From Open Source Intelligence to Decision Making: a Hybrid Approach

We provide an overview of tools enabling users to utilize data from open sources for decision-making support in weakly-structured subject domains. Presently, it is impossible to replace expert data with data from open sources in the process of decision-making. Although organization of expert sessions requires much time and costs a lot, due to insufficient level of natural language processing technology development, we still have to engage experts and knowledge engineers in decision-making process. Information, obtained from experts and open sources, is processed, aggregated, and used as basis of recommendations, provided to decision-maker. As an example of a weakly-structured domain, we consider information conflicts and operations. For this domain we propose a hybrid decision support methodology, using data provided by both experts and open sources. The methodology is based on hierarchic decomposition of the main goal of an information operation. Using the data obtained from experts and open sources, we build the knowledge base of subject domain in the form of a weighted graph. It represents a hierarchy of factors influencing the main goal. Besides intensity, the impact of each factor is characterized by delay and duration. With these parameters taken into account, main goal achievement degree is calculated, and changes of target parameters of information operation object are monitored. In order to illustrate the suggested hybrid approach, we consider a real-life example, where we detect, monitor, and analyze actions intended to discredit the National academy of sciences of Ukraine. For this purpose, we use specialized decision-making support and content monitoring software.

💡 Research Summary

**
The paper addresses the challenge of supporting decision‑making in weakly‑structured domains—specifically information operations (IO)—by combining expert knowledge with open‑source intelligence (OSINT). Recognizing that expert workshops are costly and time‑consuming, yet acknowledging that current natural‑language‑processing (NLP) and sentiment‑analysis tools are not yet mature enough to replace human expertise, the authors propose a hybrid methodology that leverages the strengths of both data sources.

The methodology consists of five main steps. First, a preliminary investigation identifies the target object of the IO and selects relevant performance indicators. Second, a group of domain experts conducts a structured session to decompose the main IO goal into a hierarchy of sub‑goals, estimating for each sub‑goal three temporal parameters: intensity (the magnitude of impact), delay (the time before the impact becomes observable), and duration (the length of the impact). These estimates are captured using the Distributed Collection and Processing of Expert Information (SDCPEI) system and an Expert Evaluation System (EES).

Third, the collected expert assessments are used to build a knowledge base (KB) in the form of a weighted directed graph. Nodes represent sub‑goals or factors, edges encode hierarchical relationships, and weights reflect the intensity values. The fourth step enriches and “fine‑tunes” this KB with data harvested from open sources—news articles, blogs, social‑media posts, official documents—through textual mining, keyword frequency analysis, and sentiment detection. Because current NLP tools cannot fully automate the extraction of structured information, a knowledge engineer manually maps raw OS data into the KB’s required format.

Finally, the system computes a Goal Achievement Degree (GAD) for each sub‑goal by integrating intensity, delay, and duration (typically via a weighted sum or product). Historical GAD values are averaged to establish a baseline. Real‑time GADs are compared against this baseline; when the current GAD exceeds a predefined threshold, the system flags a likely deterioration of the target object’s indicators (e.g., reduced funding, reputational damage). This enables early warning and supports the formulation of counter‑measures.

The authors discuss the approach’s advantages: (1) reduced reliance on expensive expert sessions, (2) exploitation of the large volume and diversity of OS data, (3) a nuanced, multi‑dimensional model that distinguishes between immediate and lingering effects, (4) reusability of the KB for future analyses, and (5) the possibility of remote expert collaboration via SDCPEI. They also acknowledge limitations: the need for periodic expert input to keep the KB up‑to‑date, the labor‑intensive preprocessing of heterogeneous OS data, and the difficulty of formulating precise queries for complex IO components.

A concrete case study illustrates the methodology. The authors examine a suspected information campaign aimed at discrediting the National Academy of Sciences of Ukraine (NAS). Budget data from 2014‑2016 show a decline in state funding for NAS, which the authors hypothesize is partially driven by a coordinated media attack. Experts decompose the overarching IO goal into fifteen sub‑goals (e.g., bureaucratic inefficiency, corruption, under‑representation of scientific achievements, attacks on individual NAS officials). Each sub‑goal receives intensity, delay, and duration estimates. Open‑source monitoring captures the frequency and sentiment of references to NAS across news portals, blogs, and social networks. By integrating these streams, the weighted graph yields GADs that correlate with the observed budget reductions. The model successfully predicts that when the GAD for “corruption” and “negative media coverage” surpasses certain levels, the NAS’s funding indicator deteriorates, providing a quantitative early‑warning signal.

In conclusion, the paper demonstrates that a hybrid expert‑OSINT framework can deliver cost‑effective, dynamically updated decision support for complex, weakly‑structured problems such as information warfare. While current NLP capabilities limit full automation, the authors propose future research directions including advanced language models, automatic knowledge‑graph construction, and real‑time feedback loops to further reduce human overhead and improve predictive accuracy.