TL;DR
This paper introduces a novel annotation scheme with 16 distinct labels to categorize sentences in patent descriptions based on their discursive roles. By creating and analyzing a human-annotated corpus, the authors aim to improve our understanding of patent language and lay the groundwork for automated analysis tools.
What Problem Does This Paper Address?
Patent descriptions are crucial for understanding inventions, securing patents, and assessing competition. While the structure of scientific articles has been extensively studied, patent descriptions, despite their increasing importance and volume, have received less linguistic attention. Existing work often focuses on specific parts like claims or abstracts, overlooking the rich information within the full description. This paper addresses the need for a systematic way to analyze the discursive structure of entire patent descriptions, which can aid in patent writing, feature extraction, and potentially detecting anomalies in patent drafting.
Methods
The core of this work is the development of a 16-label annotation scheme designed to capture the discursive roles of sentences within patent descriptions. This scheme was iteratively refined through an annotation process involving two annotators. Initially, 12 labels were proposed based on the typical sections of a patent description (e.g., Technical Field, Background Art, Summary, Embodiments). This was expanded to 15 labels after initial annotation revealed the need to distinguish between "Advantage" and "Problem" related to the invention versus prior art. A further refinement added a 16th label, "Section subtitle," to handle variations in how sub-sections are introduced. The annotation was performed at the sentence level, with annotators having access to context. Inter-annotator agreement was measured using Cohen's kappa.
Datasets
The authors created a new benchmark dataset consisting of 16 patent descriptions. These were selected from applications published by the European Patent Office (EPO) and classified using the Cooperative Patent Classification (CPC) system. Two applications were randomly selected per CPC domain, resulting in two datasets for evaluation and inter-annotator agreement analysis.
Key Insights and Ideas
- Patent language is unique: Patent descriptions have a distinct writing style and structure that requires a specialized annotation scheme, differing from that used for scientific articles.
- Discursive roles are complex: Identifying the precise discursive role of a sentence can be challenging, especially when distinguishing between prior art and the invention itself, or when sentences contain dual statements.
- Annotation scheme refinement is crucial: The iterative process of annotation and discussion led to the refinement of the label set, highlighting the practical difficulties in defining clear boundaries between categories.
- Agreement varies by label: Labels related to titles and factual statements (e.g., "Technical field," "Rephrased claim") showed higher agreement than those requiring more interpretation (e.g., "Reference advantage," "Figure description").
- "Other" label is a safety net: The "Other" label, for sentences that don't fit any category, was used sparingly, suggesting the 16-label scheme is largely comprehensive.
Key Results
The study achieved a notable inter-annotator agreement score of 0.69 (Cohen's kappa) for the entire dataset after revisions, which is comparable to similar work in scientific discourse analysis. The annotation process identified specific pairs of labels that were frequently confused, such as "Reference vs. Embodiment" and "Embodiment vs. Figure description," providing valuable insights into the nuances of patent language. The majority of annotations fell into categories like "Figure description," "Embodiment," and "Rephrased claim," aligning with the typical goals of patent drafting (explaining the invention and its novelty).
Limitations and Future Work
The authors acknowledge that their sample dataset of 16 documents is relatively small and may not be sufficient for training robust classification models. They also suggest that having a lead annotator to supervise the process could improve consistency and reduce annotation errors. Future work includes expanding the dataset, employing active learning for more efficient annotation, and potentially developing multi-labeling or multi-layer annotation schemes to handle ambiguous sentences and provide a more hierarchical understanding of patent structure. The ultimate goal is to create a valuable linguistic resource for tasks like patent classification, novelty detection, and computer-assisted patent drafting.
Summary updated