GPSR

Source code used for GPSR Challenge at RoboCup 2015-2017@home

Docs

1. Voice Command
   We are using Kaldi speech engine.

2. Tagging
   nltk default perceptron tagger had poor performance. We used Two-step Convolutional Neural Network to overcome this problem.

3. Normalization Convert polite expression, and conjuction to normal, and separated sentence for normalization.

4. Classification of command Classify the command to "deliver", "grasp", "follow", "answer", "search", "move, "manipulation" categories.

5. Pronounciation fix of noun
   For understanding non-native speaker speech, we used CMUDict + Ngram for word similarity matching and search for word candidate.

6. Extract location and objective
   Extract location and ojective by the five sentence structures of English grammer.

Library

scripts/CommandAnalyzer.py have library functionality. Include to your project to use it.

Dependencies

nltk

1. Brown Corpus
2. WordNet
3. CMUDict
4. Tagger

nlpnet
kaldi_ros
NGram