ITL: Interpretable Document Alignment with Structured Reference Frameworks

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

cs.CLAug 27, 2026

ITL: Interpretable Document Alignment with Structured Reference Frameworks

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

Organizations: School of Engineering Pablo de Olavide University ES-41013 Seville, Spain

Abstract

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 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 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.

Figures & tables

Panel of Figure 1 — Figure 1: Schematic representation of the textual-unit–concept affinity matrix $\Lambda(D,SRD)$ . Rows correspond to the textual units $S_{i}\in U(D)$ and columns to the concepts $c_{j}\in SRD$ , each described by its reference textual unit $R_{j}$ and characterized by its terminological profile $RTS_{j}$ . The aggregation operator $\mathcal{A}(\cdot)$ reduces each column of $\Lambda(D,SRD)$ to the document-level affinity $\lambda_{j}(D,SRD)$ , jointly forming the vector $\lambda(D,SRD)$ .

Panel of Figure 2 — Figure 2: ITL Architecture.

Term	Type
reduc	IND - VB
inequ	IND - NN
countri	IND - NNS
inequ countri	2G
reduc inequ	2G
reduc inequ countri	3G

Table 1: Reference Term Set for G10: "Reduce inequality within and among countries".

Element	Choice	Version / Reference
Language	English	—
Number of concepts ( $m$ )	17 (goals)	—
Runtime	Python	3.9
NLP toolkit	NLTK	3.9.1 Bird et al. (2009)
Sentence segmentation	Punkt tokenizer	Kiss and Strunk (2006)
Tokenization	Treebank word tokenizer	Marcus et al. (1993)

Table 2: Parameters and components of the reference instantiation on the SDGs.

		Frequency
Term	Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17	Discriminability
promot inclus	COC				1				1								1		0.612239
sustain	IND - JJ		1				1	1	2	1		1	1		1	1	1	1	0.163710
ensur	IND - VB			1	1		1	1					1						0.431939

Table 3: Frequency and discriminability of terms

Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
IND	2	9	5	9	6	6	7	8	9	3	7	5	6	8	17	14	8
2G	1	8	4	8	5	5	6	8	8	2	6	4	5	7	17	14	7
3G	0	7	3	7	4	4	5	7	7	1	5	3	4	6	16	13	6
COC	1	52	16	52	25	25	34	40	52	4	34	16	25	33	107	104	43

Table 4: Frequencies of terms by type and goal

Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
IND	0.500000	0.049724	0.121951	0.049724	0.090909	0.090909	0.072165	0.060606	0.049724	0.272727	0.072165	0.121951	0.090909	0.064000	0.018085	0.024433	0.059259
2G	0.250000	0.198895	0.243902	0.198895	0.227273	0.227273	0.216495	0.212121	0.198895	0.272727	0.216495	0.243902	0.227273	0.224000	0.144681	0.158813	0.207407
3G	0.000000	0.464088	0.243902	0.464088	0.303030	0.303030	0.360825	0.424242	0.464088	0.090909	0.360825	0.243902	0.303030	0.448000	0.723404	0.635253	0.414815
COC	0.250000	0.287293	0.390244	0.287293	0.378788	0.378788	0.350515	0.303030	0.287293	0.363636	0.350515	0.390244	0.378788	0.264000	0.113830	0.181501	0.318519

Table 5: Term-type specificity weights by Sustainable Development Goal.

Term	Type	G2	G3	G4	G6	G7	G8	G9	G11	G12	G14	G15	G16	G17
promot inclus	COC	0	0	0.175891979	0	0	0.185526784	0	0	0	0	0	0.111121991	0
sustain	IND - JJ	0.008140316	0	0	0.014882712	0.011814132	0.009921808	0.008140316	0.011814132	0.019964598	0.010477440	0.002960695	0.003999926	0.009701291
ensur	IND - VB	0	0.052675393	0.021477735	0.039267143	0.031170878	0	0	0	0.052675393	0	0	0	0

Table 6: Importance of terms distributed among goals.

	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
G1	$0.375$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$
G2	$0$	$0.261\,899$	$0.000\,503$	$0.000\,205$	$0.000\,904$	$0.000\,196$	$0.000\,155$	$0.003\,895$	$0.002\,024$	$0$	$0.000\,155$	$0.000\,263$	$0$	$0.000\,138$	$0.001\,957$	$0.001\,235$	$0.000\,128$
G3	$0$	$0.000\,556$	$0.293\,464$	$0.016\,824$	$0$	$0.001\,402$	$0.001\,113$	$0.000\,678$	$0.000\,556$	$0$	$0$	$0.001\,881$	$0$	$0$	$0.000\,202$	$0.000\,273$	$0$
G4	$0$	$0.000\,205$	$0.008\,953$	$0.259\,792$	$0$	$0.000\,517$	$0.000\,410$	$0.005\,085$	$0.002\,429$	$0$	$0.000\,426$	$0.000\,693$	$0$	$0$	$0.000\,075$	$0.002\,927$	$0$
G5	$0$	$0.000\,939$	$0$	$0$	$0.308\,535$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$	$0$
G6	$0$	$0.000\,204$	$0.001\,317$	$0.000\,537$	$0$	$0.296\,393$	$0.013\,059$	$0.000\,248$	$0.000\,204$	$0$	$0.000\,295$	$0.007\,789$	$0$	$0.000\,262$	$0.005\,013$	$0.000\,100$	$0.000\,243$

Table 7: Affinity matrix

Semantic Alignment

See all 15 papers

A Unifying Framework for Concept-Based Representational Similarity
Jun 8, 2026Grégoire Dhimoïla, Agustin Martin Picard, Victor Boutin +2Representational Similarity Analysis Cross-Modal Alignment
Brown University · ENS Paris Saclay · DEEL - IRT Saint Exup´ery +2
Scaling Unsupervised Word Alignment to Documents via Structural Constraints
Aug 21, 2026Michelle Wastl, Jannis Vamvas, Rico Sennrich
Department of Computational Linguistics, University of Zurich
MATCHA: Matching Text via Contrastive Semantic Alignment
May 26, 2026Siran Li, Ece Sena Etoglu, Carsten Eickhoff +1LLM Evaluation Semantic Textual Similarity
University of Tübingen

Jun 8, 2026J/K move · Enter open · S save

A Unifying Framework for Concept-Based Representational Similarity

Grégoire Dhimoïla, Agustin Martin Picard, Victor Boutin, Thomas Fel+1

Brown University · ENS Paris Saclay · DEEL - IRT Saint Exup´ery+2

Representational Similarity Analysis Cross-Modal Alignment Concept Alignment Sparse Autoencoders

Open paper AskPDF arXiv

Learned representations across models and modalities often exhibit striking structural similarities, suggesting shared underlying concept decompositions. However, concept alignment remains poorly defined: existing approaches optimize different objectives under the same terminology, obscuring what is actually aligned. We propose a unifying framework that decomposes alignment along two axes: what is aligned (representations vs. concepts) and at what level (instance-wise vs. distributional). This induces four corresponding properties---instance-wise and distributional variants of translation and concept consistency---and reveals precisely which of these guarantees existing methods provide. We further introduce InterVenchA, an intervention-based benchmark that separately measures extraction quality, translation quality, and concept consistency---which we validate against human perception of concept alignment. Using this framework, we show that commonly assumed equivalences between alignment objectives fail in practice: optimizing one property does not reliably recover the others, and, on real embeddings, purely unsupervised objectives fail to recover meaningful instance-level alignment. As a constructive consequence, a model, the Coupled SAE (CoSAE), matches or exceeds specialized alignment methods such as crosscoders and Universal SAEs on their own target metrics, by jointly enforcing the properties identified by our framework without any architectural innovation. Moreover, as little as 1% paired data is sufficient to recover instance-level alignment when anchoring distributional objectives. Overall, concept alignment is fundamentally multi-objective, and progress requires treating it as such. Our framework and benchmark provide the means to define, measure, and compare concept alignment methods on this basis.

Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
IND	2	9	5	9	6	6	7	8	9	3	7	5	6	8	17	14	8
2G	1	8	4	8	5	5	6	8	8	2	6	4	5	7	17	14	7
3G	0	7	3	7	4	4	5	7	7	1	5	3	4	6	16	13	6
COC	1	52	16	52	25	25	34	40	52	4	34	16	25	33	107	104	43

Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
IND	2	9	5	9	6	6	7	8	9	3	7	5	6	8	17	14	8
2G	1	8	4	8	5	5	6	8	8	2	6	4	5	7	17	14	7
3G	0	7	3	7	4	4	5	7	7	1	5	3	4	6	16	13	6
COC	1	52	16	52	25	25	34	40	52	4	34	16	25	33	107	104	43

Abstract

Semantic Alignment

Explore similar work

A Unifying Framework for Concept-Based Representational Similarity

Scaling Unsupervised Word Alignment to Documents via Structural Constraints

MATCHA: Matching Text via Contrastive Semantic Alignment

Type	G1	G2	G3	G4	G5	G6	G7	G8	G9	G10	G11	G12	G13	G14	G15	G16	G17
IND	2	9	5	9	6	6	7	8	9	3	7	5	6	8	17	14	8
2G	1	8	4	8	5	5	6	8	8	2	6	4	5	7	17	14	7
3G	0	7	3	7	4	4	5	7	7	1	5	3	4	6	16	13	6
COC	1	52	16	52	25	25	34	40	52	4	34	16	25	33	107	104	43