Embedded Files
We propose a novel deep neural network to predict the severity of bug reports. We generate word embeddings and pass
them through GRU based model. We compare different attention mechanisms and word embeddings
A Project by - RIJUL DAHIYA
Mission of the project
Mission of the project
Software development and deployment is often infested with bugs. Bugs lead to bug
reports. Then comes the task of prioritizing bug reports. Widely, manual priority is given
to these reports using domain knowledge and expertise but the traiger. This as
expected is an inaccurate and legacy approach which leaves a scope for significant
optimization. We propose a Hierarchical Attention based recursive deep learning
approach. Our method entails taking data from linux and chrome databases and then
passing it through a deep learning model. We experimented with various embedding
techniques namely GLOVE. Fasttext and word2vec to generate word embeddings,
which were then sent via a Hierarchical Attention network. The hierarchical attention
network performs well on datasets with well-structured data (Linux bugs), but it does not
perform well on datasets with no such structure. The document representation obtained
from the hierarchical attention network was then concatenated with the VALDER
sentiment score which was used as input to the softmax layer for predictions.
Research Questions:
Research Questions:
We would try to find out what is the effect of adding an emotion value score to our
model. Does emotion play an important role in predicting the priority class of our bug
report. We would also try to compare different types of word embeddings such as
word2vec to find out which gives the best results.
Model Architecture:
Model Architecture:
In our approach we used the model proposed by Volodymyr Lyubinets,Taras
Boiko,Deon Nicholas,”Automated labeling of bugs and tickets using attention-based
mechanisms in recurrent neural networks”. We used the word2vec model to get word
embeddings which were passed through a Hierarchical Attention network. Hierarchical
attention works effectively on datasets with well-structured data (Linux bugs). However,
the hierarchical attention network does not provide good results on datasets without
such structure. To overcome this problem, two changes were made to the hierarchical
attention network model:
Appending Emotion Score:
Appending Emotion Score:
VADER (Valence Aware Dictionary and sEntiment Reasoner) is a lexicon and rule-
based sentiment analysis tool that is tuned in to social media sentiments. VADER
employs a mix of A sentiment lexicon is a collection of lexical characteristics (e.g.,
words) that are categorized as positive or negative based on their semantic orientation.
VADER not only reports the Positivity and Negativity scores, but also how positive or
negative an emotion is.
Conclusion and future work
Conclusion and future work
● Currently the best results that we got was using the model in which emotion
scores were concatenated with the document vector representation that was
used as an input to the final softmax layer.
● We can try running the above model for the chromium dataset as well
● On comparing different models by just changing the word embeddings, we
observed that glove embeddings gave the best results
Questions?
Questions?
Contact errijuldahiaya(at)gmail.com to get more information on the project
Page updated
Google Sites
Report abuse