This PhD thesis was conducted as a CIFRE collaboration between the LIRIS laboratory of INSA Lyon (France) and Gauc SAS, part of Auvalie Group, a Lyon-based company that provides business intelligence tools to
technology startups. The objective of this PhD is to propose a method for detecting weak signals automatically in continuous streams of unstructured text, from the construction of the representation on which the detection operates to the detection itself. A weak signal is an early, low-intensity indication of a change to come, such as a company entering a new market or a competitor repositioning itself. Such indications appear in public sources, but they are rare relative to the volume of text that surrounds them, and finding them by hand is not feasible at the scale organizations need to monitor.

We propose an approach for the automatic detection of weak signals in dynamic temporal knowledge graphs built from text. Our proposal is divided into three main tasks. The first task constructs a consistent knowledge graph from raw documents without a predefined ontology. The second task extends this construction to a dynamic temporal knowledge graph, in which every fact carries a validity period, and the graph grows as documents arrive. The third task reads the resulting graph and reports the weak signals it contains. Our proposal combines information extracted by large language models with resolution mechanisms that operate on pre-trained embeddings and on the structure of the graph.

The knowledge graph construction part is a zero-shot, incremental pipeline named iText2KG, composed of four modules. The Document Distiller rewrites raw documents into structured semantic blocks according to a user-defined blueprint, which reduces noise and guides extraction without imposing a fixed ontology. The main contribution in the first task is the resolution of entities and relations as documents are read: the Incremental Entities Extractor maintains a global entity set and matches each locally extracted entity against it using cosine similarity over pre-trained embeddings, and the Incremental Relations Extractor applies the same mechanism to relations. The Graph Integrator produces the final graph in a graph database. Evaluated on scientific articles, CVs, and websites, iText2KG obtains lower false discovery rates for entity and relation resolution on structured documents and performs better than the baseline methods across all three scenarios.

We propose ATOM, an adaptive and optimized approach for the construction of dynamic temporal knowledge graphs, which addresses four limitations of iText2KG: non-exhaustive extraction, instability across runs, limited scalability, and the absence of temporal modeling. Module~1 decomposes each document into atomic facts, that is, short self-contained statements each conveying one piece of information. Module~2 extracts 5-tuples (subject, relation, object, validity start, validity end) from these atomic facts in parallel, and applies a dual-time model that separates the observation time, when a document was ingested, from the validity period, when the fact holds. Module~3 merges the resulting temporal knowledge graphs in parallel through iterative pairwise binary merges. ATOM obtains 18% higher factual-temporal exhaustivity and 33% higher stability than direct extraction from full paragraphs, and reduces latency by more than 90% over the baseline methods.

Finally, we propose C-Unseen, a self-interpretable framework that detects weak signals in the graph constructed by ATOM. It first defines a weak signal in this setting as a rare subgraph whose content is in tension with the dominant narrative of a snapshot and whose tension continues in later snapshots. It then detects such subgraphs with two modules. The Rare Subgraphs Extractor establishes the dominant narrative of each snapshot, isolates the quintuples in tension with it, and spans them with a connecting subgraph built from the shortest paths between their entities. The Weak Signal Alerter compares that subgraph with those retained from earlier snapshots and labels it a weak signal when its content advances a tension already visible before. Each alert is delivered with its subgraph and a natural-language interpretation, so that an analyst can verify it rather than accept it.

Additionally, the rationale of the proposed approach has been verified with a set of experiments on Wiki-OpenAI, a benchmark for weak signal detection introduced in this thesis, and on an anonymized variant of it, and the complete pipeline has been applied to competitive intelligence data.