Content Information Units (CIUs) Overview

• Content Information Units (CIUs) are task-grounded units that break discourse and images into small, semantically interpretable elements for clinical and analytical applications.
• CIUs are operationalized using fixed ontologies or contextual models like BERT pipelines, mapping visual and textual data to predefined or emergent information units.
• Extraction methods span rule-based, statistical, and LLM-driven approaches, enabling quantitative assessments in cognitive impairment, summarization, and image segmentation research.

Searching arXiv for the specified paper and closely related work on CIUs/SCUs to ground the article in current literature. Content Information Units (CIUs) are task-grounded units of meaning used to decompose discourse or perceptual content into smaller, clinically or analytically interpretable elements. In cognitive-linguistic picture description, CIUs are often operationalized as predefined picture-relevant entities, actions, or higher-level situation descriptors whose sequence can be used to reconstruct a speaker’s visual narrative path through an image (Ng et al., 30 Sep 2025). In aphasic discourse analysis, the closely related established term is often Correct Information Units, referring to words that are accurate, relevant, and informative about the topic (Pittman et al., 12 Nov 2025). In summarization research, an analogous construct appears as Summary Content Units (SCUs), concise factual units used to evaluate or rank summaries by semantic content rather than surface overlap (Nawrath et al., 2024). Across these domains, the common function of CIU-like units is to transform global discourse or image descriptions into smaller information-bearing components that support measurement, comparison, and automation.

1. Conceptual scope and terminological variants

The meaning of CIU depends on domain and annotation tradition. In the Cookie Theft picture-description literature, CIUs are discrete, clinically meaningful picture elements and actions tied to a fixed ontology for a single image (Ng et al., 30 Sep 2025). In the automated spatio-semantic graph work for Cookie Theft, the ontology contains 23 predefined CIUs covering people, places, objects, actions, and higher-level situation descriptors such as boy, girl, woman, kitchen, outside, cookie, jar, stool, sink, plate, dishcloth, water, window, cupboard, dishes, curtains, boy taking/stealing, boy or stool falling, woman drying/washing plates, water overflowing, action performed by girl, woman unconcerned by overflowing, and woman indifferent to the children (Ng et al., 30 Sep 2025). In this formulation, CIUs are not discovered from unrestricted language; they are a fixed, picture-grounded inventory.

In aphasia research, the term used is typically Correct Information Units, following Nicholas and Brookshire. There, CIUs are counted at the word level after transcription of connected speech, with coders identifying which intelligible words are accurate, relevant, and informative about the topic, while excluding fillers, repetitions, and tangential remarks (Pittman et al., 12 Nov 2025). This makes CIUs a discourse-level informativeness measure rather than a picture-grounded ontology.

In summarization, the closest parallel is the SCU. SCUs are described as concise human-written sentences or phrases that describe a single fact and decompose summaries into semantically interpretable factual units (Nawrath et al., 2024). A plausible implication is that CIUs, Correct Information Units, and SCUs are best understood as a family of related information-unit formalisms rather than a single universal standard. They share the same general function—factoring content into smaller meaningful units—but differ in granularity, annotation target, and whether the units are manually authored, lexically triggered, or model-induced.

2. CIUs in clinical picture description and the visual narrative path

The most developed recent CIU framework in cognitive impairment assessment treats Cookie Theft descriptions as ordered traversals over picture-grounded content units (Ng et al., 30 Sep 2025). The underlying idea is that not only what is described, but also where those concepts lie in the image and in what order they are verbalized, can reflect visuospatial processing, attentional allocation, and organizational abilities. Earlier work represented these descriptions as spatio-semantic graphs whose nodes were CIUs and whose edges captured transitions in narrative order, enriched by each CIU’s picture location (Ng et al., 2 Feb 2025).

In this setting, each predicted or manually annotated CIU is linked to a known visual referent in the Cookie Theft image. The paper on automated extraction and ordering states that each CIU has an X-Y coordinate and a quadrant assignment, even though explicit coordinates are not printed in the text (Ng et al., 30 Sep 2025). This enables transcript content to be projected back into picture space, turning a verbal description into a path through the image.

That projection supports a set of spatio-semantic features designed to characterize the recovered visual narrative path. The features reported are Avg. X, Std. X, Avg. Y, Std. Y, Total path distance, Unique nodes, Total path / Unique nodes, Nodes, Self cycles, Cycles, Self cycles (quadrants), and Cross ratio (quadrants) (Ng et al., 30 Sep 2025). These features quantify spatial spread, traversal distance, repetition, and inter-quadrant versus intra-quadrant movement. In clinical interpretation, longer paths, more repeated nodes, and more local cycling can indicate a less organized visual-semantic path, while broader and more efficient traversal suggests more coherent organization.

The rule-based predecessor to the BERT-based system implemented this same basic idea through dictionary matching, lemmatization, and a manually constructed lexical inventory tied to the 23 Cookie Theft CIUs (Ng et al., 2 Feb 2025). That pipeline mapped transcript words to CIU labels and then converted the resulting sequence into spatio-semantic graphs using fixed image coordinates. Its main contribution was not direct label-level accuracy evaluation, but the observation that automated CIU sequences could still yield graph features that differentiated cognitively impaired from unimpaired speakers (Ng et al., 2 Feb 2025).

3. Operationalization and extraction methodologies

CIU extraction methods vary substantially in what counts as a unit and how the mapping from raw input to units is performed. In the Cookie Theft BERT pipeline, the task is formulated as contextualized sentence-level multi-label classification over the fixed ontology of 23 CIUs (Ng et al., 30 Sep 2025). The architecture is described as follows: the input text is processed through a BERT model to generate contextual embeddings, those embeddings are aggregated via mean pooling, and a linear classification layer produces logits across the 23 predefined CIU classes. Predicted CIUs are those with sigmoid probability above 50\%, and within each sentence they are ordered by descending logit value to construct temporal order (Ng et al., 30 Sep 2025). This system does not tag spans; it predicts sentence-level presence for each class.

The training objective combines binary cross-entropy over the 23 labels with an auxiliary pairwise ranking loss that enforces higher logits for earlier CIUs in the ground-truth sequence, using margin $m=1$ and weight $\lambda = 0.1$ in the combined objective (Ng et al., 30 Sep 2025). The model uses bert-base-uncased, hidden size 768, 12 transformer layers, and dropout 0.2. Fine-tuning ran for 50 epochs with AdamW, using learning rate $2\times10^{-5}$ for BERT parameters and $10^{-3}$ for the classifier, on 2,783 descriptions from 1,352 unique speakers, with 5-fold cross-validation and external validation on 256 transcripts from 235 unique speakers (Ng et al., 30 Sep 2025).

By contrast, the earlier Cookie Theft automation work is rule-based. It removes punctuation, lemmatizes transcripts using spaCy, and matches words against CIU-specific dictionaries built from word-frequency histograms across the datasets (Ng et al., 2 Feb 2025). This approach supports only the synonyms and paraphrases manually included in the lexicon. Pronouns and broad lexical triggers are included directly in the dictionaries, but there is no explicit coreference resolution, learned disambiguation, or probabilistic scoring (Ng et al., 2 Feb 2025).

In aphasic discourse analysis, the task is token-level rather than sentence-level. The study on practical machine learning for aphasic discourse uses cleaned CHAT transcripts from the Cat Rescue picture-description task, tokenizes them with spaCy, and labels tokens on two dimensions: WORD vs. NOT WORD and, conditional on WORD, CIU vs. NOT CIU (Pittman et al., 12 Nov 2025). The five models evaluated are SVM-linear, SVM with RBF kernel, Decision Tree, k-Nearest Neighbor, and Random Forest, using token-level character n-grams, local context, and handcrafted linguistic features (Pittman et al., 12 Nov 2025). This formulation makes CIU detection a local token classification problem rather than a fixed-ontology picture grounding problem.

In summarization, SCU approximations are constructed differently again. Human SCUs remain the gold standard, but approximations include STUs based on semantic role triplets, SMUs derived from AMR subgraphs, SGUs generated by LLMs, and coarse sentence splits used as a baseline (Nawrath et al., 2024). The more recent SCURank framework operationalizes SCUs as LLM-extracted “small sentences” representing subsentences of candidate summaries, then clusters semantically equivalent units with all-mpnet-base-v2 embeddings and HDBSCAN (Wang et al., 21 Apr 2026). This suggests a broader methodological pattern: CIU-like units can be manually defined, lexicon-triggered, classifier-predicted, parser-derived, or LLM-induced, depending on task structure.

4. Quantitative performance and downstream validity

The strongest direct CIU extraction results currently reported for Cookie Theft come from the BERT-based spatio-semantic pipeline. Under 5-fold cross-validation, the abstract reports 93\% median precision, 96\% median recall, and 24\% sequence error rates for CIU detection and ordering (Ng et al., 30 Sep 2025). Sequence quality is measured with a Levenshtein-based sequence error rate decomposed into insertions, deletions, and substitutions, with reported speaker-level components of approximately 11\% insertions, 10\% deletions, and 6\% substitutions, yielding the overall 24\% sequence error rate (Ng et al., 30 Sep 2025).

Per-CIU detection performance is high for many concrete object and action categories. Reported precision/recall pairs include window: 96.4 / 99.2, jar: 96.5 / 96.9, stool: 96.0 / 97.7, water: 96.4 / 98.3, and boy: 95.0 / 98.2 (Ng et al., 30 Sep 2025). Performance is weaker for semantically complex or inferential classes such as boy taking/stealing: 75.7 / 80.1, woman unconcerned by overflowing: 66.4 / 74.6, and woman indifferent to the children: 63.8 / 66.1 (Ng et al., 30 Sep 2025). The paper attributes these difficulties to semantic complexity, broader contextual dependence, and lower training frequency.

External validation on W-ADRC compares spatio-semantic features derived from BERT-predicted CIUs against those derived from the prior dictionary baseline, using Pearson correlation with manually derived ground truth. All reported correlations are significant at $p<0.05$, and the BERT pipeline improves correlation on every reported feature, for example Avg. X: 0.95 vs 0.80, Std. X: 0.90 vs 0.61, Total path distance: 0.97 vs 0.85, Self cycles: 0.88 vs 0.62, and Cross ratio (quadrants): 0.64 vs 0.31 (Ng et al., 30 Sep 2025). This is important because it shows that better CIU extraction propagates to better downstream spatial and graph features, not merely better local label metrics.

Clinical validity was evaluated with ANCOVA on combined WRAP and Pitt data, comparing 1062 cognitively unimpaired speakers against an impaired group of 24 MCI and 189 dementia speakers, with age, gender, education, and unique nodes as covariates when applicable (Ng et al., 30 Sep 2025). The broad pattern reported is that BERT-derived features are “closely aligned” with those from manual annotation, while the dictionary baseline often produces larger $F$-values that likely reflect systematic over-tagging rather than better construct fidelity (Ng et al., 30 Sep 2025).

Aphasic discourse CIU detection remains harder. In the Cat Rescue study, all models achieved near-ceiling performance on WORD vs. NON-WORD classification, with 0.995 accuracy across models, but CIU vs. NON-CIU performance was substantially lower and more variable (Pittman et al., 12 Nov 2025). The best CIU classifier was k-NN with 0.824 accuracy, 0.889 F1, and 0.787 AUC, while SVM-rbf achieved the highest AUC at 0.797 (Pittman et al., 12 Nov 2025). The contrast between the easy lexicality task and the harder informativeness task is central: CIU status is not a shallow lexical property.

In summarization evaluation, gold human SCUs still outperform automatic approximations. However, SGUs provide the best approximation quality among automatic alternatives, while STUs and SMUs remain competitive (Nawrath et al., 2024). In SCURank, automatically extracted SCUs from candidate summaries improve ranking-based distillation and outperform traditional metrics and LLM-based ranking methods across several reported datasets and evaluation measures (Wang et al., 21 Apr 2026). A plausible implication is that content-unit-based representation remains valuable even when the units are approximate, provided the approximation is semantically faithful enough for the downstream objective.

5. Information-theoretic and segmentation perspectives on content units

Not all work relevant to CIUs is clinical or discourse-specific. In image segmentation, Uniform Information Segmentation proposes replacing uniform-size superpixels with connected regions whose information content is bounded by a threshold $T$ (Achanta et al., 2016). The segmentation problem is formulated as minimizing the number of connected segments while enforcing $I(A_k)\le T$ for each segment, where information is defined through Shannon self-information under a chosen probabilistic model (Achanta et al., 2016). These segments, called adaptels, become smaller in complex regions and larger in homogeneous ones. Although this work is not semantic in the clinical-discourse sense, it provides a formal model of content-adaptive unitization: segments are defined by bounded information rather than fixed geometry.

The adaptel framework is relevant because it formalizes a shift from uniform size to uniform information budget. That is close to a CIU interpretation in which units should track content density rather than area. The method uses the multivariate double exponential distribution, a single information threshold $T$ as the only hyperparameter, and a greedy growth algorithm with reported linear complexity in the number of pixels (Achanta et al., 2016). The authors explicitly describe the output as smaller and denser segments in areas of high complexity and larger segments in homogeneous regions. This suggests a model of CIUs as bounded-information units, though the paper itself concerns low-level image statistics rather than semantics.

A different information-theoretic perspective appears in the multi-scale Shannon-based pattern analysis work. That paper argues that ordinary symbol-level Shannon information can overestimate the internal information content of structured patterns, especially when repetition occurs at higher scales (Pocze, 2023). It defines a scale-dependent spectrum by partitioning a discrete pattern into non-overlapping blocks of length $r$, computing Shannon information on block sequences, normalizing against a scale-specific maximum, and taking the minimum across scales to obtain $\lambda = 0.1$0 (Pocze, 2023). The method therefore treats information-bearing units as potentially larger than atomic symbols.

This multi-scale view is not a semantic CIU theory, but it is directly relevant to the notion that the “right” information unit may be scale-dependent. The paper’s examples show that periodic or block-repetitive patterns can have much lower effective information than symbol-level entropy suggests, because the meaningful recurring units are larger repeated blocks (Pocze, 2023). A plausible implication is that CIUs, particularly in structured media, should often be treated as emergent units whose granularity depends on recurrent organization rather than being fixed in advance.

6. Limitations, controversies, and future directions

A persistent limitation across CIU research is domain specificity. The Cookie Theft BERT system is built around 23 predefined CIUs for a single image, making it highly interpretable for that task but not directly reusable for another elicitation picture without a new ontology, spatial mapping, and annotated training data (Ng et al., 30 Sep 2025). The earlier rule-based approach shares the same dependence on picture-specific lexicons and fixed spatial referents (Ng et al., 2 Feb 2025). This constrains generalizability even when within-task performance is strong.

Another limitation concerns semantic complexity. The classes most difficult for the BERT system are precisely the inferential and pragmatic ones, such as woman unconcerned by overflowing and woman indifferent to the children (Ng et al., 30 Sep 2025). In aphasic discourse, token-level CIU detection is difficult because informativeness depends on context, pragmatic relevance, stimulus grounding, and discourse structure, none of which are well captured by shallow local-feature models (Pittman et al., 12 Nov 2025). The Cat Rescue study explicitly notes that current models do not encode discourse structure, topic flow, coreference, prosody, acoustic cues, or visual grounding (Pittman et al., 12 Nov 2025).

There is also a methodological controversy about what level of agreement matters most. The 2025 dictionary-based Cookie Theft study evaluates CIU automation primarily through downstream group differences in graph features, not direct extraction metrics such as precision, recall, or kappa (Ng et al., 2 Feb 2025). By contrast, the 2025 BERT study adds direct extraction and ordering evaluation, then shows superior downstream fidelity as well (Ng et al., 30 Sep 2025). This suggests that downstream clinical usefulness and label-level agreement are related but distinct validation targets.

In summarization, a similar issue appears in the gap between human SCUs and automatic approximations. Human SCUs remain best for difficult summary-level ranking of short summaries, even though SGUs are judged high quality and often competitive (Nawrath et al., 2024). SCURank shows that automatically extracted SCUs can be effective for candidate ranking and distillation, but its units are candidate-derived and consensus-weighted rather than source-grounded or manually verified (Wang et al., 21 Apr 2026). If all candidates omit an important fact, the framework cannot recover it. If a hallucinated fact appears in multiple candidates, consensus can make it appear important (Wang et al., 21 Apr 2026).

Future directions reported in the CIU literature include applying spatio-semantic features to other neurodegenerative disorders for broader clinical generalizability (Ng et al., 30 Sep 2025), testing CIU automation on novel transcripts and additional discourse genres (Pittman et al., 12 Nov 2025), and exploring LLMs, neurosymbolic, and knowledge-graph approaches for richer CIU identification in aphasic discourse (Pittman et al., 12 Nov 2025). More broadly, the convergence of picture-grounded CIUs, discourse-level Correct Information Units, and summary-level SCUs suggests an emerging research program centered on content-unit representations that are interpretable, automatable, and semantically faithful. What differs across domains is not the need for content units, but the ontology, granularity, and operational mechanism by which those units are defined.