Home   ||   Project   ||   Publication   ||   Miscellaneous

Speech Translation

This project is to build a real-time closed-captioning system for broadcast news speech to help people get latest world news immediately. This system consists of ASR (Automatic Speech Recognition) and MT (Machine Translation) sub-systems. I am taking charge of deep neural network-based language model and translation model parts. Our ASR sub-system, transcribing TED speeches to texts took part in the IWSLT TED share task and achieved best performance in 2012 and 2013.

Related Entity Finding

This project is to investigate the problem of related entity finding. Given the name and homepage of an entity, as well as the type of the target entity, the system is required to find related entities that are of target type and homepages of the related entities. Our system took part in TREC 2009 and 2010 and obtained good performance.

Multilingual Question Answering

Given a natural language question, the system returns the exact answers extracted from the given document collection. Currently, our systems can answer FACTOID English and Chinese questions (What is the 3-character airport code for Dulles Airport?), complex English and Chinese questions (What are the hazards of global warming?). The English-Chinese cross-lingual complex QA system was the best among all English-Chinese participants’ systems of the NTCIR-2008.

Named Entity Recognition

The target was to recognize product names, person, location, organization, time, and number expressions from Chinese text. Our system took part in 863 Evaluation Conferences (domestic conference in China) in 2003 and 2004, and performances on person, location, and organization recognition were the best among all participants’ systems.

(click to see large picture)

Speech Document Retrieval

This project was funded by Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan. Speech documents with lower speech recognition (SR) is quite challenging in this task. Therefore, using the intermediate results of SR, i.e., lattice, is our solution. The proposed algorithm achieved satisfactory performances in terms of Japanese speech collection.

Kyoto Sightseeing Dialogue System

The spoken dialogue system, is being used in Kyoto’s tourist information center, is a robot that understands what we speak and gives pertinent replies. Asking the machine by “Tell me about the Golden Pavilion,” it would, for example, give you some information gathered from Wikipedia. If you asked, “How do I get there,” it would answer “You can go there by bus” and display the bus timetable.