Using Convolution Neural Networks, fundMyLife classifies financial questions posed by our users. And through our sentiment analysis module tracks user satisfaction during the fact finding phase.
3. Problem Statement
• How do consumers know which financial services
to use for each life stage?
• Who can consumers ask when they do not know
anything about loans, credits, savings, interests
etc.?
5. Agenda of Talk
• Using NLP to predict question tags to match user to
the right financial advisers
• Using sentiment analysis to score questions and
improve matching
6. Schematic Overview of fML DL
Questions
Tag
Prediction
Sentiment
Analysis
External
Validation
Training
Internal
Validation
word2vec
GloVE
Single Layer
Multiplicative
LSTM
Output
OutputCNN
Personal
Finance
Insurance
7. Using NLP to match users to advisers
AXA
GE
AIA
Insurance A
Insurance A
Insurance A
“Insurance A”
Tag Prediction
9. • Similar to 20 newsgroup dataset problem*
• Used Word embeddings (word2vec nn)/(GloVe nmf)
instead of one-hot encoding
• Framework: Keras 2.0.2 (Tensorflow 1.0)
*Lang, K. (1995). NewsWeeder : Learning to Filter Netnews ( To appear in ML 95 ). In Proceedings of the 12th
International Machine Learning Conference (ML95). http://doi.org/10.1.1.22.6286
Brief Implementation Details
10. Architecture - 1D CNN with word embedding
• Hyperparameters
• 4 merged ngrams of width 3, 5, 7, 9 convolution(ngrams)->maxpool->dropout
• 300 feature maps each embedded word vector
• Datasets (average sentence length ~ 40; 90th Percentile 90 words)
12. Sentiment analysis to score questions and advisers
0.1103234
Sentiment Scoring Monitor sentiment
change over time
between adviser
and user
13. Results
"Exclusion due to fibroid or
cysts?. I would like to know
can the insurance company
exclude me from disorder of
uterus if I have fibroid or
cysts? this is really unfair
for me and I am still
young…"
15. • Kim, Y. (2014). Convolutional Neural Networks for
Sentence Classification. arXiv: 1408.5882.
• Lang, K. (1995). Newsweeder: Learning to filter
netnews. Proceedings of the 12th international
conference on machine learning.
• Radford, A., Rafal J., Ilya S. (2017) Learning to
Generate Reviews and Discovering Sentiment.
arXiv:1704.01444.
References