AAI_2025_Capstone_Chronicles_Combined

‭ResolveAI‬

‭Model Architecture‬

‭The model is built using a sequential architecture that transforms tokenized text input into‬ ‭dense vector representations via an embedding layer, followed by bidirectional LSTM layers to‬ ‭capture contextual information from both past and future tokens. The first bidirectional LSTM‬ ‭layer has 64 units, while the second has 32 units. This progressive reduction of units helps in‬ ‭extracting hierarchical features while still prioritizing efficiency. The architecture features‬ ‭dropout regularization layers after both bidirectional LSTM and Dense layers to help prevent‬ ‭overfitting.‬ ‭For initial testing, the dropout rate was set to 0.5 to ensure robustness while training, but‬ ‭this is one of the several layers that we will be optimizing during hyperparameter tuning.‬ ‭GlobalMaxPooling1D was applied to reduce dimensionality while still being able to preserve‬ ‭critical information, which makes the model more efficient. After the bidirectional LSTM layers,‬ ‭a dense layer with 128 units and a final dense layer with a single unit are used to transform the‬ ‭extracted features from the BiLSTM layers into a final classification decision.‬ ‭The model’s input consists of preprocessed and padded textual data extracted from key‬ ‭fields (subject, body, and answer), along with additional encoded metadata “type”, while the‬ ‭output is a binary label indicating ticket urgency. The bidirectional LSTM layers use tanh and‬ ‭sigmoid activations by default because tanh effectively captures the underlying patterns in‬ ‭sequential data by mapping values between -1 and 1, which allows for a more balanced‬ ‭representation of positive and negative relationships (Baheti P., (2021). The sigmoid activation‬ ‭for the bidirectional LSTM layers is used in the gating mechanisms to control the flow of‬ ‭information, which ensures that the model retains relevant features and discards unnecessary‬ ‭ones.‬

‭15‬

63