Speech Concept Bottleneck Model (SCBM)

• SCBM is a text-classification model that maps input text to a human-interpretable vector of adjective relevance scores, enhancing transparency.
• It employs a frozen LLM to generate concept vectors via descriptive adjectives, ensuring that the classification decision is fully explainable.
• Designed for sexism detection in social media, SCBM provides both local and global interpretability while balancing performance with structured concept constraints.

Searching arXiv for the specified paper and closely related concept bottleneck work to ground the article. arxiv_search(query="(Labadie-Tamayo et al., 28 Jul 2025)", max_results=5) arxiv_search(query="Speech Concept Bottleneck Model (Labadie-Tamayo et al., 28 Jul 2025)", max_results=10) The Speech Concept Bottleneck Model (SCBM) is a text-classification architecture introduced in the FHSTP@EXIST 2025 system for sexism detection in social-media posts. It addresses a stated tension between opaque dense representations from large pretrained encoders and more transparent but weaker interpretable methods by inserting a concept bottleneck between the input text and the downstream classifier. In SCBM, an input text $x \in X$ is mapped to a human-interpretable vector of adjective relevance scores $c \in \mathbb{R}^K$, and prediction is made only from that concept vector. The reported implementation targets the CLEF 2025 EXIST benchmark, especially the tasks of sexism identification, source intention classification, and sexism categorization in tweets, and uses descriptive adjectives as the operative “speech concepts” (Labadie-Tamayo et al., 28 Jul 2025).

1. Problem setting and conceptual rationale

SCBM was developed for the fifth Sexism Identification in Social Networks (EXIST) challenge at CLEF 2025, with emphasis on Task 1: identifying and classifying sexism in social-media textual posts. The system is described for three subtasks: Subtask 1.1, Sexism Identification in Tweets; Subtask 1.2, Source Intention in Tweets; and Subtask 1.3, Sexism Categorization in Tweets (Labadie-Tamayo et al., 28 Jul 2025).

The central design choice is the “concept bottleneck.” Rather than allowing the classifier to consume a dense hidden representation directly, SCBM forces all downstream predictions to depend on a fixed set of human-interpretable concepts. In this formulation, the concepts are descriptive adjectives $\{a_1,\ldots,a_K\}$ intended to capture affective, evaluative, and discursive dimensions of a tweet, with examples including “misogynistic,” “condescending,” and “invasive.” Each adjective is treated as a concept that may or may not apply to a given tweet, and the model assigns each concept a relevance score in $[0,1]$.

This design yields two intended consequences. First, it preserves the use of LLMs for semantic encoding. Second, it makes post-hoc explanations available at both local and global levels because the classifier’s input is explicitly interpretable. A plausible implication is that SCBM should be read less as a replacement for LLM-based semantic processing than as a structural constraint on how that processing can affect a decision.

2. Speech concepts and adjective lexicon construction

In SCBM, “speech concepts” are defined as a fixed lexicon of descriptive adjectives. The adjective set is task-conditioned at construction time but fixed during classification. To obtain the lexicon, GPT-o3-mini was prompted separately for each EXIST subtask, using in-context examples and class definitions, and asked to generate 50 relevant adjectives per task. After deduplication across subtasks, the final lexicon contains $K=132$ adjectives (Labadie-Tamayo et al., 28 Jul 2025).

These adjectives are treated as atomic concepts. The reported formulation does not introduce an additional embedding layer or token-level concept encoder for them; instead, concept relevance is derived directly from LLM outputs. This is a notable aspect of the model definition: the concept space is human-readable from the outset, and interpretability is anchored in a lexicon rather than in latent dimensions that are later given a semantic label.

Because the adjective lexicon is automatically generated, its coverage is broad but not guaranteed to be exhaustive. The paper explicitly notes that domain-specific nuance may be omitted. This suggests that SCBM’s interpretability is conditioned not only on the classifier architecture but also on the adequacy of the chosen concept inventory.

3. Architecture and inference pipeline

SCBM uses a two-stage pipeline consisting of a concept extractor $g_\theta : X \rightarrow \mathbb{R}^K$ and a classifier $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$ (Labadie-Tamayo et al., 28 Jul 2025).

For an input text $x$, the concept extractor produces

$g_\theta(x) = c \in \mathbb{R}^K,$

where the $j$-th component $c \in \mathbb{R}^K$0 measures the relevance of adjective $c \in \mathbb{R}^K$1 to the text. The implementation uses a frozen, 8-bit quantized instance of Llama3.1-8b-instruct. For each adjective $c \in \mathbb{R}^K$2, the model issues the prompt:

“Tell me if the adjective [a_j] describes the content of the following text: [x]?”

Let $c \in \mathbb{R}^K$3 denote the set of affirmative tokens and $c \in \mathbb{R}^K$4 the probability that the LLM generates token $c \in \mathbb{R}^K$5 as its first token. SCBM defines

$c \in \mathbb{R}^K$6

Affirmative tokens are described as including forms such as “Yes,” “Si,” and “_yes,” among others. The concept vector is therefore

$c \in \mathbb{R}^K$7

The classifier stage applies a trainable relevance gate followed by a multi-layer perceptron. Let $c \in \mathbb{R}^K$8 be a learned gating vector initialized to ones, and define

$c \in \mathbb{R}^K$9

where $\{a_1,\ldots,a_K\}$0 denotes element-wise multiplication. Prediction is then produced by

$\{a_1,\ldots,a_K\}$1

The MLP uses two hidden layers of sizes $\{a_1,\ldots,a_K\}$2 with ReLU activations, and the output layer is task-dependent: sigmoid units for multi-label prediction and softmax for multi-class prediction.

A common misconception would be to treat SCBM as an end-to-end jointly trained concept bottleneck architecture. In the reported formulation, however, the concept extractor is frozen. Only the relevance gates and classifier parameters are optimized. The concept space is therefore externally induced by LLM prompting rather than learned from supervision during downstream training.

4. Training objectives, optimization, and data regime

Because $\{a_1,\ldots,a_K\}$3 is frozen, SCBM training reduces to optimizing the classifier parameters $\{a_1,\ldots,a_K\}$4 together with the gating vector $\{a_1,\ldots,a_K\}$5. For a training pair $\{a_1,\ldots,a_K\}$6, the model computes $\{a_1,\ldots,a_K\}$7, then $\{a_1,\ldots,a_K\}$8, and finally $\{a_1,\ldots,a_K\}$9. There is no explicit concept-prediction loss $[0,1]$0 in the reported system. The total objective is therefore

$[0,1]$1

This absence of a trainable concept extractor is one of the main structural differences between SCBM and concept-bottleneck variants that explicitly supervise intermediate concepts (Labadie-Tamayo et al., 28 Jul 2025).

The classification loss depends on the subtask. For Subtask 1.1, binary sexism identification,

$[0,1]$2

where $[0,1]$3 is the predicted probability of the class “SEXIST.”

For Subtask 1.2, multiclass source intention,

$[0,1]$4

with $[0,1]$5 a one-hot vector over $[0,1]$6.

For Subtask 1.3, multi-label sexism categorization,

$[0,1]$7

The reported SCBM hyperparameters are learning rate $[0,1]$8, RMSProp optimization, batch size 32, and 300 epochs, with early stopping on macro-F1 on the development set. The data regime uses the EXIST2025 training set of 6,920 tweets, a development set of 1,038 tweets, and a test set of 2,076 tweets. Augmentation with EXIST2022, comprising 11,345 tweets, is reported for the transformer baseline rather than for SCBM. For Subtasks 1.2 and 1.3, runs addressing class imbalance apply random undersampling of the majority class.

The system also explores annotator metadata. In those runs, the prompt to $[0,1]$9 is enriched with a persona string such as “You are a 45-year-old Latino male with a Bachelor’s degree.” Multiple concept vectors can then be generated for a single tweet, one per annotator persona, and predictions are aggregated by majority vote.

5. Interpretability mechanism and explanatory granularity

SCBM’s interpretability follows directly from the requirement that every prediction flow through the adjective vector $K=132$0 and the relevance gate $K=132$1. Explanations are therefore phrased in terms of concept activations rather than hidden-state saliency or token attribution (Labadie-Tamayo et al., 28 Jul 2025).

At the local, instance-level scale, one examines the effective importance scores $K=132$2 for a given test tweet and selects the top 10 adjectives. The paper reports examples in which tweets labeled SEXIST are associated with adjectives such as “harassing,” “degrading,” and “misogynistic,” and describes these explanations as coherently reflecting the content. The local explanation mechanism is therefore not ancillary; it is structurally identical to the representation consumed by the classifier.

At the global, class-level scale, the method averages $K=132$3 over correctly classified training instances belonging to class $K=132$4. The top 10 adjectives under this class-conditional average serve as a global explanation for that class. For the DIRECT intent class, the reported most salient adjectives include “patriarchal,” “dehumanizing,” “chauvinistic,” and “androcentric.” This highlights stable semantic patterns associated with the class under the model.

A common misconception is to equate interpretability with mere feature listing. In the SCBM formulation, interpretability is operationalized through a constrained prediction pathway: the same intermediate quantities that support explanation also determine the classifier input. A plausible implication is that SCBM offers a stronger form of explanation than purely post-hoc methods that do not constrain the prediction process itself.

6. Benchmark performance, SCBMT extension, and limitations

The paper evaluates three models: SCBM, SCBMT, and a fine-tuned XLM-RoBERTa transformer model. Results are reported under both Soft-Soft and Hard-Hard evaluation protocols, using ICM-Soft, ICM-Soft Normalized, and Cross-Entropy. The key summary in the paper emphasizes Soft-Soft ICM-Soft Normalized, where higher is better, and Cross-Entropy, where lower is better (Labadie-Tamayo et al., 28 Jul 2025).

For Subtask 1.1, binary sexism detection, the reported scores are approximately as follows: SCBM, ICM-Soft Norm $K=132$5 and Cross-Ent $K=132$6; SCBMT, ICM-Soft Norm $K=132$7 and Cross-Ent $K=132$8; XLM-RoBERTa, ICM-Soft Norm $K=132$9 and Cross-Ent $g_\theta : X \rightarrow \mathbb{R}^K$0, with 4th place overall. The abstract also states that XLM-RoBERTa, fine-tuned on provided data augmented with prior datasets, ranks 6th for English and Spanish and 4th for English in the Soft-Soft evaluation.

For Subtask 1.2, source intention, the reported scores are SCBM, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$1 and Cross-Ent $g_\theta : X \rightarrow \mathbb{R}^K$2; SCBMT, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$3 and Cross-Ent $g_\theta : X \rightarrow \mathbb{R}^K$4, with 7th place; and XLM-RoBERTa, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$5 and Cross-Ent $g_\theta : X \rightarrow \mathbb{R}^K$6.

For Subtask 1.3, sexism categorization, the reported scores are SCBM, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$7; SCBMT, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$8, with 14th place; and XLM-RoBERTa, ICM-Soft Norm $g_\theta : X \rightarrow \mathbb{R}^K$9. Cross-Entropy is stated as not applicable for the multi-label setting.

The reported comparative pattern is consistent across the benchmark discussion. The plain transformer excels at coarse binary classification, whereas SCBMT’s fusion of adjective concepts with XLM-RoBERTa embeddings yields superior performance on the finer-grained multiclass and multi-label subtasks. SCBM alone lags in absolute accuracy but meets the system’s interpretability goals.

SCBMT extends SCBM by combining the gated concept vector with a projected transformer embedding. If $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$0 denotes the transformer contextual embedding, SCBMT defines

$f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$1

The transformer weights remain frozen; only the projection $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$2, the gates $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$3, and the classifier parameters $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$4 are learned. The reported training regime for SCBMT uses learning rate $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$5 and 16 epochs. The paper describes this as a way to reduce computational cost relative to full transformer fine-tuning while balancing interpretability and performance.

The limitations section identifies three main constraints. First, SCBM’s reliance on frozen LLM calls per adjective incurs latency. Second, the automatically generated adjective lexicon, although broad at $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$6, may omit domain-specific nuance. Third, the current formulation does not sharpen the relation between concepts and decisions through additional constraints such as sparsity or mutual-information regularization. The proposed future directions are correspondingly explicit: train a lightweight concept predictor $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$7 that approximates the LLM outputs and supports end-to-end learning with an explicit concept-prediction loss $f_\phi : \mathbb{R}^K \rightarrow \mathcal{Y}$8; incorporate expert-curated concepts or hierarchical concept taxonomies; and experiment with alternative interpretability constraints. These proposals suggest that SCBM, as reported, is both a working benchmark system and a template for more tightly integrated concept-based text classification.