Human Activity Role Identification using Feature Vector and Encoding Techniques on Natural Language Sentences

Anam Arshad, Vivek Tiwari, Mayank Lovanshi, Rahul Shrivastava

Abstract:

Role Identification has the potential to enhance activity recognition
applications since it adds more information. Most of the works in
the field of activity recognition and role identification are mainly
dominated by models that use images and videos. The existing
datasets of human activity are not capable of role identification. In
this view, this work attempt to develop a novel Human Activity
Role Identification Dataset and a novel Computational Recurrent
Model that takes textual data as input. Additionally, various feature
vector generation methods like N-Grams extraction, Unique word
extraction, and Word2Vec are used to encode the input data into
feature vectors that describe the relationship between sequences of
words. To determine the fundamental roles, these feature vectors are
trained on various types of Recurrent Neural Networks (i.e. RNN,
LSTM, GRU). The proposed model is validated on evaluation metrics
such as Precision, Recall, F1 Score, etc., using Recurrent Neural
Networks like RNN, LSTM, and GRU. Hence, the combination of
LSTM with unique word extraction methods outperforms with an
F1 Score, precision and recall by 0.44, 0.36 and 0.58 respectively. So
this role identification work may help to bind roles with entity and
objects in human activity recognition.