{"id":"d625ae67-7731-448d-bc5b-3a339fc8bce6","arxiv_id":"1909.00360","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AMBER frames emotion representations as ordered attribute descriptor sets with ambiguity functions, encompassing categorical, numerical, and ordinal labels in one formalism.","lead":"This paper offers a unified mathematical framework, AMBER, for describing emotion labels as ordered sets equipped with functions that encode ambiguity. The goal is to give affective computing a common language for comparing categorical, numerical, and ordinal emotion representations.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"AMBER's per-descriptor ambiguity function (Eq. 3) cannot encode joint ambiguity across attribute descriptors, so the Section 7 universality claim is unsupported without a formal constraint mechanism.","rationale":"The paper is a clearly written position piece with useful worked examples; the formal core is simple and self-contained. My concern does not attack the usefulness of treating ambiguity as distributions per label dimension, nor the insight that point estimates are special cases. It targets the precise universality sentence in Sec. 7 and the claim that implicit assumptions can be made explicit and comparable. Equation (3) quantifies ambiguity independently on each attribute descriptor; nothing in the paper specifies how ambiguity that lives on combinations of descriptors (e.g., exclusivity or correlation between labels) is represented. The special cases in Box 1 and Box 3 add ad hoc constraints, but a general framework needs a general constraint mechanism. Without one, the representation either conflates distinct emotion distributions (marginal insufficiency) or is rendered trivially universal by allowing arbitrary constraints. This is exactly the kind of assumption the paper wants to make explicit, so it is load-bearing. A conditional verdict is appropriate: the universality claim should be qualified or the framework extended with joint ambiguity functions or a constraint language. The reader's weakest assumption already points at the same expressive-power issue, so my analysis agrees with it.","tokens_in":18570,"tokens_out":8328,"duration_ms":79361,"concrete_test":"Take a two-category space (happy, sad) and write both (A) the categorical distribution p(happy)=0.5, p(sad)=0.5, with exactly one label assumed, and (B) the product distribution in which happy and sad are independent Bernoulli(0.5) over {present, absent}. Express each in the notation of Eq. (3) following the paper's own construction for categorical labels in Sec. 5.1. If (A) and (B) yield the same AMBER tuple, then the representation is not faithful; the fix would be to add a formal joint-ambiguity component or an explicit constraint language, and to re-derive the Sec. 5.1 emotion-profile and Sec. 5.2 distribution examples with that component. Also test whether the constraint ∑_n ξ_n(I)=1 from Box 1 is derivable from Eq. (3) alone; it is not.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Sec. 7: 'every emotion representation scheme employed in affective computing till date can be described within the proposed AMBER framework') rests on Eq. (3), which defines Ψ(t) as a set of N independent functions ξ_{n,t}: X^{(n)} → [0,∞). The framework therefore represents ambiguity as a tuple of per-descriptor scores. It does not define a joint ambiguity function over X^{(1)}×...×X^{(N)}, nor a general language for cross-descriptor constraints. This is not a corner case: for the categorical emotion-profile representations the paper itself discusses (Sec. 5.1, Box 1), 'either happy or sad, 50/50, no other option' and 'happy and sad each independently present with probability 0.5' produce identical ξ values under Eq. (3) (ξ_happy(I)=ξ_sad(I)=0.5, ξ_happy(O)=ξ_sad(O)=0.5), although they are different distributions over label combinations. The paper's special constraints, e.g. Eqs. (9)-(11), show that something beyond Eq. (3) is needed, but the general framework never formalizes how such joint/relational structure is expressed. If arbitrary constraints are permitted, the universality claim becomes unfalsifiable; if only Eq. (3) is intended, it is false for any scheme with non-independent ambiguity across descriptors. Either way the Section 7 claim needs a precise statement of how joint ambiguity is represented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the AMBiguous Emotion Representation (AMBER) framework, a mathematical language intended to describe categorical, numerical, and ordinal emotion representation schemes used in affective computing, including time-varying and ambiguity-aware variants. The framework represents an emotion at time t as a set of elements chosen from a finite number of ordered attribute descriptor sets, and extends this to per-descriptor ambiguity functions. The authors reinterpret existing schemes, including categorical labels, emotion profiles, blended emotions, dimensional labels, annotation distributions, and ordinal/rank labels, through AMBER, and argue in Section 7 that every emotion representation scheme employed in affective computing to date can be expressed in the framework. The paper is a conceptual contribution with no empirical evaluation.","tokens_in":18923,"tokens_out":6109,"duration_ms":54107,"significance":"If the universality claim could be made precise, AMBER would give affective computing researchers a common vocabulary for exposing and comparing the implicit assumptions of emotion labels, addressing a real need in the field. The paper's strengths are the clarity of the per-descriptor formalism, the worked re-descriptions in Boxes 1 through 4 and Section 5, and the explicit separation of ambiguity (perception and expression) from uncertainty (model imprecision). The central limitation is that the framework's core definition in Eq. (3) only captures per-descriptor ambiguity, and the paper does not formally account for joint structure across attribute descriptors; this gap is load-bearing because the universality claim rests on it.","major_comments":[{"comment":"The definition of Ψ(t) as a set of N independent functions ξ_{n,t} on each descriptor X^(n) cannot represent joint ambiguity across descriptors. For example, the two representations 'either happy or sad with probability 0.5 each' and 'happy and sad each independently present with probability 0.5' produce identical ambiguity functions under Eq. (3): ξ_happy(I) = ξ_sad(I) = 0.5 and ξ_happy(O) = ξ_sad(O) = 0.5 in both cases, even though they encode different distributions over label combinations. Since the framework introduces special constraints such as Eqs. (9)-(11) only for the blended-emotions example, it lacks a general mechanism for expressing cross-descriptor constraints. This makes the Section 7 universality claim either false (if only Eq. (3) is intended) or unfalsifiable (if arbitrary constraints are allowed). Please specify how joint and relational structure is represented in the general framework, or qualify the universality claim accordingly.","section":"Section 4.2, Eq. (3)"},{"comment":"The statement 'every emotion representation scheme employed in affective computing till date can be described within the proposed AMBER framework' is asserted after only a few illustrative examples. No enumeration of representation families is provided, and the term 'described' is not given a formal meaning. In particular, it is not clear what class of constraints such as Eqs. (9)-(11) is permitted, and whether the framework is closed under their composition. The paper should either provide a systematic account of the space of emotion representations and a proof sketch of universality, or replace the universal claim with a weaker statement about the examples presented.","section":"Section 7"},{"comment":"The proposed ambiguity-aware treatment of relative ordinal labels is not an instance of Eq. (3). The text in Section 5.3 and Figure 6B suggests defining an ambiguity function over the outcomes of each pairwise comparison, but Eq. (3) defines functions on a single descriptor X^(n), not on pairs of time points or comparisons. Similarly, the discussion of temporal dynamics in Section 5.2 mentions constraints between Ψ(t) and Ψ(t-ε) only as an open question. The framework needs a formal account of how such relational and temporal structures are expressed within AMBER.","section":"Section 5.3"}],"minor_comments":[{"comment":"The text says ξ maps to '[0, ∞)' but then describes it as associating each element with a 'positive real number'; since Eq. (4) assigns the value 0, the term should be 'non-negative real number'.","section":"Section 4.2"},{"comment":"The requirement that every X^(n) be an ordered set is applied to categorical labels, for which the two elements {O, I} have no natural order. The paper imposes an arbitrary order without discussing whether this order carries any semantic meaning.","section":"Section 4.1"},{"comment":"The distinction between ambiguity and uncertainty introduced in Section 1 is never formally reflected in the framework; the same ambiguity function appears to serve for both annotator disagreement and model prediction uncertainty. The paper should clarify whether ξ can represent one, the other, or both, and how the distinction is operationalized.","section":"Section 1 and Section 4.2"},{"comment":"The text contains a minor typo: 'an more general' should read 'a more general' when referring to the emotion profile representation.","section":"Section 5.1"},{"comment":"The Iverson bracket notation in Eq. (10) is defined immediately after its first use; moving the definition before the equation would improve readability. Additionally, the phrase '[condi t i on]' has unusual spacing in the preprint.","section":"Box 1"}],"recommendation":"major_revision","confidential_remarks":"This is a position/framework paper from well-known authors in affective computing. The central idea is useful and the examples are generally correct, but the universality claim is not supported by the formal apparatus. A major revision that either formalizes cross-descriptor and temporal constraints within Eq. (3) or explicitly restricts the claimed scope would be needed before publication. The editor should also consider whether the paper's contribution is best framed as a proposal with a strong claim or as a more modest unifying notation, since the latter may be more defensible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a sound, readable framework paper that does real work in making implicit assumptions in emotion labeling explicit. The core idea—represent a label space as a finite set of ordered attribute descriptors and represent ambiguity as a score per element of each descriptor—cleanly covers categorical, numerical, and ordinal labels, and the worked re-descriptions of blended emotions, emotion profiles, distributional labels, and relative ordinal labels are accurate. The paper earns credit for showing how a categorical label is a special case of the framework (a delta ambiguity function), and for exposing assumptions like Euclidean distance, orthogonal dimensions, and annotator mapping. I have started using the terminology when thinking about labels.\n\nSoft spots, in order. The universality claim in Section 7 is the load-bearing problem. Equation (3) defines ambiguity as a tuple of independent per-descriptor functions, so it cannot by itself represent joint structure across descriptors: the state \"either happy or sad with probability 0.5 each\" and the state \"happy and sad co-occurring, each with marginal 0.5\" give identical ξ values, though they are different distributions over label combinations. The paper's own Box 1 constraints and the discussion of Gaussian processes show that something beyond Eq. (3) is needed for relational, cross-descriptor, or temporal structure, but the general constraint mechanism is never formalized. If arbitrary constraints are allowed alongside Eq. (3), \"every scheme can be described\" becomes hard to falsify; if only Eq. (3) is intended, it is false for any scheme with joint ambiguity. Either way, Section 7 needs either a precise definition of the full representation (descriptors plus allowed constraints plus ambiguity functions) and a systematic argument or enumeration, or a weaker claim. This is fixable, and it does not sink the descriptive value of the framework.\n\nMinor: the ambiguity/uncertainty distinction is clear and useful, but the paper does not give criteria for choosing among ambiguity functions; that is honestly left open. Self-citations are used as illustrations, not evidence, which is fine for a conceptual paper. The citation list is broad and appropriate.\n\nWho benefits: affective computing researchers designing annotation schemes, comparing label formats, or translating between categorical and dimensional labels. It will not change empirical results on its own, but it gives the field a vocabulary it lacks. Deserves a serious referee round, with the revision centered on the universality claim.","headline":"AMBER gives affective computing a genuinely useful common vocabulary for emotion labels, but the Section 7 universality claim is broader than the formalism in Eq. (3) supports.","tokens_in":19379,"tokens_out":4301,"would_cite":true,"duration_ms":41868,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AMBER describes every emotion representation scheme used in affective computing, from single categorical labels to time-varying distributions over numerical and ordinal labels, in one mathematical language.","keywords":["emotion representation","affective computing","ambiguity","categorical labels","numerical labels","ordinal labels","annotator disagreement","AMBER framework"],"falsifier":"Find a published emotion representation whose label space cannot be written as a finite set of ordered attribute descriptors with non-negative per-element ambiguity functions—for example, one whose meaning depends on the joint distribution or correlation between dimensions, such as a valence–arousal trace where the correlation itself carries information—and show that it cannot be reconstructed from the per-descriptor ambiguity functions alone. Such a case would refute the claim that every scheme fits AMBER without adding new structure.","tokens_in":18426,"feed_emoji":"🎭","tokens_out":5920,"duration_ms":53204,"temperature":0.7,"pith_summary":"AMBER is a proposed common mathematical language for emotion labels. It claims that every way affective computing has represented emotion—categorical labels, numerical dimensions such as arousal and valence, ordinal ratings or rankings, and time-varying versions of all three—can be described by the same two ingredients: a finite set of ordered attribute descriptors and an ambiguity function assigning a non-negative number to each element of each descriptor. The point is to make implicit assumptions explicit, so researchers can compare schemes, see what treating annotator disagreement as noise actually assumes, and design representations that carry ambiguity rather than averaging it away. If the claim holds, affective computing gains a rigorous way to reason about ground truth and about conversions between label types.","feed_headline":"Every emotion label scheme fits one mathematical framework","feed_subtitle":"AMBER turns categorical, numeric, and ordinal labels into ordered attribute sets with ambiguity functions, exposing hidden assumptions.","key_machinery":"The central objects are attribute descriptors and ambiguity functions. Each attribute descriptor $X^{(n)}$ is an ordered set $\\{x^{(n)} \\mid \\alpha^{(n)} \\preceq x^{(n)} \\preceq \\beta^{(n)}\\}$: categorical schemes use two-element sets $\\{O, I\\}$ per emotion category, numerical schemes use intervals such as $[-1,1]$ for valence and arousal, and ordinal schemes use finite ordered sets without any distance notion. The ambiguity function $\\xi_{n,t}: X^{(n)} \\to [0,\\infty)$ assigns a non-negative number to every element of each descriptor at time $t$, encoding how likely that element is; equation (4) shows that single-valued representations are exactly the special case where the function is 1 on one element and 0 elsewhere. This pair of ingredients does the work of unifying the three label spaces and of turning previously hidden assumptions into explicit mathematical restrictions.","core_discovery":"The central claim is that the AMBER framework can describe every emotion representation scheme employed in affective computing to date. An emotion at time $t$ is represented as a set of functions $\\xi_{n,t}: X^{(n)} \\to [0,\\infty)$ over attribute descriptors, where each descriptor $X^{(n)}$ is an ordered set with lowest element $\\alpha^{(n)}$ and highest element $\\beta^{(n)}$. Single-valued categorical, numerical, and ordinal labels are special cases in which the ambiguity function equals 1 for the chosen element and 0 for all others; blended emotions, emotion profiles, distributional labels, and multi-annotator traces are less constrained instances of the same structure. The consequence is that implicit choices—mutual exclusivity of categories, Euclidean distance in the arousal–valence plane, the assumption that annotators transcribe identically, or the assumption that labels do not vary within an interval—can be stated explicitly and compared across schemes.","pith_inferences":["If the universality claim holds, one could build a representation calculus in which schemes are compared by their assumption load, and a researcher choosing a label scheme could see exactly which mathematical restrictions they are buying into.","A testable extension would be to define a formal measure of the representational capacity of a scheme—for instance, the number of independent ambiguity functions or the richness of the ordered sets—and use it to predict when one annotation scheme can substitute for another.","The framework suggests that future affective benchmarks might report full ambiguity functions instead of averaged labels, making inter-annotator disagreement a first-class evaluation target rather than noise to be removed.","One could extend AMBER to compare ambiguity functions across modalities, treating each modality as a separate attribute descriptor and using the framework to reason about multimodal emotion perception."],"forward_implications":["Researchers can express any existing emotion representation in a shared notation, making categorical, numerical, and ordinal schemes directly comparable.","Ambiguity-aware training becomes a well-defined task: the loss can be a divergence between the predicted ambiguity function and the target ambiguity function, such as Kullback–Leibler divergence when both are probability distributions.","Conversions between schemes—for example, from Likert-type annotations to continuous $[-1,1]$ predictions, or from numerical labels to categorical clusters—can be analyzed for the distance metrics and orthogonality assumptions they silently introduce.","Treating the averaged annotation as ground truth is revealed as a special assumption of certainty rather than a neutral default, so datasets and models that ignore inter-annotator disagreement become open to explicit scrutiny.","Time-varying ambiguity functions provide a natural language for modeling annotator reaction lag, drift, and systematic disagreement over time."],"supporting_citations":[{"why":"Supplies the catalogue of emotional states and the primary/secondary emotion distinctions that categorical schemes rely on.","marker":"[13]"},{"why":"Supplies a Gaussian mixture model of emotion prediction uncertainty, an example of a numerical ambiguity function that AMBER subsumes.","marker":"[19]"},{"why":"Supplies the primitives-based emotion evaluation approach, a precursor to the emotion profile representation.","marker":"[27]"},{"why":"Supplies a Gaussian soft-label treatment of perception uncertainty, used as an example of a constrained ambiguity function.","marker":"[30]"},{"why":"Defines emotion profiles, a central example of an ambiguity-aware categorical representation that AMBER generalizes.","marker":"[49]"},{"why":"Provides relative ordinal labels derived from pairwise comparisons between time-continuous traces, the basis for equation (13) and Figure 5.","marker":"[54]"},{"why":"Provides the circumplex model of affect, a theoretical grounding for dimensional representations.","marker":"[60]"},{"why":"Introduces blended emotions with major and minor categorical labels, an example of a constrained ambiguity-aware representation.","marker":"[71]"},{"why":"Argues for the ordinal nature of emotion perception and supplies the motivation for treating ordinal labels as a distinct representation space.","marker":"[75, 76]"}],"fun_headline_variants":["AMBER unifies categorical, numeric, ordinal labels","One math framework for all emotion label schemes","Emotion labels from discrete to fuzzy fit AMBER","AMBER makes emotion label assumptions explicit","A single formalism for every emotion representation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that every emotion representation can be faithfully captured by a finite number of ordered attribute descriptors, with all ambiguity expressed as a non-negative number attached to each element of each descriptor; if emotions require relations between descriptors, context-dependent meaning, or a structure that an ordered set cannot express, the universality claim fails.","fun_headline_variants_meta":{"raw":{"variants":["AMBER unifies categorical, numeric, ordinal labels","One math framework for all emotion label schemes","Emotion labels from discrete to fuzzy fit AMBER","AMBER makes emotion label assumptions explicit","A single formalism for every emotion representation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1474,"prompt_tokens":914,"completion_tokens":560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":492}},"tokens_in":530,"tokens_out":560,"duration_ms":5796,"temperature":1.0,"reasoning_tokens":492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:54:11.351767+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a published emotion representation whose label space cannot be written as a finite set of ordered attribute descriptors with non-negative per-element ambiguity functions—for example, one whose meaning depends on the joint distribution or correlation between dimensions, such as a valence–arousal trace where the correlation itself carries information—and show that it cannot be reconstructed from the per-descriptor ambiguity functions alone. Such a case would refute the claim that every scheme fits AMBER without adding new structure.","supporting_citations":[{"cited_title":"and Cornelius, R","cited_arxiv_id":null,"evidence_quote":"Supplies the catalogue of emotional states and the primary/secondary emotion distinctions that categorical schemes rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies a Gaussian mixture model of emotion prediction uncertainty, an example of a numerical ambiguity function that AMBER subsumes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the primitives-based emotion evaluation approach, a precursor to the emotion profile representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies a Gaussian soft-label treatment of perception uncertainty, used as an example of a constrained ambiguity function."},{"cited_title":"J., and Narayanan, S","cited_arxiv_id":null,"evidence_quote":"Defines emotion profiles, a central example of an ambiguity-aware categorical representation that AMBER generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides relative ordinal labels derived from pairwise comparisons between time-continuous traces, the basis for equation (13) and Figure 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the circumplex model of affect, a theoretical grounding for dimensional representations."},{"cited_title":"and Devillers, L","cited_arxiv_id":null,"evidence_quote":"Introduces blended emotions with major and minor categorical labels, an example of a constrained ambiguity-aware representation."}],"review_version":1}