AAI_2025_Capstone_Chronicles_Combined

Cinema Analytics and Prediction System

11

To address these limitations, transformer-based models like BERT (Bidirectional Encoder

Representations from Transformers) have been developed. Introduced by Devlin et al. (2019),

BERT leverages a deep bidirectional transformer architecture to generate context-aware

embeddings. Unlike TF-IDF, BERT understands the semantics of entire sentences, allowing it to

detect deeper thematic similarities. This makes it especially powerful for tasks involving plot

understanding, genre prediction, and nuanced recommendation, and it has been widely

adopted in academic and industry NLP pipelines.

Recurrent neural networks, such as Long Short-Term Memory (LSTM) models, have also

been used in text classification and recommendation tasks. LSTMs are well-suited for handling

sequential data and have been shown to capture long-range dependencies in natural language

(Hochreiter & Schmidhuber, 1997). When combined with pre-trained GloVe word embeddings

(Pennington et al., 2014), LSTMs can provide more meaningful input representations than raw

token sequences, though they may still underperform compared to transformer-based models

when it comes to modeling context at scale.

Hybrid architectures that integrate contextual embeddings with structured data have also

emerged as an effective method for recommendation tasks. For example, Liu et al. 2024,

proposed a multimodal recommender that fuses text embeddings from language models with

numerical metadata (e.g., ratings, runtime) to improve predictive accuracy. These approaches

are particularly beneficial for domains like movies, where metadata such as popularity and

runtime often correlate with genre or audience preferences. Models that combine BERT with

normalized metadata have been shown to outperform text-only models (Yang et al., 2024), this

178

Internal