When a language app decides when to show you a word again, it relies on a model of how fast you forget. The best-known model, Half-Life Regression (HLR), assumes every word fades along the same exponential curve and only lets each word and learner have its own speed. Our starting observation is that HLR is already a proportional hazards model, the classic survival model in which covariates scale a shared baseline risk of an event over time. Here the event is forgetting, and HLR's baseline happens to be a constant forgetting rate. Seen that way, the obvious generalization is to let the data choose the curve's shape instead of fixing it as exponential. That lets us ask a question current models cannot express: does vocabulary learned inside a professional context, such as veterinary terms taught through clinical scenarios, fade differently, and not just more slowly, than words drilled from a list?
The question matters for domain-specific language teaching. Advocates of teaching language inside a discipline (English for Specific Purposes, "Chinese+", content-and-language-integrated learning) claim that tying terms to a learner's existing conceptual knowledge makes them stick. The evidence is mostly motivational and survey-based. If context-anchored terms have a forgetting risk that drops over time while list-learned terms keep a constant risk, the two need different review schedules, and a scheduler that assumes one exponential shape will over-review one kind and under-review the other. We propose to build the model, validate it on large existing practice logs, and test the shape hypothesis directly in a randomized retention study in a specialized-language course.
Core Claim
Retention data from review-based learning are not "time until forgetting" data. At a review after delay $\Delta$ we only see whether the learner still remembers. Survival statisticians call this current-status data: you learn whether the event happened by the inspection time, never when it happened. Under a proportional hazards model, the probability of recall at delay $\Delta$ for a learner–item pair with features $x$ in item class $g$ is
$$ p(\Delta \mid x, g) = \exp\!\big(-\Lambda_g(\Delta)\, e^{\beta^T x}\big), $$
where $\Lambda_g$ is a nondecreasing baseline cumulative forgetting hazard for items of class $g$ (for example, context-anchored technical terms versus general or list-learned words), and $\beta$ weights learner and item features such as past correct and incorrect counts.
HLR from A Trainable Spaced Repetition Model for Language Learning models recall as $p = 2^{-\Delta/h}$ with half-life $h = 2^{\Theta^T x}$. Rewriting it gives $p = \exp(-(\ln 2)\,\Delta\, e^{-(\ln 2)\Theta^T x})$. That is exactly the model above with one linear baseline $\Lambda(\Delta) = (\ln 2)\Delta$ and $\beta = -(\ln 2)\Theta$.
Our proposal makes two changes to HLR. First, it replaces the linear baseline with a flexible monotone spline estimated from data. Second, it lets the baseline differ by item class. The scientific claim is testable. Suppose teaching technical terms through professional schemas changes only how fast they fade. Then the class baselines are proportional, $\Lambda_{\text{domain}}(\Delta) = c\,\Lambda_{\text{general}}(\Delta)$, and adding a simple domain indicator to HLR is enough. Suppose instead it changes the shape of forgetting. Then the baselines cross or diverge, and no rescaling of HLR can capture it.
The correct likelihood is a binomial likelihood with a complementary log-log link, not Cox's partial likelihood. This has a practical side effect. The risk-set computation that makes Cox regression expensive on millions of records, which FastSurvival solves with exact linear-time derivatives, never arises. The model fits with ordinary minibatch or Newton-type optimization on the 12.9M-trace Duolingo data.
Relation to Prior Work
Four papers motivate the gap by showing that forgetting matters for prediction while leaving its shape unmodeled. HLR builds a trainable exponential forgetting curve and reports over 45% lower mean absolute error than the Leitner and Pimsleur schedules. Its authors say the exponential form may be suboptimal, and that fine-grained word tags made predictions too pessimistic on rare words. We keep HLR's features and likelihood-based spirit but free the curve shape.
DAS3H handles forgetting through hand-chosen time windows (24 hours, 7 days, 30 days) inside a multi-skill model and reports AUC 0.826 on Algebra05. That is effectively a coarse step-function baseline. Ours is continuous, monotone and stratified.
The SLAM 2018 systems predict word-level errors as snapshot classification. The winning LSTM-plus-gradient-boosting ensemble in Second Language Acquisition Modeling: An Ensemble Approach (AUC 0.861 on en_es) loses accuracy further from its training data. The context-feature logistic model in Context Based Approach for Second Language Acquisition models no elapsed time at all. Neither estimates a forgetting curve, which is what a scheduler needs.
LECTOR adjusts half-lives heuristically for interference between similar words and is evaluated only in simulation. Its recall and review-burden metrics supply our scheduling evaluation, but it does not estimate curve shape from human data.
The domain-teaching literature supplies the hypothesis rather than the method. Research on Narrow Context ESP Course Design in Adult Education describes phased vocational curricula (for example, animal care leading to clinical consultation) and reports engagement gains. An Application of the Input Hypothesis to the Teaching Design of English Reading Class in Senior High School argues for anchoring new language in existing schemas. EduCSW shows that computer-science terminology can be identified and embedded automatically in natural mixed-language text. We reuse its term-identification approach to label technical items. None of these works measures retention over time, which is the gap this project fills.
On the survival side, DeepSurv replaces the linear risk score with a neural network and reports C-index 0.867 versus 0.816 for linear Cox on METABRIC. It constrains our design in one respect: it assumes observed event times and uses the partial likelihood, so it cannot be applied to recall logs as-is. We treat a neural risk score as an optional extension under the correct current-status likelihood. The time-varying-covariate work in Generating Survival Times Using Cox Proportional Hazards Models with Cyclic and Piecewise Time-Varying Covariates and A Federated Cox Model with Non-Proportional Hazards supplies the tools to model non-proportionality if our shape test finds it.
Research Program
The first component is the estimator. We represent each $\Lambda_g$ as a monotone spline in $\log \Delta$ with nonnegative coefficients on an increasing basis. This family contains the exponential (HLR) curve and power-law curves as special cases, so likelihood-ratio comparisons against both are nested. We fit by maximizing the binomial complementary log-log likelihood.
For rare technical terms, per-word parameters are shrunk toward their item class and morphological family. This targets the rare-word pessimism HLR reported. The ablation is to remove the shrinkage.
One subtlety needs explicit handling. In production logs, the scheduler picks review times based on predicted recall. Current-status estimation is unbiased only if the inspection time is independent of the true forgetting time given the covariates. Because the Duolingo scheduler's inputs are essentially HLR's features, which we condition on, the assumption is plausible there. We nonetheless treat it as the main threat to validity for observational data, and it is why the decisive test uses randomized delays.
The second component is the proportionality test. We fit a single shared baseline and class-specific baselines. We then test whether the log ratio $\log\Lambda_{\text{domain}}(\Delta) - \log\Lambda_{\text{general}}(\Delta)$ is flat in $\Delta$, using a bootstrap over learners. A flat ratio means domain anchoring acts as a rate multiplier only. A sloped ratio means it changes the curve shape.
The third component translates fitted curves into schedules. For each class we compute the review interval that keeps predicted recall above a target. We then compare total review burden and end-of-horizon recall against schedules implied by HLR and by the flexible predictors, in a LECTOR-style simulation driven by our fitted human curves rather than synthetic learners.
Baselines
The naive reference is a class-level constant recall rate, plus the Leitner and Pimsleur schedules HLR compared against. Any model that fails to beat these on held-out recall is uninformative.
The strongest forgetting-specific competitor is HLR itself, refit on identical train/test splits with the same features, the same tuning budget, and the authors' published loss. The file reports its improvement only relatively, as over 45% lower mean absolute error than Leitner and Pimsleur. No reported absolute MAE for HLR on the Duolingo 12.9M dataset was found on Lacuna, so we will report our own reproduction.
We add two cheap variants a reviewer would ask for. The first is HLR with a power-law decay. The second is HLR with a domain-class indicator added to its features, which is the rate-only hypothesis in model form.
The strongest flexible predictor is a gradient-boosted tree model given elapsed time, practice counts and item-class features, with learner IDs excluded to avoid the cold-start failure noted by the SLAM ensemble authors. The file's evidence that this family is strong is the SLAM 2018 ensemble's AUC of 0.861 on en_es in Second Language Acquisition Modeling: An Ensemble Approach. That number is for next-exercise error prediction, not delayed recall, so it is context rather than a target. All models get the same features and a matched hyperparameter search budget.
Two ablations isolate the key ingredients. The first forces $\Lambda$ to be linear, which recovers HLR under the binomial likelihood and separates baseline flexibility from the change of loss. The second uses one shared baseline, which isolates the value of stratification.
We count the proposal as beating the baselines under three conditions. It must reduce held-out log-loss and long-delay calibration error relative to HLR and its variants beyond bootstrap uncertainty. It must stay within about one AUC point of the tree model while producing monotone, interpretable curves. And in the randomized study, it must detect a shape difference that the rate-only HLR variant cannot represent.
Why not just…?
Why not just use gradient-boosted trees with elapsed time as a feature? They will probably predict delayed recall about as well as our model, and we will not pretend otherwise. What they lack is structure. Their implied recall curves need not decrease with delay, they give no estimate of forgetting shape per class, and they offer no test of proportionality. We will measure how often the tree model's implied curves are non-monotone and how its schedules compare in the burden simulation. If it matches our model on scheduling too, the case for the survival model rests only on interpretability and on the randomized shape finding.
Why not just swap HLR's exponential for a power law? That tests a second fixed shape, not the question of whether shapes differ across kinds of vocabulary. Our spline baseline nests both forms, so the likelihood-ratio tests will show directly whether a single power law suffices. If it does, and the class ratio is flat, we will recommend power-law HLR and report that.
Why not just add a "technical term" indicator to HLR? This is the most serious alternative, because it is exactly right if domain anchoring only changes the forgetting rate. The proportionality test in the randomized study decides it. A flat log-ratio of class baselines means this alternative wins. A sloped ratio means it cannot express what is happening.
Why not just fit a standard Cox model or DeepSurv to time until forgetting? No time until forgetting is ever observed, only recall at inspection times. Treating the first failed review as the event time pushes every lapse to the scheduler's chosen review date and biases the curve toward whatever the scheduler did. The current-status likelihood is the correct one. It also makes the partial-likelihood scaling machinery of FastSurvival unnecessary.
Proposed Experiments
The first experiment validates the method on the Duolingo 12.9M interaction dataset used by HLR. We will label word types as technical or general using an EduCSW-style term-identification pass. General courses contain few technical terms, so the aim here is to show that the flexible baseline improves held-out log-loss and calibration in the long-delay bins relative to the linear-baseline ablation, and to report the fitted baseline shape. Metrics are mean absolute error on recall rate (for comparability with HLR), AUC, log-loss, and reliability curves by delay bin.
The second experiment checks transfer on the SLAM 2018 en_es, es_en and fr_en tracks. We build current-status observations from days since a token was last seen and add our predicted recall as a feature to the gradient-boosted model. If the feature raises AUC, the curve captures information the snapshot models miss.
The third experiment is decisive: a prospective randomized retention study in a specialized-language course, for example veterinary or business English or a Chinese+ technical module. Each learner meets about 60 technical terms. Half are taught inside narrow-context professional scenarios following the phased ESP design; half are taught as glossed lists. Assignment of terms to conditions is counterbalanced across learners. Each term is tested exactly once per learner, at a delay drawn at random from 1 to 60 days, so the independence assumption holds by design.
With roughly 200 learners this yields about 12,000 independent current-status observations, enough to estimate two class baselines and the slope of their log-ratio. The pre-registered outcome is whether that slope differs from zero. If it does, domain-anchored and list-learned terms need different review schedules. If it does not, we will report that anchoring's benefit, if any, is a pure rate effect captured by HLR with an indicator.
The fourth experiment converts the fitted curves into schedules. In a simulation seeded with the study's estimates, we compare reviews needed to hold 90% recall over 100 days against HLR-based and tree-based schedules. The comparison point is LECTOR's reported 90.2% recall success in LECTOR. That figure comes from a synthetic benchmark, so it serves as context, not a like-for-like target.
References
- A Trainable Spaced Repetition Model for Language Learning
- FastSurvival: Hidden Computational Blessings in Training Cox Proportional Hazards Models
- DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills
- Second Language Acquisition Modeling: An Ensemble Approach
- Context Based Approach for Second Language Acquisition
- LECTOR: LLM-Enhanced Concept-based Test-Oriented Repetition for Adaptive Spaced Learning
- Research on Narrow Context ESP Course Design in Adult Education
- An Application of the Input Hypothesis to the Teaching Design of English Reading Class in Senior High School
- EduCSW: Building a Mandarin-English Code-Switched Generation Pipeline for Computer Science Learning
- DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network
- Generating Survival Times Using Cox Proportional Hazards Models with Cyclic and Piecewise Time-Varying Covariates
- A Federated Cox Model with Non-Proportional Hazards