📄 Smart PV: A Software Application for Managing English Phrasal Verbs

Summary

Phrasal verbs (PVs) are verbs combined with one or two particles that shift the verb's usual meaning. Their polysemy and flexible word order make them hard for non-native readers and translators. The authors first build a characterization of English PVs that extends the scheme of Lavin and Sánchez (1989). It has three main groups: adverbial, prepositional and adverbial-prepositional. These are subdivided by:

• whether a direct object (DO) is present,
• where the DO sits relative to the particle,
• whether passive transformation (PT) is possible,
• whether the DO coincides with the prepositional object (PO).

Further features cover interchangeable or optional particles, reflexivity, register, regional variety and nominal forms. This scheme drives Smart_PV, a web application with two parts. Guided forms let users enter PVs into a database, and a detector scans text by matching inflected verb forms followed by the stored particles. Validation used lexicons of 40, 60 and 80 PVs over 60 texts, and detection coverage improved as the lexicon grew. The tool is pitched as support for teaching, translation and text mining. Choosing the right sense in context is left as future work.

Key points

• The classification has groups A (adverbial), B (prepositional) and C (adverbial-prepositional). They split into subgroups such as A-1a/A-1b/A-1c (DO movable, before, or after the particle), B-1a/B-1b and C-1a/C-1b (DO matches the PO or not), and A-2/B-2/C-2 (no PT). Extra features are labelled D–I: equivalent particles, optional particles, reflexive, style (formal/casual/slang/taboo), geographic context (American/British/Australian English) and nominal transformation into phrasal names.
• One verb-particle combination (VPC) can yield several PVs with different classes. For example, "go off" is A-2 in one sense and B-2 in another.
• Detection works in three steps:
  • each text word is compared against five stored verb forms: infinitive, past, past participle, gerund and third person;
  • the following words are then checked against the first and, if needed, the second particle;
  • the particle search stops at a punctuation mark or at another verb.
• The evaluation applied a lexicon of at most 80 PVs to 60 texts. That lexicon contained the 25 PVs that Trebits (2009) identified as most frequent in European Union documents. It reported three rates: search efficiency (detected/total PVs), search deficiency (missed/total PVs) and search optimization (1 − false VPC hits/all detected VPCs). Ten of the most frequent detected PVs also appear in Trebits's top 25.
• The database schema stores four attributes: Type (A/B/C), PT (yes/no), Subtype (particle position for A; DO=PO for B/C) and Reflexive.

Interesting ideas

• I1. Store passivizability as a lexical diagnostic on each PV entry, since whether "the picnic was called off" works signals that the object is a true argument of the verb+particle unit.
• I2. Separate the surface verb-particle combination from its sense entries, each with its own syntactic class, so one string like "go off" maps to several PVs with different argument structures.
• I3. Distinguish whether the direct object is the same as the prepositional object ("account for her absence") or a separate argument ("take him for a blind man"). This is a compact way to encode one-object versus two-object PV frames.
• I4. Limit the particle search window by stopping at punctuation or the next verb, a cheap guard against linking a verb to a distant, unrelated particle.
• I5. Track a false-positive rate (detected combinations that are not PVs) alongside coverage as the lexicon grows, so lexicon expansion is judged on precision as well as recall.

For Mova

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

• I1 (frames): VerbNet-based role assignment could consult a PT flag to decide whether the PV's object should get a core Patient/Theme role, including for passive sentences.
• I2 (world): the compiled verb-class lexicon could key multiword verbs by form and list senses with distinct frames, letting the frame stage choose by observed syntax (intransitive vs. prepositional object).
• I3 (frames): this flag could tell the role mapper whether the pobj of a prepositional verb fills a core role or an additional one, making argument linking explainable.
• I4 (tree): it could serve as a sanity check or fallback when the parser's compound:prt / obl attachment of a particle crosses a clause or verb boundary, flagging suspicious attachments for repair.
• I5 (learn): when growing multiword-verb lists from silver data, report both rates per lexicon size to catch expansions that add noise.