📄 Detecting Intentional Lexical Ambiguity in English Puns

Summary

The paper addresses automatic analysis of puns, where a single word deliberately carries two meanings at once. It covers two tasks: deciding whether a sentence is a pun, and locating the ambiguous target word. The authors assume every pun contains two overlapping groups of words, each pointing to a different abstract meaning. To find these groups, they map each word, and each collocation found through part-of-speech patterns, to the 34 Sections of Roget's Thesaurus. A sentence then becomes a 34-dimensional count vector. An SVM with an RBF kernel, trained on 2,480 collected puns against ordinary sentences drawn from NLTK corpora, aphorisms and proverbs, reaches an F-measure of about 0.73. For target location they compare rule-based methods: a sense-based method that uses the overlap of the two dominant Sections, and a positional "last word" rule. The positional rule often wins, possibly only because of how the dataset is built. The system, PunFields, took part in SemEval-2017 Task 7 and did best on pun classification. The work shows that a small, readable semantic feature space can capture deliberate ambiguity.

Key points

• Resource: Roget's Thesaurus, used at the level of its 34 Sections (subdivisions of its six Classes). Thesaurus collocations such as idioms and proverbs are also looked up, using POS patterns like verb+particle and adjective+noun.
• Data: 2,480 puns gathered from the Internet, contrasted with random 5–25-word sentences. Two contrast sets were used: "Mixed styles" (Brown, Reuters, Web) and "Belles lettres" (Gutenberg plus aphorisms and proverbs). Each pun was tested against one of these sets, split into halves for training and testing.
• Classification: Scikit-learn SVM with an RBF kernel gave the best and most balanced results. Pun F-measure was about 0.73, and all models scored higher against the literary contrast set. The authors attribute part of this to puns sharing topics, for example recurring names such as John Deere. They probed this by sorting the vector values to remove which-Section information.
• Location: the "last word" heuristic was strong for homographic puns. For heterographic puns it scored 0.39, against a SemEval baseline of 0.57. PunFields' current location precision is 0.5145 for homographic and 0.3879 for heterographic puns.
• Other SemEval winners used different approaches: Duluth (disagreement among WordNet variants), Idiom Savant (a combination including word2vec), JU_CSE_NLP (HMM plus a dependency network), N-Hance (Wikipedia PMI peaks) and UWaterloo (11 location features).

Interesting ideas

• I1. A sentence can be represented as counts over a few dozen coarse thesaurus categories, which yields a compact feature space where every dimension has a human-readable name.
• I2. Deliberate ambiguity can be modelled as two dominant semantic fields, with the target word being the item that belongs to both (or, for heterographic puns, to only one).
• I3. Multiword units are checked against a lexicon before senses are assigned, because the joke often lives in a fixed phrase like "change one's mind" rather than in a single word.
• I4. Sorting feature values destroys category identity while keeping the distribution, which tests whether a classifier learned topic rather than structure.
• I5. A trivial positional rule (the target sits near the end) proved hard to beat, which shows the value of always reporting dumb structural baselines.

For Mova

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

• I1 (world): Project each sentence onto coarse WordNet or FrameNet domains to get an explainable topic profile, which can sanity-check entity and sense choices in the world of the text.
• I2 (frames): The prag crate could flag puns and wordplay when a token's senses split across the two strongest domains of its context, feeding the irony and implicature layers.
• I3 (tokens): Look up idioms and phrasal verbs as lexical units at lemma level, possibly triggered by UD patterns, so that later sense and frame assignment does not treat their parts literally.
• I4 (learn): Add similar feature-scrambling negative controls to evaluations, revealing when a learned component relies on dataset topic leakage.
• I5 (answer): When evaluating readers on fables or tests, compare against position-only baselines to confirm that answers come from understanding rather than layout cues.