DiscoGraMS: Enhancing Movie Screen-Play Summarization using Movie Character-Aware Discourse Graph

Abstract: Summarizing movie screenplays presents a unique set of challenges compared to standard document summarization. Screenplays are not only lengthy, but also feature a complex interplay of characters, dialogues, and scenes, with numerous direct and subtle relationships and contextual nuances that are difficult for machine learning models to accurately capture and comprehend. Recent attempts at screenplay summarization focus on fine-tuning transformer-based pre-trained models, but these models often fall short in capturing long-term dependencies and latent relationships, and frequently encounter the "lost in the middle" issue. To address these challenges, we introduce DiscoGraMS, a novel resource that represents movie scripts as a movie character-aware discourse graph (CaD Graph). This approach is well-suited for various downstream tasks, such as summarization, question-answering, and salience detection. The model aims to preserve all salient information, offering a more comprehensive and faithful representation of the screenplay's content. We further explore a baseline method that combines the CaD Graph with the corresponding movie script through a late fusion of graph and text modalities, and we present very initial promising results.

• The paper introduces CaD Graphs, a novel representation that captures complex character interactions and non-linear narrative structures in movie scripts.
• It employs a late fusion model combining a GNN for graph encoding and an LED for textual encoding, significantly improving precision and recall metrics.
• Extensive ablation studies validate each component's contribution, offering practical insights for advancing screenplay summarization in NLP applications.

Analyzing DiscoGraMS: A Novel Approach to Movie Screenplay Summarization

The task of summarizing movie screenplays, as explored in the work titled "DiscoGraMS: Enhancing Movie Screen-Play Summarization using Movie Character-Aware Discourse Graph," is complicated by the unique narrative structures inherent in cinematic scripts. This paper introduces an innovative methodology that addresses the complexities of screenplay summarization through the development of DiscoGraMS—a system leveraging a movie character-aware discourse graph (CaD Graph). It aims to enable improved summarization by preserving salient relationships among characters, scenes, and dialogues.

Key Contributions

The paper presents three primary contributions:

1. Introduction of CaD Graphs: The characterization of movie scripts as CaD Graphs represents a pioneering step in screenplay summarization research. These graphs encapsulate crucial narrative elements and semantically relevant relationships, including intricate character interactions and plot developments that challenge traditional text-based approaches. This graphical representation is vital in managing non-linear plot structures, such as flashbacks and parallel storylines, which are prevalent in screenwriting.
2. Late Modality Fusion Model: The study proposes a methodological advancement through a late fusion model that unites CaD Graphs with the script's textual content. This fusion leverages a Graph Neural Network (GNN) and a Longformer Encoder-Decoder (LED) model, which work in tandem to integrate structural and narrative information, thereby enhancing the overall summarization process.
3. Ablation Studies: Through comprehensive ablation studies, the authors demonstrate the individual and combined efficacy of the model components—namely, the GNN-based CaD Graph encoding and the LED textual encoding. This analysis shows that each component contributes distinct value to the summarization task, with the GNN component notably enhancing performance metrics.

Results and Evaluation

The experimental evaluation utilizes well-established benchmarks for summarization, including ROUGE and BERT scores. The proposed LGAT model exhibits superior performance across all baseline models. Notably, LGAT achieved significant improvements in precision and recall measures, as indicated by the BERT scores, although it underperformed on the ROUGE-L metric due to the constrained context window imposed by hardware limitations. These metrics underscore the model's ability to maintain narrative coherence and capture semantically significant content, thanks to its graph-based representation and processing.

Implications and Future Directions

This research holds important implications for NLP applications beyond screenplay summarization, including tasks like question-answering, genre identification, and salience detection. It highlights the utility of knowledge-based text representations and encourages further exploration of graph-based modeling for complex narrative structures. However, the study acknowledges limitations, such as the absence of co-reference resolution which may limit the model’s interpretive abilities. Future work should address these challenges, potentially involving larger context windows and more robust character-relationship mappings to refine output quality further.

In sum, the DiscoGraMS framework proposes a forward-thinking approach to the challenge of screenplay summarization by effectively handling narrative complexity with character-aware discourse graphs. This work not only advances the field of NLP but promises broader applications in narrative analysis and computational filmmaking studies.