---
title: 'CueTaxo: Hierarchical Support Taxonomy'
url: https://www.emergentmind.com/topics/cuetaxo
type: topic
---

# CueTaxo: Hierarchical Support Taxonomy

Searching arXiv for the CueTaxo paper and closely related context.
CueTaxo is a hierarchical taxonomy for the controlled elicitation of missing support information in posts from Online Mental Health Communities (OMHCs). Introduced as part of the MH-COPILOT framework, it operationalizes three support attributes—**event**, **effect**, and **requirement**—into a structured questioning policy that asks targeted follow-up questions when a post is incomplete or under-specified. In the associated system, CueTaxo sits between assessment and generation: the model first detects which support attributes are present and how clearly they are expressed, and the taxonomy then determines what kind of question should be asked, for which attribute, and in what form [2508.16788].

## 1. Conceptual definition and scope

CueTaxo is described as both “a dedicated taxonomy of support attributes” and a “custom taxonomy” that “enables a structured and interpretable approach to prompting users with targeted questions for improving support attributes” [2508.16788]. Its immediate motivation is the observation that many OMHC posts remain unanswered not necessarily because support is unavailable, but because the posts omit contextual cues that supporters need in order to respond effectively. CueTaxo is designed to convert that incompleteness into a controlled elicitation problem.

The taxonomy is organized around three support attributes. **Event** denotes what happened; **effect** denotes what impact the event had, including emotional, mental, or practical consequences; and **requirement** denotes what support the user wants. CueTaxo does not replace these attributes. Rather, it organizes and operationalizes them by mapping their degree of completeness to appropriate follow-up question types [2508.16788].

A common misconception is to read CueTaxo as a flat label inventory. The paper defines something more structured: a hierarchy over joint attribute-completeness states. Another misconception is to treat it as a standalone generation model. In MH-COPILOT it is instead the policy layer that constrains and interprets question generation. This suggests that its primary significance lies in controllability and interpretability rather than in free-form text generation.

## 2. Hierarchical organization and completeness levels

CueTaxo consists of **five levels**, each representing “varying degrees of completeness for the three support attributes within a post.” The paper defines **Level 1** as “none present,” **Level 5** as “all well-described,” and **Levels 2–4** as intermediate cases of partial expression [2508.16788]. The underlying per-attribute intensity labels are:

\[
v_i \in \{0,1,2\}
\]

with \(0\) for absent, \(1\) for moderate, and \(2\) for present.

The hierarchy uses sublabels such as **A**, **B**, and **C** to distinguish different mixed states of incompleteness. Questions are generated only for absent or moderate attributes; “level 5A is not shown” because no elicitation is needed when all relevant attributes are already well-described [2508.16788].

| CueTaxo node | Operational condition | Representative prompt family |
|---|---|---|
| Level 1A | Attribute absent with little grounding context | “Can you tell me what happened?” |
| Levels 2A–2C | Partial expression with contextual cue \(X\) | “How did X make you feel?” |
| Levels 3A–3C | Mixed incompleteness across attributes | “What happened that you want X?” |
| Level 4A | Attribute present but moderate | “Can you elaborate more on X?” |
| Level 5A | All required attributes well-described | No question generated |

At **Level 1A**, CueTaxo uses generic fallback prompts because no reliable contextual anchor is available. The paper gives three canonical forms: for **event**, “Can you tell me what happened? You can be as specific as you like.”; for **effect**, “Could you describe the specific effect the event has had on you?”; and for **requirement**, “What kind of support or help you feel would be most beneficial?” [2508.16788].

At **Levels 2A–2C**, the taxonomy begins to use a contextual placeholder \(X\), described as a user-mentioned entity. This is the point at which CueTaxo becomes explicitly relational across attributes. For example, if **event** is missing but some **requirement** content is present, the template may become “What happened that you want X?” If **effect** is absent but an event cue is available, the question may be “How did X make you feel?” The paper emphasizes this design choice as necessary for absent-information cases where generic prompts are too vague [2508.16788].

At **Levels 3A–3C**, CueTaxo handles more complex mixed states in which multiple attributes are incomplete but partially grounded. Here the same attribute may admit more than one contextualized template, such as “What made you feel X?” or “What happened that you want X?” for **event**, depending on which other attribute provides the anchor. At **Level 4A**, the target attribute is already present but only moderately expressed, so the taxonomy falls back to a simpler elaboration request: “Can you elaborate more on X?” [2508.16788].

The hierarchy therefore functions as a state-dependent template system. Its taxonomy is not over discourse topics; it is over the completeness of relational support cues.

## 3. Relationship to REDDME and the annotation framework

CueTaxo is tightly coupled to the paper’s dataset, **REDDME**, which contains **4,760 posts** from mental health subreddits annotated for the span and intensity of the three support attributes [2508.16788]. The taxonomy is not only a modeling device; it is also embedded in the dataset construction procedure. Annotators produced rationale spans, intensity labels, and taxonomy-based reference questions.

The support-attribute spans are explicitly marked. The paper lists the special tags as: **event** \(\langle es \rangle, \langle ee \rangle\), **effect** \(\langle efs \rangle, \langle efe \rangle\), and **requirement** \(\langle rs \rangle, \langle re \rangle\). These spans are later used both in prompting and in contextual grounding for the placeholder \(X\) [2508.16788].

CueTaxo-informed question annotation is central to the supervision signal. Annotators were instructed not only to identify what is missing, but also to write reference questions that encourage the user to provide the absent or underdeveloped content. The paper states that “CueTaxo assists in the selection of the ideal template question,” and gives the example “What happened that you wanted X?” where \(X\) is filled from the rationales [2508.16788]. This means the taxonomy is human-authored before it is model-generated.

The dataset analysis further motivates the taxonomy’s practical necessity. The paper reports **2125 posts with event absent or moderate**, **2781 for effect**, and **2976 for requirement** [2508.16788]. This suggests that incomplete support attributes are not marginal annotation cases but a substantial property of OMHC data.

The annotation structure also appears reliable at the attribute level. The paper reports \(\kappa = 0.885\) for event, \(\kappa = 0.886\) for effect, and \(\kappa = 1.000\) for requirement [2508.16788]. These are not direct agreement scores on CueTaxo nodes, but they support the stability of the attribute schema on which CueTaxo depends.

## 4. Role inside MH-COPILOT

In MH-COPILOT, CueTaxo is the intermediate control layer between diagnosis and question generation. The pipeline begins with **Contextual attribute-span identification (CSpan)**, framed as token-level entity extraction using RoBERTa. For a post \(P\), the input consists of tokens \(\{t_1, t_2, \dots, t_n\}\) and labels \(\{a_1, a_2, \dots, a_n\}\), where the labels correspond to event, effect, or requirement spans [2508.16788].

A second module predicts an intensity vector

\[
V = \{ v_\text{event}, v_\text{effect}, v_\text{req} \}
\]

with each

\[
v_i \in \{0,1,2\}.
\]

The paper states that this output “guides the question-generator module by providing information on how well the attributes are expressed using our taxonomy” [2508.16788]. CueTaxo is the mapping from that intensity configuration to a prompt-selection decision.

The taxonomy is then “embedded in the language model via prompting.” The prompt includes the original post, marked spans, intensity scales for each support attribute, and an output schema of the form:

\[
\{ \text{event\_question: , effect\_question: , requirement\_question: } \}.
\]

At this stage CueTaxo determines whether a question should be asked, which attribute should be targeted, and whether the system should use a generic or contextualized template [2508.16788].

A simple but important operational rule governs the entire module: “Questions are generated only for moderate or absent intensities.” A plausible implication is that CueTaxo implements a sparse intervention strategy rather than indiscriminately asking for more detail. If \(v_i = 2\), the attribute is treated as sufficiently described and typically receives no question.

After supervised fine-tuning on CueTaxo-consistent reference questions, MH-COPILOT applies reward-based refinement. The reward function is given as

\[
r = \sum_{i \in [1,2,3]} r_{Q_i}^{CC} \cdot r_{Q_i}^{CG} \cdot r_{Q_i}^{EA} \cdot r_{SA},
\]

where \(Q_1, Q_2, Q_3\) correspond to questions for event, effect, and requirement; \(r^{CC}\) is a category-correctness score; \(r^{CG}\) is contextual grounding; \(r^{EA}\) is empathy; and \(r_{SA}\) is structure adherence [2508.16788]. CueTaxo is directly implicated in \(r_{SA}\), since structure adherence is defined relative to the taxonomy template and is scored as \(1\) for correct format and \(0\) for deviations.

Preference optimization then uses DPO over candidate responses produced by the taxonomy-guided generator. The paper defines the preference dataset as

\[
\mathcal{D}=\{x^{(i)},y_p^{(i)},y_{np}^{(i)}\},
\]

with the usual preferred and non-preferred completions, and gives the corresponding DPO objective [2508.16788]. CueTaxo thus affects both prompt construction and reward shaping.

## 5. Question behavior, examples, and empirical contribution

The paper’s examples show that CueTaxo operates by asking only about the attributes that remain incomplete. In a post where **event** is moderate but **effect** and **requirement** are well-described, the reference question is: “Can you elaborate more on why you are taking adderall?” with no additional questions for the other slots [2508.16788]. In another example, where **effect** is absent and **requirement** is moderate, the reference questions become “How did doing yoga make you feel?” and “Can you elaborate more on what can help you change your bad habits?” [2508.16788].

The appendix comparison between prompting **without CueTaxo** and **with CueTaxo** clarifies the taxonomy’s function. For the post “My partner can get very angry at times and it makes me feel uncomfortable… Any help would be really appreciated,” the non-taxonomy prompt generates an **effect** question: “How did your partner's anger make you feel?” By contrast, CueTaxo redirects the follow-up to the underdeveloped **event** attribute: “Can you elaborate more on your partner's anger?” [2508.16788]. The paper’s interpretation is that effect and requirement were already adequately conveyed, so the taxonomy shifts elicitation toward the missing cue.

The ablation study provides the strongest quantitative evidence that CueTaxo is a substantive component rather than superficial prompt engineering. Full MH-COPILOT achieves **R1 89.30**, **R2 84.50**, **RL 88.88**, **BERTScore-F1 98.74**, and **METEOR 93.84**. Removing reward modeling yields only small declines, but removing CueTaxo causes a substantial drop to **R1 68.20**, **RL 65.58**, **BERTScore-F1 95.48**, and **METEOR 76.98** [2508.16788]. The paper explicitly concludes that this shows “the essence of taxonomy in the whole generative framework.”

The effect also transfers across base models. Reported ROUGE-L values improve from **69.71** to **88.31** for Mistral, from **64.14** to **85.98** for Phi-3, and from **68.40** to **86.83** for Llama-3 when CueTaxo is added [2508.16788]. This suggests that the taxonomy is not merely compensating for a single model’s idiosyncrasies.

A plausible interpretation is that CueTaxo contributes less by increasing lexical overlap than by improving semantic targeting. The paper itself warns that BLEU and ROUGE can over-reward templated similarity, but argues that the taxonomy steers the system toward “under-addressed or missing entity types” [2508.16788]. That claim is consistent with the qualitative examples.

## 6. Limitations, dependencies, and broader interpretation

CueTaxo’s main limitation is error propagation from upstream modules. The paper explicitly notes that “errors at the early phase get escalated to subsequent steps” [2508.16788]. If CSpan extracts the wrong spans, or if the intensity classifier assigns the wrong level, the taxonomy may trigger a question for an attribute that is already well-described or miss one that is actually absent.

A second limitation is the trade-off between structure and flexibility. CueTaxo gains control through templates and explicit levels, but that same regularization can make outputs more formulaic. The paper discusses this indirectly through its comments on BLEU and ROUGE and through examples where model outputs become repetitive or misdirected even when the taxonomy is present [2508.16788].

A third limitation concerns contextualized prompting under attribute absence. CueTaxo’s \(X\)-based templates are more informative than generic requests, but they rely on accurate rationales and correct cross-attribute grounding. This suggests that the taxonomy works best when span extraction and rationale identification are sufficiently precise.

The paper also notes model-specific variation. Although “Mistral+CueTaxo yields performance comparable to MH-Copilot,” qualitative analysis shows limitations “pertaining to the context and LLM’s understanding of the relation between support attributes” [2508.16788]. CueTaxo therefore cannot be understood as a complete substitute for model capacity; it constrains generation, but does not eliminate reasoning failures.

Finally, CueTaxo is specific to Reddit mental-health posts and to the event–effect–requirement schema used in REDDME. This suggests portability is possible but not automatic. The hierarchy is a domain-specific elicitation policy grounded in OMHC help-seeking behavior, not a general-purpose taxonomy of dialogue acts.

Taken together, CueTaxo is best understood as a taxonomy of **support-attribute completeness states** rather than a taxonomy of topics or concepts. Its significance lies in turning missing or moderate support cues into controlled, interpretable, and context-aware follow-up questions. Within MH-COPILOT, it is the component that converts assessment into action: span identification and intensity classification estimate what is missing, and CueTaxo determines how that missing information should be elicited [2508.16788].

Source: https://www.emergentmind.com/topics/cuetaxo