Page 129 - Kaleidoscope Academic Conference Proceedings 2020
P. 129
Industry-driven digital transformation
Figure 7 shows the accuracy, precision, recall and F1-score
of the model when it was tested against 100 messages
containing 50 positive and 50 negative messages.
The detailed accuracy results with precision, recall and F1-
score of the complete classification module as a whole is
given in Table 2.
Table 2 – Chat classification module accuracy
Chat Classification Actual
Module Positive Negative
Positive 46 4
Predicted Negative 4 46
Figure 6 – Variation of model accuracy
Table 2 shows the actual and predicted classes for the
complete classification module. Special attention was paid to
As the accuracy of classification increases with the size of the reducing false negatives. Reducing false negatives has in turn
training data, it is possible to achieve very high accuracies if increased the false positives. However, this does not affect the
the training data can be enriched further with more relevant overall objective of the project as child safety is not to be
data. This will enable us to develop an ideal classification compromised at any cost. Figure 8 shows the accuracy of the
model. overall classification module.
The detailed accuracy results in terms of precision, recall and
F1-score of the model are given in Table 1.
Table 1 – SVM model accuracy
Actual
SVM Model
Positive Negative
Positive 42 1
Predicted
Negative 8 49
Table 1 shows that the actual and predicted classes for the
SVM model. From the results given in Table 1, it is possible
to see that the false-positive prediction is very much less
compared false-negatives. This is acceptable as the main
objective of this work is to predict any inappropriate intention
of users involved in chatting.
Figure 8 – Accuracy of the classification module
The classification module was trained using 2000 epochs.
The final accuracy of the classification module is 92%. From
Figure 8, it can be seen that the accuracy achieved by the
proposed technique is comparable with that of the other
existing techniques.
6. CONCLUSIONS
In this paper, the authors have presented a technique that
would enable end users to elicit the appropriateness of the
intention of the chat messengers. The proposed technique
consists of multiple stages that work together for identifying
similar chat messages and group them to enrich the semantic
content of the messages, carry out disambiguation for
Figure 7 – Accuracy results of the model semantically rich words contained in the messages, and
finally to classify the messages as appropriate and
inappropriate. The experiments carried out for verifying the
– 71 –