📄 Improving Automated Prediction of English Lexical Blends Through the Use of Observable Linguistic Features

Summary

Lexical blends such as combinations of two words into one (keeping the start of the first and the end of the second) are notoriously hard to predict. Earlier computational attempts, a multi-tape FST by Deri & Knight (2015) and a character-embedding LSTM by Gangal et al. (2017), got fewer than half of blends right.

This paper takes a different route. It enumerates every candidate that joins a prefix of word 1 with a suffix of word 2, cutting only at grapheme/phoneme alignment points. Each candidate is described by interpretable phonological features based on Arndt-Lappe & Plag (2013), plus a BLICK phonotactic score. Simple learners (LASSO, second-order Polynomial regression, Random Forest) assign each candidate a probability, and the highest-scoring candidate is chosen.

Across three corpora, the Polynomial model with the full feature set clearly beats the previous benchmarks, reaching up to 74% exact matches. The work shows that explicit, observable linguistic features can outperform latent-space neural models on rare, irregular word-formation processes.

Key points

• Data came from three corpora, filtered to blends that keep the start of word 1 and the end of word 2: Deri & Knight (322 blends), Gangal et al. (1092), and Shaw (2014), curated from Thurner's 1993 dictionary (1096). All trials used 10-fold cross-validation.

• Features cover:

  • syllable lengths and medial overlap;
  • distances from word edges to primary stress and to the switchpoint;
  • whether the switchpoint falls on an onset, nucleus or coda boundary;
  • preservation of stress, segments and syllables;
  • the BLICK markedness score, a MaxEnt grammar trained on the CMU pronouncing dictionary.

• Results for the best learner (Polynomial, full features), measured by exact match and Levenshtein edit distance:

  Corpus	This paper	Previous benchmark
  Deri & Knight	64.42% (edit distance 0.72)	45.39% (edit distance 1.59)
  Gangal et al.	59.51%	48.75%
  Shaw	74.13%	—

• Only LASSO failed to beat the LSTM benchmark on the Gangal corpus. The full feature set beat a hand-pruned, low-collinearity subset, even though the full set is highly collinear.

• Error analysis of 100 wrong outputs showed a bias toward keeping too much material. Word 1 was over-preserved 40 times versus cut too short 15 times; for word 2 the figures were 33 versus 18.

Interesting ideas

• I1. Instead of decoding a form step by step, enumerate every structurally allowed candidate and let an interpretable scorer rank them; the candidate space itself encodes linguistic constraints.
• I2. Candidates whose feature vectors are identical to an earlier candidate are discarded, so the learner never sees indistinguishable positive and negative examples.
• I3. A separate phonotactic plausibility score (BLICK) serves as one feature, so the ranker can reject sound sequences that are possible in spelling but un-English.
• I4. Errors are split by direction (too much versus too little material kept), which exposes a systematic length bias rather than just a lower accuracy number.
• I5. Comparing model outputs with human speakers' brand-new coinages is proposed as truly held-out data, useful even for testing large language models that may have seen existing blends.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): Unknown words could be analysed by generating all prefix+suffix splits against known lemmas and ranking them with explicit features, giving an explainable guess at what a novel blend is made of.
• I2 (learn): When training rankers on silver data, removing feature-identical duplicates with conflicting labels would cut noise and make rule induction cleaner.
• I3 (back): When generating names or novel word forms, a phonotactic well-formedness score could filter out unpronounceable outputs.
• I4 (learn): Evaluation reports could classify errors by direction, such as over- versus under-attachment or spans that are too long versus too short, to show which repair rules are needed.
• I5 (learn): Freshly coined test items could act as a contamination-free control when comparing Mova against its LLM teacher tiers.