PhoniTale: Cross-Lingual Mnemonic Generation

• PhoniTale is a cross-lingual mnemonic system that generates Korean cues for English words using explicit phonological processing and neural transliteration.
• It employs a retrieval-first architecture that adapts English phonetics to Korean constraints before using an LLM for concise, semantically coherent cue generation.
• Evaluation shows PhoniTale outperforms baselines in phonetic similarity and context completeness, though it faces challenges with lexical flexibility and long-term retention.

PhoniTale is a cross-lingual mnemonic generation system for native Korean speakers learning English vocabulary. It operationalizes the classic keyword mnemonic method in a setting the literature had largely neglected: typologically distant language pairs in which phonological and orthographic correspondences are weak. Rather than asking a LLM to invent sound-alike cues freely, PhoniTale first derives Korean keyword sequences through explicit phonological processing, then uses an LLM only to verbalize those retrieved keywords as a short Korean mnemonic cue. The system was introduced in “PhoniTale: Phonologically Grounded Mnemonic Generation for Typologically Distant Language Pairs” (Kang et al., 7 Jul 2025).

1. Problem setting and linguistic motivation

PhoniTale addresses vocabulary acquisition for Korean-native learners of English, a direction that earlier automated keyword-mnemonic work had largely neglected. The paper argues that this direction is harder than the more familiar case of native English speakers learning other languages, because English and Korean are typologically distant in both phonology and orthography. Earlier automated systems such as TransPhoner, SmartPhone, and overgenerate-and-rank methods using LLMs were developed largely for language pairs with greater phonological overlap and often for English-speaking learners (Kang et al., 7 Jul 2025).

The system is grounded in the classical keyword mnemonic paradigm: for a target second-language word, one finds a familiar first-language word or phrase that sounds similar, then constructs a memorable scene linking that sound-alike cue to the target meaning. PhoniTale formalizes this for English-to-Korean mnemonic generation. In the paper’s notation, the input is an English target word $w_{\mathrm{L2}}$ and its meaning $\ell$; the output is a Korean keyword sequence $\mathcal{W}_{\mathrm{L1}}$ approximating the English pronunciation and a Korean verbal cue $c^*$ embedding those keywords in a short mnemonic expression (Kang et al., 7 Jul 2025).

The paper identifies four major mismatches that make this task difficult. English uses linear orthography, whereas Korean uses 2D syllable blocks. Korean disallows consonant clusters and therefore inserts epenthetic vowels when adapting English forms. Some English phones such as /θ/ lack direct Korean counterparts. The stop systems also differ: Korean uses a three-way lenis/fortis/aspirated distinction, whereas English uses a voiced/voiceless contrast with aspiration as an allophone. These mismatches mean that there is often no single Korean word that cleanly approximates an English target, and that English phonological structure must be reshaped before it can become a plausible Korean cue (Kang et al., 7 Jul 2025).

A recurrent misconception is to treat PhoniTale as simply an LLM-based mnemonic generator. The system is explicitly designed against that assumption. Its defining claim is that English–Korean mnemonic generation must be phonologically grounded before verbal creativity becomes useful, because naive phonetic matching and purely LLM-based keyword invention tend to break down in this cross-lingual setting (Kang et al., 7 Jul 2025).

2. Retrieval-first architecture and phonological grounding

PhoniTale has two phases: keyword sequence retrieval and verbal cue generation. The retrieval stage begins with an English phoneme sequence $P_{\mathrm{L2}}$. In the architecture example, the word “squander” is converted to IPA with the eng-to-ipa library, which is based on the CMU Pronouncing Dictionary. That English phoneme string is then transliterated into a Korean-adapted phoneme sequence $\widehat{P}_{\mathrm{L1}}$ by a neural seq2seq model with attention: a bidirectional LSTM encoder plus a unidirectional LSTM decoder, each with 256 hidden units, following Bahdanau-style attention. The encoder yields a 512-dimensional contextual representation. Training combines token prediction with a contrastive consistency term, and the total loss uses $\lambda = 0.1$ (Kang et al., 7 Jul 2025).

After transliteration, the system performs syllable-aware segmentation. Because Korean listeners are highly sensitive to syllable structure, the transliterated string is syllabified under Korean constraints rather than treated as a flat sequence. The boundary predictor is another bidirectional LSTM with 256 hidden units per direction, operating on embedded IPA tokens augmented with binary vowel masks. It predicts a syllable sequence $$\widehat{\boldsymbol{\sigma}} = (\sigma_1, \sigma_2, \ldots, \sigma_l)$$. For “autopsy,” the Korean-adapted /otʰapsi/ is segmented into [o, tʰap, si]. A second stage groups these syllables into at most two segments $(S_1, S_2)$ for longer English words. This upper bound is deliberate: one keyword per syllable is too fragmented in this setting, whereas one or two Korean keywords better balance phonetic coverage and cue coherence (Kang et al., 7 Jul 2025).

Keyword matching is performed against a Korean lexicon drawn from the Basic Korean Dictionary dataset (basic_korean_dict). Each segment $S_i$ and each candidate Korean word are mapped to PanPhon 22-dimensional phonological feature vectors. Segment–keyword similarity is defined as

$\ell$0

where $\ell$1 adds hand-tuned bonuses for syllable overlap, initial-syllable match, early-phone alignment, and substring inclusion. Initial-syllable alignment receives the strongest weight. The tuned values are $\ell$2, $\ell$3, $\ell$4, and $\ell$5. The best keyword for each segment is selected by

$\ell$6

and candidate segmentations are ranked by the average best segment-level similarity (Kang et al., 7 Jul 2025).

Only after retrieval does the LLM enter the pipeline. The second phase adapts the overgenerate-and-rank framework of Lee et al. (2024), but simplifies it for English–Korean. First, it removes the two-step “generate a story, then summarize it” process and uses direct generation. Second, it drops Age-of-Acquisition ranking, arguing that L2 AoA does not transfer well to cross-lingual familiarity. The prompt instructs the model to generate one short Korean sentence that contains the retrieved keywords exactly in the provided order, includes the target word exactly once marked with angle brackets, and returns the result in JSON. Generated cues are ranked only by context completeness: the target word is masked, GPT-4o proposes five probable fills, and the average cosine similarity between FastText embeddings of those guesses and the true target word is computed using FastText embeddings trained on naver_shopping_corpus (Kang et al., 7 Jul 2025).

This retrieval-first decomposition is the paper’s central design principle. English is first adapted to Korean phonology, then syllabified under Korean constraints, then matched to real Korean words, and only then verbalized. The paper contrasts this with OGR-style systems that let the LLM generate many keywords and then rank them. PhoniTale instead constrains the LLM to sentence construction, reducing hallucinated nonwords and preserving lexical validity (Kang et al., 7 Jul 2025).

3. Data resources and implementation context

The transliteration module was trained on 2,870 English–Korean word pairs with aligned IPA transcriptions. English words were drawn from Magoosh GRE and Princeton Review materials. Korean transliterations were extracted from the AHA Dictionary, converted to Korean IPA with hangul_to_ipa, manually verified, and syllabified. For comparative mnemonic evaluation, the paper used 36 vocabulary words sampled from the KSS vocabulary book, a Korean mnemonic vocabulary book aimed at advanced standardized tests (Kang et al., 7 Jul 2025).

The pipeline integrates several external resources. English pronunciations are obtained with eng-to-ipa backed by CMUdict. PanPhon supplies articulatory feature vectors. KoGPT2-base-v2 is used for perplexity estimates of Korean sentence fluency. GPT-4o is used for cue generation, context-completeness probing, and LLM-as-judge correctness scoring. EXAONE3.5 is tested as an open alternative in ablation. Candidate keywords are retrieved from the Basic Korean Dictionary dataset, and FastText embeddings are trained on naver_shopping_corpus for semantic ranking of cue contexts (Kang et al., 7 Jul 2025).

The learner-facing evaluation environment was implemented as a custom web platform with learning, testing, and feedback phases. Participants completed three rounds of 12 English words each, for 36 total items. In the learning phase, each trial displayed the English word, Korean meaning, audio at 2 s and 7 s, the Korean keyword sequence, and the verbal cue, with color underlining for phonological alignment and a 30-second limit. Testing consisted of recognition and generation tasks, followed by Likert-scale ratings of helpfulness, coherence, and imageability (Kang et al., 7 Jul 2025).

The system’s overall methodological position is therefore hybrid rather than monolithic. It combines symbolic phonological constraints, neural transliteration, lexicon-constrained retrieval, and LLM verbalization. This suggests a broader design pattern for mnemonic generation in typologically distant settings: phonological plausibility is treated as an intermediate computational object rather than as a post hoc filter (Kang et al., 7 Jul 2025).

4. Evaluation results and empirical profile

Module-level validation supports the architecture. The seq2seq transliteration model achieved a Character Error Rate of 3.95% and an Exact Match Rate of 75.56% on the training dataset. The paper also reports that syllabification reached a boundary-level $\ell$7 of 1.00 against manual annotation on transliteration outputs. For keyword matching, the paper presents qualitative ablations showing that cosine similarity alone often produces poorer syllabic matches than the full structural scoring function. For example, with “demolish,” cosine-only matching retrieved keywords with IPA /pi.mil.li/ and /Swi/, whereas the full scoring function retrieved /tE.mul.lim/ and /Swi/, which better preserved Korean-perceived syllable structure (Kang et al., 7 Jul 2025).

The main automatic evaluation compares KSS, OGR, and PHT. Keyword sequences are evaluated by phonetic similarity, keyword omission, and keyword modification; verbal cues are evaluated by context completeness and perplexity. KSS achieved phonetic similarity 0.74, omission 3.7%, context 0.38, and perplexity 553.92. OGR achieved phonetic similarity 0.86, omission 3.4%, modification 24.8%, context 0.29, and perplexity 691.01. PHT achieved phonetic similarity 0.95, omission 0%, modification 3.3%, context 0.39, and perplexity 433.41. On these automatic metrics, PHT is best on every reported measure (Kang et al., 7 Jul 2025).

The verbal-generation ablations reinforce the prompt design. Replacing OGR’s story-then-summary prompt with PhoniTale’s direct-generation prompt improved context completeness from 0.29 to 0.39 and reduced perplexity from 490.13 to 433.41, holding other components fixed. Replacing GPT-4o with EXAONE3.5 yielded context 0.38 versus 0.39 and perplexity 450.30 versus 433.41, indicating that an open Korean-capable model can approximate the stronger proprietary model in this setting (Kang et al., 7 Jul 2025).

Human evaluation is more nuanced. Recruitment began with 167 Korean-native adults with intermediate English proficiency; after screening and filtering, 51 participants remained, evenly assigned to KSS, OGR, and PHT, with 17 in each condition. Appendix statistics reported no significant group differences in vocabulary familiarity ($\ell$8), age ($\ell$9), or education level ($\mathcal{W}_{\mathrm{L1}}$0). Correctness was scored with GPT-4o as judge rather than Levenshtein distance because recognition answers allowed semantic variation and generation answers contained typos (Kang et al., 7 Jul 2025).

In recognition, KSS had the highest mean correctness, followed by PHT and OGR, but no between-group differences were statistically significant. In generation, PHT achieved the highest correctness, followed by KSS and OGR. PHT significantly outperformed OGR with $\mathcal{W}_{\mathrm{L1}}$1, and the difference between PHT and KSS was not statistically significant. The paper treats this as its strongest practical result: PhoniTale matches human-authored mnemonics on short-term recall while surpassing a prior automated baseline, especially when learners must actively produce the English target (Kang et al., 7 Jul 2025).

Subjective preference diverged from recall outcomes. KSS received mean ratings of 3.50 for helpfulness, 3.64 for coherence, and 3.68 for imageability. OGR received 2.35, 2.33, and 2.41. PHT received 2.40, 2.26, and 2.44. KSS was significantly preferred over both automated systems on all three dimensions ($\mathcal{W}_{\mathrm{L1}}$2). PHT was numerically but not significantly above OGR on helpfulness and imageability, whereas OGR was numerically but not significantly above PHT on coherence. The paper interprets this dissociation as evidence that mnemonics judged more elegant or imaginative are not always the most effective for recall (Kang et al., 7 Jul 2025).

System	Automatic evaluation	Human evaluation
KSS	Phonetic 0.74; Omission 3.7%; Context 0.38; Perplexity 553.92	Recognition highest mean correctness; no significant advantage over PHT; Helpfulness/Coherence/Imageability 3.50/3.64/3.68
OGR	Phonetic 0.86; Omission 3.4%; Modification 24.8%; Context 0.29; Perplexity 691.01	Lowest generation correctness; PHT significantly better at $\mathcal{W}_{\mathrm{L1}}$3; 2.35/2.33/2.41
PHT	Phonetic 0.95; Omission 0%; Modification 3.3%; Context 0.39; Perplexity 433.41	Generation highest; not significantly different from KSS; 2.40/2.26/2.44

5. Qualitative behavior, strengths, and failure modes

The paper’s qualitative examples illustrate the intended retrieval-first behavior. In the appendix architecture example, English /skwɑndər/ for “squander” is transformed into Korean-adapted /sWkʰwantə/, syllabified, then segmented into /sWkʰwan/ and /tə/. The retrieved Korean keywords are approximately /sæ.gwan/ (“customs”) and /tə/, and the generated cue is translated as “Wasted more time at customs.” The example shows the system’s preferred operating regime: a short Korean sentence built from real Korean words whose sound contour approximates the English target and whose scene evokes the target meaning (Kang et al., 7 Jul 2025).

A more explicit contrast appears with “frivolous.” OGR retrieves three keywords corresponding roughly to “blue,” “field,” and “Ross,” yielding a cue like “The reckless woman who was scolded by Ross in the blue field,” with perplexity 689.3. PHT retrieves two keywords corresponding to “flute” and “ballad,” producing “He, who was singing a ballad in time with the sound of the flute, acted rashly,” with perplexity 231.0. The paper uses this contrast to argue that limiting retrieval to at most two segments improves coherence relative to syllable-by-syllable decomposition (Kang et al., 7 Jul 2025).

Other examples show the system’s operating range: “reckon” yields keywords “레게, 컨셉”; “demolish” yields “대물림, 쉬”; and “reimburse” yields “레임덕, 스무.” Across these cases, the paper identifies a recurring pattern: successful mnemonics preserve onset alignment, use only one or two keywords, and embed those keywords in a semantically pointed scene. This suggests that phonological fidelity and cue coherence are not independent objectives but coupled constraints on mnemonic usefulness (Kang et al., 7 Jul 2025).

The failure modes are equally explicit. Because candidate keywords come from a fixed dictionary, retrieval may be forced toward awkward nouns, rare words, or culturally specific names when the lexicon lacks a flexible Korean item. The case study highlights “felon” and “mayhem” as words where KSS outperformed PHT. For “felon,” PHT retrieved “Peleus” and “London,” producing a more awkward sentence, whereas KSS used a more flexible cue based on a conjugated verb plus noun. The authors attribute such errors to the dictionary’s restriction to base forms, which limits inflectional and part-of-speech flexibility available to human writers and freer LLM baselines (Kang et al., 7 Jul 2025).

The paper also identifies conditions under which PHT is especially strong. For English words containing /r/, such as “reckon” and “render,” PhoniTale often retrieves Korean /l/-initial keywords, which are closer than KSS choices beginning with /n/. Because onset and syllable structure are weighted heavily in the retrieval score, the system can sometimes produce stronger phonological approximations than human-authored mnemonics (Kang et al., 7 Jul 2025).

6. Limitations, significance, and future directions

PhoniTale’s limitations are stated directly. The system is validated only on English–Korean, and the transliteration module would require language-specific adaptation for other phoneme inventories and phonotactic systems. Evaluation measures only short-term recall; there is no delayed post-test, so long-term retention remains unknown. The vocabulary comes from advanced standardized-test materials, leaving open whether the method generalizes to beginners or everyday vocabulary. The Korean dictionary contains mainly base verb forms rather than rich inflectional paradigms, which reduces lexical diversity and naturalness (Kang et al., 7 Jul 2025).

These constraints qualify the scope of the paper’s claims. PhoniTale is not presented as a universally transferable mnemonic engine, nor as proof that LLMs alone can solve cross-lingual mnemonic design. Rather, it demonstrates that explicit phonological adaptation, Korean-aware syllabification, lexicon-based retrieval, and constrained verbalization can produce automated mnemonics that are competitive with human-authored ones in a difficult language-pair setting. This suggests that mnemonic generation for typologically distant languages benefits from treating learner phonological perception as a first-class modeling target rather than as an incidental by-product of prompt engineering (Kang et al., 7 Jul 2025).

The paper’s future directions follow directly from its limitations: add flexible part-of-speech handling in retrieval, extend the method to more language pairs, and run longitudinal studies to test lasting retention. More broadly, PhoniTale contributes a methodological alternative to fully LLM-driven mnemonic generation. It decomposes the task into phonological adaptation, Korean-aware syllabification, structural lexical matching, and constrained verbalization. In that sense, its significance lies less in a single model component than in the claim that robust mnemonic generation for difficult language pairs requires explicit phonological mediation between source-word form and learner-usable cue (Kang et al., 7 Jul 2025).