Financial Industry Business Ontology (FIBO)

• FIBO is a formal ontology that defines and categorizes financial concepts using RDF/OWL with unique IRIs and hierarchical taxonomies.
• It supports semantic search and classification through graph transformations, BFS label assignment, and integration with machine learning models.
• Combining ontology-driven reasoning with techniques like Random Forest and embeddings, FIBO improves hypernym ranking and semantic similarity evaluation.

The Financial Industry Business Ontology (FIBO) is a formal ontology developed to represent financial industry concepts, taxonomies, and relationships. It provides a shared semantic layer for structuring knowledge in finance, supporting interoperability, semantic search, and more precise classification of financial terms. FIBO is typically encoded in RDF/OWL, enabling machine reasoning over taxonomies such as asset classes, instruments, regulatory entities, and related constructs.

1. Formal Representation and Structure

FIBO is authored using the Web Ontology Language (OWL), facilitating the representation of financial concepts as classes and the relationships among them as object properties. Each OWL class in FIBO is uniquely identified by an IRI (Internationalized Resource Identifier), and subclass relations encode domain-specific taxonomies. For computational applications, FIBO is frequently converted into a directed acyclic graph (DAG), where each class constitutes a node and subclass relations define the directed edges. This transformation supports efficient traversal and ancestor search for downstream tasks. In the implementation described in "JSI at the FinSim-2 task: Ontology-Augmented Financial Concept Classification" (Perdih et al., 2021), each FIBO class becomes a node in a NetworkX MultiDiGraph $G=(V,E)$, where $V$ is the set of FIBO class IRIs, and $E$ consists of directed edges for OWL subclass relations.

2. Ontology-Augmented Financial Concept Classification

FIBO's structural properties provide a foundation for semantic classification of financial terms. The workflow typically consists of mapping a candidate term to one or several FIBO nodes, followed by generalization to the nearest ancestor label among a selected set of hypernyms. For example, in the FinSim-2 shared task, ten hypernym labels—such as “Bond,” “Fund,” or “Equity Index”—serve as classification targets.

The process involves:

• Concept-to-node mapping: Identifying FIBO node(s) corresponding to the term.
• Ancestor search (generalization): Performing a breadth-first search (BFS) upwards along superclass edges until a label in the set of valid hypernyms is found.
• Default handling: Assigning a default "sink" label (e.g., "Credit Index") to terms not mappable or generalizable within the ontology hierarchy.

This procedure is formalized as follows: Let $M_0$ denote the set of mapped nodes for input term $t$. Iteratively, compute $M_{k+1} = \bigcup_{v\in M_{k}} Parents(v)$ and halt at smallest $k$ such that $M_k \cap L \ne \emptyset$, with $L$ the target hypernyms. If no label is found after traversing the entire height of the hierarchy, a default hypernym is assigned (Perdih et al., 2021).

Pseudocode for BFS Label Assignment

$E$0

3. Integration with Representation Learning and Machine Learning

FIBO is further leveraged in hybrid semantic classification systems that combine ontology-based reasoning with machine learning. A representative approach employs distributional word embeddings (e.g., Word2Vec) to generate dense vectors for concept terms. Each term $t = w_1 \ldots w_m$ is embedded as $V$0, where $V$1 is the word embedding, and the Random Forest classifier is trained on these vectors with ontology-derived labels.

The merged decision process prioritizes the direct ontology label, augmented by a probability-ranked list from the Random Forest. The combination algorithm first places the ontology-derived hypernym atop the ranking, followed by the classifier’s descending ranked predictions, eliminating any duplicates (Perdih et al., 2021).

4. FIBO in Deep Semantic Similarity and Hypernym Ranking

In the context of semantic similarity for hypernym extraction and ranking, FIBO serves as both a source of annotated definitions and the backbone for negative-sample construction. In "Learning Semantic Text Similarity to rank Hypernyms of Financial Terms" (Ghosh et al., 2023), the FIBO hierarchy is modeled as a tree with three tiers: super-class (Tier 0), first-level child (Tier 1), and leaf hypernyms (Tier 2). Preprocessing ensures that only labels precisely matching a controlled hypernym set are retained.

For negative sampling:

• For each (term, hypernym) pair, positive and structured negative samples are generated.
• Two hierarchical functions are defined: $V$2 Tier 0, $V$3 Tier 1.
• Negative hypernym labels with shared Tier 1 (first child) or Tier 0 (root) ancestry to the gold label are scored less severely than those with no hierarchical overlap, using empirically-tuned scaling factors in similarity assignments.

The table below (derived from (Ghosh et al., 2023)) summarizes the scoring for negative samples:

Category	Negative score
Same first child (FC)	$V$4
Same root only (FR)	$V$5
Neither FC nor FR	$V$6

Base embeddings for both terms and hypernyms are computed using FinBERT and FinISH within a Siamese Sentence-BERT architecture, trained with Multiple-Negative Ranking Loss and Online Contrastive Loss. Inference scores for each pair $V$7 are calculated as $V$8, where $V$9 are cosine similarities from individual models.

5. Empirical Evaluations and Impact

Empirical results demonstrate FIBO’s utility for improving hypernym ranking and classification tasks in the financial domain. In (Perdih et al., 2021), the ontology-only BFS generalization method achieved 0.88 accuracy on an internal validation set, while a merged ontology and Random Forest solution maintained the same strict accuracy but improved average label ranking to 1.19.

Evaluation on the FinSim-2 task test set reports:

System	Accuracy	Average Rank
Ontology only	0.88	n/a
Random Forest only	0.85	1.37
Merged (Ontology+RF)	0.81	1.32

For semantic similarity ranking (Ghosh et al., 2023), FIBO-based negative sampling yielded further improvements:

Model	Mean Rank	Accuracy
FinBERT (no negative sampling)	1.117	0.928
SFinBERT_neg (w/ FIBO sampling)	1.086	0.947
SFinHyp_neg (FinISH + FIBO sampling)	1.072	0.952
Ensemble (A+B w/ FIBO)	1.053	0.967

Notably, the addition of FIBO-based hierarchical negatives led to an accuracy improvement of approximately 0.019 for FinBERT and 0.035 for the ensemble compared to non-FIBO approaches.

6. Preprocessing, Data Augmentation, and Use Cases

Effective integration of FIBO into supervised machine learning pipelines often necessitates preprocessing steps, including lowercasing, whitespace normalization, and token overlap thresholding to align ontology labels with downstream hypernym targets (Ghosh et al., 2023).

In the context of task design, FIBO has been used both as a source of annotated training data (over 11,800 term-definition pairs in (Ghosh et al., 2023)) and as a constraint on candidate output labels, ensuring that only valid ontology hypernyms are considered during classification or ranking.

End-to-end examples, such as mapping "Government Bond Index Linked" to "Equity Index" via graph traversal and then confirming its semantic alignment via word embeddings and classifier ranking, illustrate the integration of symbolic and distributional signals (Perdih et al., 2021). For deep similarity tasks, embedding FIBO definitions using domain-tuned LLMs enables the correct ranking of hypernyms such as labeling "callable bond" with "Bonds" (Ghosh et al., 2023).

7. Significance and Limitations

FIBO provides a robust, machine-readable backbone for semantic reasoning in finance, supporting both direct ancestry-based classification and sophisticated negative sampling for neural semantic models. Its integration yields measurable improvements in concept classification and semantic similarity ranking, particularly as demonstrated in competitive shared tasks like FinSim-2 and FinSim-3.

A limitation is that the effectiveness of FIBO in deep learning models derives chiefly from its use in negative sampling and label structure, rather than as a direct regularization or constraint at inference time. This suggests that hierarchical information is most efficiently exploited during representation learning, rather than in final prediction. In practice, only those portions of FIBO aligning with controlled label sets are utilized, and there is no reported precision or recall decomposition isolating the impact of FIBO versus other training corpora (Ghosh et al., 2023).

FIBO’s adoption in academic and applied settings reflects its utility for taxonomy-driven concept classification, semantic disambiguation, and as a benchmarking resource for financial domain language understanding.