Interlocking-free Selective Rationalization Through Genetic-based Learning

Tue, 01 Jul 2025 00:00:00 +0000

Research setting

Selective rationalization builds text classifiers that explain themselves. Besides the label, the model returns a highlight, also called a rationale: the subset of input tokens on which its decision is based. The most popular design is the select-then-predict (SPP) architecture. A generator (the selector) reads the text and marks each token as selected or not, and a predictor classifies the text from the selected tokens alone, so the highlight is by construction what the decision rests on.

The select-then-predict architecture on an example string built with the rules of the paper’s Toy dataset, where class aba is defined by the hidden pattern a, b, a. The predictor sees only the highlight, so the highlight is a faithful explanation of the label.

The model learns from labels only: no gold highlights are given during training. Regularization terms push highlights to be short (sparsity) and made of connected tokens (contiguity).

Motivation

Training the generator and the predictor together leads to a problem called interlocking. If the generator gets stuck on a poor selection, the predictor adapts to that selection, which in turn pushes the generator to keep it. The two modules then settle in a poor solution. Prior work only mitigates interlocking with sampling, extra guidance modules, heuristics, or additional regularization.

Approach	Interlocking-free	No feature-based heuristics	No extra modules or tools	Predictor feedback reaches the generator
Differentiable sampling (Bao et al., 2018; Bastings et al., 2019)	no	yes	yes	yes
Weight sharing, FR (Liu et al., 2022)	no	yes	yes	yes
External guidance (Yu et al., 2021; Sha et al., 2023; Hu and Yu, 2024)	no	yes	no	yes
Heuristic generator pre-training (Jain et al., 2020)	yes	no	no	no
Three-stage training (Li et al., 2022)	partially	yes	yes	yes
This paper (GenSPP) This work	yes	yes	yes	yes

How GenSPP relates to the prior work discussed in the paper (Sections 1, 3, and 4).

Objective. Remove interlocking altogether by training the generator and the predictor disjointly, without heuristics, extra regularization, or architectural changes.

Approach

GenSPP keeps the select-then-predict architecture and changes only how it is trained. The generator is not trained with gradients: a genetic algorithm searches over its weights. Each candidate set of generator weights is an individual, and the set of individuals is the population. To score an individual, GenSPP freezes it and trains a brand new predictor on its highlights, so a generator is never judged by a predictor that has adapted to an earlier selection. The score, called fitness, comes from the validation set: the lower the classification loss and the highlight regularization, which weigh equally, the higher the fitness.

1. Initialize

Input text
Generator Optimized
Highlights
Predictor
Label

A population of I generators is initialized at random (I = 50 in the experiments). Each individual is one full set of generator weights.

2. Train predictor

Input text Data
Generator Frozen
Highlights Data
Predictor Trained
Label Data

For each individual, a predictor is initialized from scratch and trained with gradient descent to minimize the classification loss, while the generator weights stay frozen.

3. Fitness

Input text Data
Generator Frozen
Highlights Data
Predictor Frozen
Label Data

The trained predictor is evaluated on the validation set with the generator still frozen. Fitness is a minimization problem turned into a score: the lower the classification loss and the highlight regularization, which weigh equally, the higher the fitness.

4. Select parents

Input text
Generator Optimized
Highlights
Predictor
Label

From the population of I generators, I/2 pairs of parents are drawn by roulette-wheel selection, where each individual is picked with probability proportional to its fitness.

5. Crossover

Input text
Generator Optimized
Highlights
Predictor
Label

One-point crossover mixes each pair. A swap point is drawn at random, and the two parents exchange the parameters beyond that point, so the I/2 pairs produce I new generators.

6. Mutation

Input text
Generator Optimized
Highlights
Predictor
Label

Each parameter of the new generators is perturbed with Gaussian noise, with mutation probability pm. Crossover explores the weight space globally, and mutation explores it locally.

7. Score offspring

Input text Data
Generator Frozen
Highlights Data
Predictor Trained
Label Data

Every new generator goes through stages 2 and 3. It is frozen, a fresh predictor is trained on its highlights, and its fitness is computed.

8. Survival

Input text
Generator Optimized
Highlights
Predictor
Label

Old and new generators compete for the I places of the next generation. Half of the places go to the fittest individuals, and the other half are filled by roulette-wheel selection.

Optimized by the genetic search Trained with gradients Kept frozen Data

The GenSPP genetic algorithm (Algorithm 1 and Section 5 of the paper). Stages 1 to 3 run once on the initial population. Stages 4 to 8 form one generation, repeated for G generations (G = 100 in the main experiments).

The fitness does not need to be differentiable, so GenSPP needs no sampling tricks and no dataset-specific sparsity target.

Results

The paper compares GenSPP with four select-then-predict systems: FR, MGR, MCD, and G-RAT. All systems share the same recurrent architecture, and every score is an average over five runs with different random seeds.

Two datasets are used, both with gold highlights that serve only for evaluation:

Toy is a synthetic dataset introduced in the paper. It has 10k random strings of 20 characters, and each of three classes is defined by one hidden pattern (aba, baa, or abc), which is the gold highlight. Strings also contain pieces of the other patterns, so the full pattern must be selected.
HateXplain is a dataset of about 20k English social media posts from platforms such as X and Gab. Each post is labeled by at least three annotators, who also marked the words that justify the label. The paper uses it for binary hate speech classification and merges the annotators’ highlights by majority voting.

Three metrics describe each system:

Highlight F1 (Hl-F1) compares the selected tokens with the gold highlight, token by token. Higher is better.
Macro-F1 (Clf-F1) measures classification quality, averaged over classes. Higher is better.
Selection rate (R) is the percentage of input tokens that the generator selects, from 0 to 100. Lower means shorter highlights.

76.02Hl-F1 on Toy, against 65.70 for the best baseline
42.62Hl-F1 on HateXplain, against 36.17 for the best baseline
6.51%selection rate on HateXplain, against 24.68% to 25.55% for the baselines

Highlight F1 (Hl-F1) on Toy (higher is better)

FR 54.07 ± 4.02
MGR 50.34 ± 11.23
MCD 65.70 ± 3.76
G-RAT 50.22 ± 7.78
GenSPP This work 76.02 ± 0.64

Highlight quality when training from scratch. Source: Table 1 of the paper (average over five seed runs).

Show as table

System	Highlight F1 (Hl-F1) on Toy
FR	54.07 ± 4.02
MGR	50.34 ± 11.23
MCD	65.70 ± 3.76
G-RAT	50.22 ± 7.78
GenSPP (this work)	76.02 ± 0.64

Highlight F1 (Hl-F1) on HateXplain (higher is better)

FR 31.15 ± 2.56
MGR 29.38 ± 4.83
MCD 27.92 ± 1.66
G-RAT 36.17 ± 1.62
GenSPP This work 42.62 ± 0.73

Highlight quality when training from scratch. Source: Table 1 of the paper (average over five seed runs).

Show as table

System	Highlight F1 (Hl-F1) on HateXplain
FR	31.15 ± 2.56
MGR	29.38 ± 4.83
MCD	27.92 ± 1.66
G-RAT	36.17 ± 1.62
GenSPP (this work)	42.62 ± 0.73

Model	Toy Clf-F1	Toy R (%)	HateXplain Clf-F1	HateXplain R (%)
FR	99.78	14.80	72.14	25.55
MGR	99.92	15.05	71.14	25.30
MCD	99.90	15.18	70.37	25.07
G-RAT	99.36	14.81	73.85	24.68
GenSPP	99.00	11.47	69.71	6.51

Classification quality and selection rate. Source: Table 1 of the paper.

GenSPP selects the most accurate highlights on both datasets, 10.3 Hl-F1 points above the best baseline on Toy and 6.5 points above it on HateXplain. Its classification macro-F1 stays comparable to the baselines, and its highlights are much shorter on HateXplain. Its scores also vary less across seed runs, while MGR and G-RAT are notably unstable on Toy. In a second test, the generator starts from a deliberately skewed state, pre-trained to select the first token according to the class label, which imitates interlocking: GenSPP recovers and reaches 74.28 Hl-F1 on Toy and 42.81 on HateXplain with G = 150 generations (Table 2 of the paper).

Takeaways

Interlocking can be removed, not only mitigated.Judging each frozen generator with a freshly trained predictor breaks the dependency between the two modules that causes interlocking.
Genetic search fits rationalization well.It needs no differentiable objective, no sampling, and no dataset-specific sparsity threshold, and it produced the best highlights on both benchmarks.
The cost is training time.A seed run takes about 36 minutes on Toy and 78 minutes on HateXplain, against about 8 and 4 minutes for the baselines. The paper notes that parallel evaluation of individuals could reduce this.

Neuroevolution | Language Technologies Lab

Interlocking-free Selective Rationalization Through Genetic-based Learning

Research setting

Motivation

Approach

Results

Takeaways