Abstract
As the first step in an automated text summarization algorithm, this work presents a new method for automatically identifying the central ideas in a text based on a knowledge-based concept counting paradigm. To represent and generalize concepts, we use the hierarchical concept taxonomy WordNet. By setting appropriate cutoff values for such parameters as concept generality and child-to-parent frequency ratio, we control the amount and level of generality of concepts extracted from the text.
Introduction
Traditional Information Retrieval (IR) systems often rely on raw word frequency to determine importance. However, this approach is "conceptually blind." For instance, if a news article mentions "apples," "oranges," and "bananas" once each, a standard word counter perceives three unrelated items. A human reader, conversely, immediately identifies the overarching topic as "fruit."
The challenge in automated summarization is that simple word counting misses these semantic relationships and generalizations. While some researchers have attempted to bridge this gap using complex statistics or deep linguistic analysis, this paper proposes a robust, knowledge-based method for identifying topics by navigating a hierarchical taxonomy of concepts.
Methodology
The proposed system utilizes a "knowledge-based concept counting paradigm" built upon WordNet, a comprehensive lexical database of English. The approach hinges on three primary mechanisms:
1. Concept Weighting and Taxonomy
Instead of merely counting individual word occurrences, the system calculates the "weight" of a concept. This weight includes the frequency of the concept itself plus the cumulative frequency of all its sub-concepts (children in the hierarchy). This allows the system to recognize that specific mentions contribute to a more general idea.
2. The Branch Ratio ($\mathcal{R}$)
The core innovation of this work is the Branch Ratio. To determine if a parent concept serves as an effective summary for its children, the system calculates the ratio of the most frequent child's weight to the sum of all children's weights.
- High Ratio: If one child dominates (e.g., "Apple" accounts for 90% of "Computer Company" mentions), the system retains the specific child concept.
- Low Ratio: If mentions are distributed across many children (e.g., Toshiba, IBM, and NEC are mentioned with similar frequency), the system generalizes to the parent concept: "Computer Company."
3. Starting Depth ($\mathcal{D}_s$)
To prevent trivial generalizations—such as categorizing every topic under the extremely broad WordNet root "entity"—the algorithm only begins its search for topics after a predefined depth in the taxonomy tree. This ensures the identified topics remain meaningful and relevant.
Experimental Setup
The system was evaluated using 50 Business Week articles from 1993-94 regarding information processing, averaging 750 words each. Professional abstracts from an online service served as the "ground truth." The system's task was to identify an "interesting wavefront" of concepts and use them to extract the 8 most representative sentences from each text.
Results and Analysis
The performance of the Concept Counting method was compared against baseline methods of random sentence selection and standard word counting. The results demonstrate a significant improvement in both recall and precision.
| Method | Recall | Precision |
|---|---|---|
| Concept Counting (Proposed) | 0.32 | 0.35 |
| Plain Word Counting | 0.23 | 0.28 |
| Random Selection | 0.18 | 0.22 |
These findings suggest that hierarchical knowledge bases provide a much clearer signal for topic identification than raw vocabulary usage alone. Notably, the system achieved these results without the use of part-of-speech taggers or complex syntactic parsers, highlighting the inherent power of structured conceptual hierarchies.
Conclusion
This work demonstrates that topic identification is essentially a search for the optimal level of abstraction. By using the Branch Ratio as a mathematical proxy for linguistic intuition, the system can effectively find where specific mentions in a text converge into broader, more representative categories. Future work will focus on incorporating linguistic tools to move from sentence extraction to the generation of novel summary structures.
Summary updated