Deep Exploration and Filtering of Text (DEFT) Denys Katerenchuk Speech Lab, Queens College, CUNY
DEFT DARPA, Department of Defense Excess of data Important data may be implicit Efficiency Goal: Develop automated deep NLP technology for efficient information processing and understanding
DEFT “Overwhelmed by deadlines and the sheer volume of available foreign intelligence, analysts may miss crucial links, especially when meaning is deliberately concealed or otherwise obfuscated” Bonnie Dorr, DARPA program manager
Overview Introduction Innovation Algorithms and Tools Results Future work
Introduction NLP has a range of tools to discover target data Information Retrieval Automated Speech Recognition Machine Translation Etc. Robust combination is needed!
Information-Rich Relational Analysis for Spoken Data Innovation Information-Rich Relational Analysis for Spoken Data
Hypothesis Combine rich speech representation and annotations with prosodic features to improve F- measure score
Motivation Speech is different from text Speech may contain disfluency and depends on ASR performance Named Entities are often OOV Prosodic information is ignored
ASR Hypothesis Representation Lattice CN
Lattice vs CN CN Much smaller in size and less processing time Better align (improve WER) Contain posterior probabilities Lattice Contain complete hypothesis
Automatic Speech Recognition KALDI (http://kaldi.sourceforge.net/) Powerful open source tool for ASR Written in C++ and can be easily modified CN and Lattice support Complete recipes for some common corpora Supports Grid Engine
Named Entity Recognition Blender Name Tagger Simple to use Works with ACE data (corpus) Using MALLET Machine Learning toolkit
Prosody Analysis AuToBI The must tool for extraction prosodic features No need for introduction
Algorithm Speech AuToBI KALDI (ASR) Prosodic features Text New Representation Clusters Blender Name Tagger (NER) Model Blender Name Tagger (NER)
ASR Results KALDI Training time ~ 78 hours WSJ Models Triphone SGMM – (better, but not ready yet) 1 best Triphone Recognizer performance: 67.37% (WER)
Results
Future work Improve current ASR model Improve current CN approach Add n-best and oracle models Extend feature set Event Recognition
Thank you!