Information Fusion-2-Text: Bridging the Gap Between Fuzzy Integrals and Natural Language

TL;DR

Researchers have developed a way to turn the "black box" of complex information fusion into simple English sentences. By applying Linguistic Protoforms to the Choquet Integral, they can summarize the performance of multiple AI models (like DCNNs) into statements like "Most sources are important" or "Few walks are observed." This reduces thousands of confusing data points into actionable insights for non-experts.

The "Curse" of Information Overload in XAI

Information fusion—specifically using the Choquet Integral (ChI)—is a powerful tool for combining outputs from different AI models. It handles non-linear interactions better than simple averaging. However, when you fuse 7 different Deep Convolutional Neural Networks (DCNNs) across 30 different image classes, you end up with over 150,000 XAI metrics.

For a human analyst, this is a nightmare. There is too much data to actually "explain" anything. The authors argue that for AI to be truly Explainable (XAI), it must speak our language, not just output tables of Shapley values.

Methodology: From Math to Protoforms

The core innovation is the transition from Model-Centric Indices to Linguistic Protoforms.

1. The Choquet Integral & Walk Visitation

The ChI isn't just a formula; it’s a journey through a "Hasse diagram" (a lattice of subsets). Each unique path the data takes is called a Walk. If a model hasn't "walked" a certain path during training, it can't be trusted when it sees that path in testing.

2. Linguistic Protoforms (LPS)

The authors use a template: $Q$ $y$'s are $P$.

• $Q$ (Quantifier): Few, Many, Most.
• $y$ (Subject): The AI models or the data walks.
• $P$ (Summarizer): Important, Observed.

Figure 1: The proposed Fusion-2-Text pipeline, transforming raw sensor data through DCNNs and Choquet Integrals into linguistic summaries.

3. Fuzzy Logic Bridge

To determine if a statement like "Most sources are important" is true, the system uses Fuzzy Sets. A source is considered "important" if its Walk Centric Shapley value exceeds a threshold ($1/N$). The "Truth Value" of the summary is then calculated using fuzzy calculus.

Case Study: Remote Sensing

The researchers tested this on the AID dataset (aerial scenes) by fusing 7 DCNNs (including ResNet, Inception, and Xception).

Key Numerical-to-Text Findings:

• Source Importance: The system generated the summary "ResNet 50 and ResNet 100 are not contributing." This is a massive insight—it means you can delete these heavy models and speed up your system without losing accuracy.
• Data Bias: It produced the summary "Few walks are observed." This revealed that despite having 10,000 images, the data was highly redundant, only triggering five specific "walks" or logic paths in the fusion model.

Table 1: The fuzzy membership parameters used to define linguistic terms.

Critical Analysis & Conclusion

Why this matters

The beauty of this research is that it recognizes that transparency $eq$ visibility. Just because we can see the weights of a model doesn't mean we understand it. By using Linguistic Protoforms, the authors provide a "compression" of logic that retains the semantic meaning while discarding the numerical noise.

Limitations

Currently, the fuzzy membership functions (what defines "many" vs "most") are manually tuned. This introduces human bias. Future versions should learn these thresholds directly from how human experts describe data.

Future Outlook

This "Fusion-2-Text" approach is a stepping stone toward Interrogable AI. Imagine asking your fusion system: "Why didn't you rely on the thermal sensor for this classification?" and receiving the answer: "Because in many cases, that source was unobserved during training for this specific class."

Takeaway: To make AI more trustworthy, we need to move beyond heatmaps and bar charts. We need AI that can write its own report card.