Abstract
Cluster analysis related to computational linguistics seldom concerns the Pragmatics level. Features of a corpus on the Pragmatics level relate to specific situations, including backgrounds, titles, and habits. To improve the accuracy of clustering for conversations collected from international students at Tsinghua University, contextual features were required. Here, we collected four hundred conversations as a corpus and built a Vector Space Model (VSM). Using the Oxford-Duden Dictionary and other methods, we modified the model and categorized the data into three groups. We testified our hypothesis through a Self-Organizing Map (SOM) neural network. The results suggest that the modified model yields a better outcome.
Introduction
Traditional text clustering often fails to distinguish between different conversational contexts because it relies too heavily on word frequency rather than situational meaning. In Natural Language Processing (NLP), documents are typically categorized using a Vector Space Model (VSM), which represents text as points in a high-dimensional space based on word counts. However, standard VSMs are often "frequency-blind" to Pragmatics—the linguistic level concerned with how context contributes to meaning.
For instance, conversations in a hospital and a bank might share common vocabulary, leading standard algorithms to produce "false positive" clusters. The most critical words for defining a situation (e.g., "scalpel" or "consignment") often appear infrequently, while high-frequency common words offer little discriminative power. This research aims to bridge this gap by incorporating situational backgrounds, titles, and habits into the clustering process.
Methodology
The study compared three distinct approaches to constructing the Vector Space Model:
- The "Default" Group: A standard VSM where values are determined solely by raw word frequency.
- The "Defined" Group: A modified VSM where 8 representative keywords were manually selected for each of the 8 categories (64 words total). If a word from this list appeared, its weight was doubled.
- The "Dictionary" Group: A VSM modified using the Oxford-Duden Dictionary. Ten words per category (80 words total) were selected based on the dictionary’s pictorial and situational organization. These words were also given double weight.
To evaluate these models, a Self-Organizing Map (SOM) neural network was employed. The SOM is ideal for this task as it maps high-dimensional data onto a 2D grid, allowing for visual inspection of how effectively conversations "clump" into distinct groups.
Dataset
The experimental corpus consisted of:
- 400 conversations collected from international students at Tsinghua University.
- 8 specific scenarios: Hospital, Restaurant, Renting House, Inside Class, After Class, Airport, Barber’s, and Bank.
- Scale: Approximately 10,000 words total, reduced to 5,000 tokens after removing "stop words."
Results and Discussion
The performance of the models was evaluated qualitatively through SOM distance maps and weight vector visualizations.
The Dictionary Group produced the most compact and distinct clusters. By using the Oxford-Duden Dictionary as a source for "contextual maps," the researchers captured Pragmatics effectively. Because this dictionary groups words by scene (e.g., items in a surgery room), it provides ready-made contextual anchors that align with human situational associations.
In contrast, the Default Group failed to produce clear boundaries. Cells were scattered across the model, making accurate categorization nearly impossible. This reinforces the idea that low-frequency words often carry the highest "information density" for categorization. By artificially boosting the weight of these situational words, the model becomes significantly more robust.
Specific categories like "Restaurant" were highly successful due to unique keywords (Food, Beverage, Waiter). However, some contexts remained "sticky"; for example, "Inside Class" and "After Class" shared too much overlapping vocabulary (teacher, student, book) to be perfectly separated even with weighting.
Conclusion
This study demonstrates that manually weighting contextual features significantly improves the accuracy of cluster analysis in conversational corpora. While effective, the current method relies on manual keyword selection. Future work should explore automating the extraction of these pragmatic features and refining the weighting sensitivity to optimize clustering outcomes further. Additionally, addressing keyword overlap—where words like "Money" appear in both "Bank" and "Airport" contexts—remains a challenge for future model iterations.
Summary updated