quality of translation are extremely high, therefore using solely the results of an …

Content truncated to 500 words