EnglishFrench Verb Phrase Alignment

1. Structural Overview

Both languages convey tense, aspect, mood, voice, and modality within the verb phrase, but the surface order and the distribution of auxiliaries differ.

1.1 Basic word order

• English: Subject (aux) V (object) (adverbial)
• French: Subject (aux) V (object) (adverbial)

While the canonical order appears similar, French often places adverbial pronouns before the verb and uses a larger set of clitic pronouns that occupy a fixed preverbal slot.

1.2 Auxiliaries and compound tenses

Both languages employ auxiliary verbs be (English) / tre (French) and have (English) / avoir (French) to form perfective aspects. However, French also uses tre as the auxiliary for many intransitive verbs, especially verbs of motion and reflexive constructions.

1.3 Passive voice

English passive uses be + past participle. French passive uses tre + past participle and agrees in gender and number.

2. Common Alignment Challenges

1. Different auxiliary selection French may use tre where English uses have (e.g., She has gone Elle est partie).
2. Pronoun placement French enclitic pronouns precede the verb, while English places objects after it. Alignment must reorder constituents.
3. Verbnoun blends French infinitive constructions (e.g., je veux manger) correspond to English I want to eat. The infinitive functions as a complement rather than a full VP.
4. Aspectual nuances The English progressive be + -ing often translates to a simple present in French, or to tre en train de. Deciding which mapping to use depends on context.
5. Negation English places not after the auxiliary; French surrounds the auxiliary with ne pas. This changes the linear span of the VP.

3. Alignment Strategies

3.1 Constituentbased alignment

Identify the minimal VP constituent in each language and align them regardless of internal word order. This works well for parallel corpora where tokenlevel alignment is not required.

3.2 Wordlevel alignment with reordering

Map each element (auxiliary, main verb, modal, negation particles, clitics) to its counterpart, inserting NULL where no direct equivalent exists. Tools such as GIZA++ or fast_align can be trained with specially crafted alignment templates.

3.3 Semanticrole alignment

Align VPs based on the roles they encode (Agent, Patient, Experiencer, etc.). This abstracts away from syntactic divergence and is valuable for downstream tasks like AMR or predicateargument extraction.

3.4 Hybrid approach

Combine constituentbased alignment for highconfidence pairs (e.g., simple present tense) with wordlevel alignment for complex constructions (passives, reflexives). Machinelearning models can be trained to select the appropriate strategy per sentence.

4. Practical Guidelines for Manual Annotation

• Mark the VP span in both languages. Include auxiliaries, negation particles, and clitics.
• Document auxiliary mismatches (e.g., have tre). Record the mapping in a separate field.
• Handle clitic pronouns as part of the VP in French, even though they precede the verb.
• Note aspectual differences. When English progressive is rendered by a simple present, add a comment indicating the implied progressive meaning.
• Align negation markers as a unit with the auxiliary, because French splits nepas.

Example annotation:

5. Automatic Extraction Techniques

Modern NLP pipelines can extract VPs using dependency parsing. Below is a typical workflow:

1. Parse both English and French sentences with a universal dependency parser (e.g., spaCy, UDPipe).
2. Identify the aux, aux:pass, neg, compound:prt, and obj relations linked to the main verb (root).
3. Collect all tokens reachable via these relations to form the VP span.
4. Apply a bilingual lexicon or pretrained crosslingual embeddings to align verbs and auxiliaries.
5. Use a reordering model (sequencetosequence or transductive alignment) to map French clitic order to English order.

Evaluation metrics such as Alignment Error Rate (AER) and Span F1 can assess the quality of the extracted alignments.

6. Resources and Further Reading

• Universal Dependencies English & French treebanks (UD EnglishEWT, UD FrenchGSD)
• VerbNet & Lefff lexical resources for verb subcategorization
• Crosslingual Predicate Alignment ACL 2021 paper by Smith etal.
• Opensource tools: spaCy, UDPipe, Moses for alignment models.

These resources provide annotated data, lexical information, and baseline models useful for building robust EnglishFrench VP alignment systems.