PaperlayerView exact revision on arXiv

Source-linked AI summary

ITL: Interpretable Document Alignment with Structured Reference Frameworks

Raúl Giráldez, Dayrelis Mena, Jesús S. Aguilar--Ruiz

arXiv:2608.27031v1cs.CL

TL;DR

Assessing document alignment with structured frameworks is difficult when evidence is distributed across heterogeneous, ambiguous texts and existing approaches provide limited traceable concept-level measures. ITL builds weighted concept profiles and textual-unit–concept affinities; in an SDG consistency study, every descriptor matched its corresponding concept most strongly, with off-diagonal affinity two orders of magnitude lower than reference affinity.

  • Problem

    Assessing alignment with structured reference frameworks is difficult because documentation is heterogeneous, often unstructured, semantically ambiguous, and costly to evaluate manually and reproducibly.

  • Method

    ITL induces weighted, concept-specific terminological profiles from an SRD and computes interpretable affinity values for every textual-unit–concept pair.

  • Results

    The SDG consistency assessment found that every descriptor received its highest affinity for the corresponding concept, while mean off-diagonal affinity was two orders of magnitude lower than mean reference affinity.

  • Takeaways & Limitations

    ITL distinguishes the conceptual profiles defined by an SRD while keeping alignment results interpretable and traceable to supporting terminology.

  • Takeaways & Limitations

    The evaluation uses descriptors identical to those defining the SRD, so empirical generalization across unseen wording, domains, and languages remains unestablished.

Abstract

from arXiv · show

Measuring alignment between documents and structured reference frameworks requires identifying conceptual evidence distributed throughout the text and reporting it through measures that are quantitative, interpretable, and traceable. Many commonly used retrieval and classification approaches return either pairwise similarity scores or one or more class labels, whereas fewer methods provide concept-level scores that are directly traceable to the terminological evidence supporting them. We present \emph{Intelligent Target Locator} (ITL), a domain-agnostic and language-portable methodology that estimates the affinity between the textual units of a target document and the concepts defined in a \emph{Structured Reference Document} ($SRD$). From the $SRD$, ITL induces concept-specific terminological profiles built from independent terms, bigrams, trigrams, and co-occurrences. Each term is assigned an importance weight that combines concept membership, term-type specificity and inter-concept discriminability. The output is a textual-unit--concept affinity matrix that can be aggregated at different levels of granularity. We conduct an internal consistency assessment using the 17 Sustainable Development Goals (SDGs), evaluating each official goal statement against the $SRD$ induced from the same set of descriptors. Every statement reached its highest affinity with the corresponding concept, and the mean affinity across the remaining concepts stayed marginal relative to the mean reference affinity. This separation indicates that ITL distinguishes the conceptual profiles of the framework. ITL thus offers a general basis for quantifying document alignment with structured frameworks while keeping each result traceable to the terminological evidence that supports it.

1 Introduction

The paper addresses the need for interpretable, traceable, and comparable measures of document alignment with structured reference frameworks, a task complicated by heterogeneous unstructured text and scattered conceptual evidence. It introduces ITL, a domain- and language-independent methodology that models framework concepts through an SRD, computes weighted affinity, and evaluates the approach using the SDGs.

  • Problem: The methodology addresses alignment as a conceptual relationship whose evidence may be distributed across a document and must remain auditable and comparable.Identifying related framework elements alone is insufficient when the relationship must also be quantified and traced to supporting evidence.
  • Method: ITL uses a Structured Reference Document as a terminological reference that represents a domain as a finite set of concepts, each described by a reference textual unit.The approach is formulated independently of domain and language.
  • Contribution: ITL provides an interpretable affinity measure that quantifies how strongly a target document shares important concepts with a Structured Reference Document.The measure is based on weighted related concepts and accounts for their importance within the reference framework.
  • Method: The paper’s two-phase methodology first induces a structured, attributed reference model and then aligns a target document against its aggregated and weighted concepts.This separates reference-model induction from document alignment.
  • Validation: The SDGs provide the consistency-study and empirical-validation setting, with the SRD instantiated from the official goal text.The study originated in the Universities for Sustainable Development project focused on the SDGs in European universities.

2 Related Work

Prior work spans retrieval, semantic representation, terminology extraction, formal framework modeling, and thematic classification, but ITL combines SRD-induced concept profiles with graded, traceable concept-level affinity. It therefore complements methods that retrieve, classify, or formalize frameworks without jointly providing this evidence-grounded alignment.

  • Related work: Related research includes lexical retrieval such as BM25, contextual and dense models including BERT [3], Sentence-BERT [4], DPR [5], and ColBERT [6], plus terminology, ontology, and classification methods.These directions respectively address document comparison, semantic matching, comparable representations, framework formalization, or category assignment.
  • Novelty: ITL preserves concept-specific term-level evidence in its scoring, unlike dense retrieval methods whose native scores generally do not decompose matches by concept.Dense representations remain complementary because they could serve as a matching component within ITL.
  • Terminology representation: Its representation extracts independent terms, bigrams, trigrams, and co-occurrences, anchoring terminology to explicit SRD concepts and weighting terms by specificity and discriminability.This contrasts with topic modeling, which induces latent topics without explicit correspondence to a predefined framework.
  • Novelty: ITL differs from automatic classification by producing graded affinity for each SRD concept instead of discrete labels or probabilities from fixed learned categories.Classification approaches commonly require annotated data and do not explicitly decompose assignments into supporting evidence.
  • Novelty: ITL computes graded affinity for every textual-unit–concept pair using terminological profiles induced directly from an arbitrary SRD, rather than fixed learned categories.The SRD supplies the reference model, while the resulting scores remain tied to the concepts defined in that document.

3 Problem Formalization

ITL formalizes document alignment as quantitative affinity between each target-document unit and each concept represented by a Structured Reference Document. Its outputs include interpretable unit-level and aggregated concept-level representations based on concept-specific terminology, with relative scores addressing SRD-dependent scale.

  • 3.1 Problem Statement: The problem is to quantify alignment between each target-document unit S_i and each SRD concept c_j, producing an interpretable textual-unit–concept affinity matrix.The matrix supports aggregation to paragraph, section, and document levels, as well as visualization such as textual-unit–concept heatmaps.
  • 3.2 Terms and Reference Term Sets: Each SRD concept receives a reference term multiset RTS_j extracted from its reference unit, preserving term multiplicities as its concept-specific terminological profile.These profiles provide the basis for evaluating target-unit affinity with each concept.
  • 3.3 Term Typology: The term representation uniformly incorporates independent terms, consecutive bigrams, trigrams, and bounded-window co-occurrences to capture isolated and compound linguistic evidence.Co-occurrences may include both directions of an ordered non-adjacent token pair.
  • 3.4 Discriminability, Specificity, Relevance, and Importance: Co-occurrence pattern spaces are quantified from observed distinct co-occurrences because they depend on both the proximity window and the text’s term arrangement.This differs from bigram and trigram pattern spaces, which are determined by the number of tokens.
  • 3.6 Absolute Affinity, Reference Affinity, and Relative Affinity: The affinity matrix and document-level vector have SRD-dependent magnitudes, so the formulation provides absolute and relative affinity vectors rather than relying on universal reference values.The dependence arises from terminological richness, term distribution across concepts, and the aggregation operator.
  • 3.7 Model Output and Intermediate Elements: The formal model output is the textual-unit–concept affinity matrix together with aggregated absolute and relative concept-affinity vectors, providing a quantitative and interpretable representation of document alignment.These outputs are illustrated in Figure 1, with complete notation collected in Appendix B.

4 Method

ITL is a reproducible, domain- and language-independent methodology that preserves traceability from document–concept affinity values to their supporting terminology. It induces a reusable reference model from a structured reference document and applies it to segmented target-document units through shared linguistic processing.

  • Method overview: ITL supports domain and language independence, interpretability through terminological traceability, and application to unstructured documents.Its design permits an SRD to be instantiated from any concept-structured framework, while retaining the evidence underlying each affinity value.
  • Method overview: The methodology comprises reference-model induction from the SRD followed by document alignment against the induced model.The reference model is computed once for an SRD and reused across target documents; alignment produces the affinity matrix and document-level aggregations.
  • Shared processing: Both phases share tokenization, tagging, filtering, stemming, and term-typology extraction, but induction processes concept descriptors whereas alignment processes target-document textual units.Figure 2 summarizes this common processing core and the overall ITL flow.
  • Input representation: The SRD associates each concept with a descriptive unit, while the target document is segmented into textual units, with text extraction preceding segmentation for unstructured inputs.These representations provide the inputs to induction and alignment, respectively.
  • Reference-model induction: The induction pipeline normalizes descriptor text, retains concept-bearing content, and extracts independent terms, bigrams, trigrams, and bounded-window co-occurrences.Independent terms preserve stem and part-of-speech identity, whereas compound terms are formed from concatenated stems; extracted frequencies are organized in a concept matrix for efficient weighting.

5 Consistency Study: Sustainable Development Goals

The consistency study instantiates ITL with the 17 Sustainable Development Goals, using their official statements as both SRD descriptors and evaluation units. Each statement most strongly matches its corresponding goal, with cross-affinities remaining marginal, while the controlled design limits conclusions about generalization and aggregation.

  • Reference instantiation: Each of the 17 official SDG statements serves both as a concept descriptor in the SRD and as an evaluation unit with a known dominant concept.This controlled configuration evaluates textual-unit affinities against the framework induced from the same statements.
  • Reference-model induction: Term importance combines profile membership, term-type specificity, and inter-goal discriminability, giving rarer terms higher discriminability and compound patterns higher specificity.A term reaches high importance when it is specific to a goal and discriminates that goal from others.
  • Alignment results: 0.18–0.38 diagonal affinities and a maximum cross-affinity of 0.020 show that every statement most strongly matches its corresponding goal, with mean off-diagonal affinity two orders lower.The largest cross-affinity remains below the lowest reference affinity of 0.18.
  • Alignment results: Differences in reference affinity reflect terminological richness: longer descriptors and more specific compound patterns generally produce higher affinities than shorter statements.These reference affinities serve as interpretive anchors for normalizing alignment with external target documents.
  • Scope and limitations: The study assesses internal consistency only because statements reproduce the descriptors used to induce the model, so it does not test unseen wording, external documents, or competing methods.External heterogeneous documents, systematic comparisons, and cases where units align with multiple concepts remain future work.
  • Scope and limitations: Affinity is evaluated per textual unit rather than aggregated at document level because the 17 statements form heterogeneous units, not a cohesive document.Consequently, absolute and relative affinity vectors are not computed in this consistency study.

6 Conclusions

ITL provides an interpretable, traceable affinity metric for quantifying alignment between target documents and structured reference documents. Its internal SDG assessment separated corresponding concepts, while cross-domain and cross-language generalization remains to be established.

  • Contributions: ITL formalizes a graded affinity metric for each textual-unit–concept pair, quantifying alignment between a target document and a Structured Reference Document.The methodology induces concept-specific terminological profiles directly from the reference document.
  • Method: Term importance combines profile membership, term-type specificity, and inter-concept discriminability, producing a traceable textual-unit–concept affinity matrix.Concept self-affinity supplies an interpretive anchor, and the values can be aggregated at different levels.
  • Internal consistency assessment: Two orders of magnitude lower mean off-diagonal affinity than mean reference affinity accompanied perfect corresponding-concept ranking across the 17 SDG descriptors.Each descriptor received its highest affinity with its corresponding concept; this assesses separation of induced profiles rather than generalization to unseen text.
  • Scope and limitations: ITL is portable in principle to normative, strategic, regulatory, or programmatic frameworks and different languages, but empirical cross-domain and cross-language generalization remains unestablished.Instantiation in other settings requires the corresponding linguistic-processing resources.
  • Applicability: Domain and language independence, together with traceable affinity values, supports explaining alignment across heterogeneous documentation against explicit reference frameworks.The stated applicability concerns scenarios requiring both quantification and explanation of alignment.
  • Future work: Future work will validate ITL on external documents, additional domains, and languages; test document-level aggregation strategies; and compare it with other alignment approaches.The planned comparisons include semantic-representation and alignment methods.

A The 17 Sustainable Development Goals

The United Nations 2030 Agenda for Sustainable Development is organized into 17 goals, whose official statements are used verbatim as concept descriptors in the reference instantiation.

  • The 2030 Agenda for Sustainable Development is organized into 17 goals.
  • Each goal’s official statement serves verbatim as the corresponding concept descriptor R_j in the reference instantiation.
  • The listed goals address themes including poverty, hunger and food security, health, education, gender equality, and water and sanitation.

B Notation Glossary

The glossary defines the notation used to formalize document alignment, covering structured reference documents, concepts, textual units, terms, and terminological profiles. It also indicates where applicable the equation or section introducing each element.

  • Documents and Textual Units: The notation centers on the SRD, its finite concept set C, concept-specific reference units R_j, and target document D.SRD denotes the Structured Reference Document encoding the target semantic domain; C = {c_1, ..., c_m} contains m concepts, each described by R_j.
  • Terms, Vocabularies, and Terminological Profiles: Terms are lexical units extracted from textual units, including unigrams, n-grams, and co-occurrences, with frequency and vocabulary notation distinguishing multiplicity from distinct terms.T(X) is a multiset of extracted terms, f(t | X) gives term frequency, and V(X) contains distinct terms.
  • Terms, Vocabularies, and Terminological Profiles: Concept-specific terminological profiles are represented by RTS_j = T(R_j), collected in TP, with V_SRD and type-specific vocabularies defining the induced term universe and term categories.V_IND(R_j) and V_COC(R_j) respectively capture distinct independent terms and distinct co-occurrences observed for concept c_j.

Term Typology.

ITL represents each preprocessed textual unit through four term types: independent terms, bigrams, trigrams, and bounded co-occurrences. Independent-term identity combines a stem with a part-of-speech tag, while co-occurrences permit limited intervening tokens.

  • Term Typology.: ITL extracts independent terms, consecutive bigrams, consecutive trigrams, and bounded co-occurrences from normalized token sequences.Co-occurrences consist of two non-consecutive tokens separated by at most Nsw intermediate tokens.
  • Term Typology.: Each textual unit is an ordered sequence of normalized tokens obtained through filtering and linguistic normalization.The sequence length LX counts the remaining tokens, and multiset union preserves term multiplicities.
  • Term Typology.: Independent-term identity is defined by the pair of a token’s stem and part-of-speech tag.Term types are represented by κ(t) ∈ {IND, 2G, 3G, COC}, with specificity normalized over this set.
  • Term Typology.: Term typology feeds ITL’s reference-model weighting, where relevance, type-dependent specificity, and cross-concept discriminability determine term importance.The importance φ(t | RTSj) is defined as the product of intra-concept relevance and inter-concept discriminability.
Loading 2608.27031v1…