[Paper Review] Assessing Phrasal Representation and Composition in Transformers

Lang Yu, Allyson Ettinger|arXiv (Cornell University)|Oct 8, 2020

Topic Modeling54 references4 citations

TL;DR

This paper systematically evaluates phrasal representation and composition in state-of-the-art transformers using controlled tests that separate lexical effects from true compositional meaning. It finds that while models strongly reflect word content, they show little evidence of nuanced phrase-level composition beyond lexical overlap, with performance dropping significantly when word overlap is controlled.

ABSTRACT

Deep transformer models have pushed performance on NLP tasks to new limits, suggesting sophisticated treatment of complex linguistic inputs, such as phrases. However, we have limited understanding of how these models handle representation of phrases, and whether this reflects sophisticated composition of phrase meaning like that done by humans. In this paper, we present systematic analysis of phrasal representations in state-of-the-art pre-trained transformers. We use tests leveraging human judgments of phrase similarity and meaning shift, and compare results before and after control of word overlap, to tease apart lexical effects versus composition effects. We find that phrase representation in these models relies heavily on word content, with little evidence of nuanced composition. We also identify variations in phrase representation quality across models, layers, and representation types, and make corresponding recommendations for usage of representations from these models.

Motivation & Objective

• To investigate whether state-of-the-art transformers perform sophisticated composition of phrase meaning beyond simple word content.
• To distinguish between representations driven by lexical overlap versus genuine compositional meaning in phrase embeddings.
• To analyze variations in phrasal representation quality across models, layers, and representation types (e.g., CLS, Avg-Phrase).
• To provide practical recommendations for selecting optimal representations from these models based on intended use (e.g., word content fidelity vs. compositional meaning).

Proposed method

• Adapted phrase similarity and paraphrase classification benchmarks using human judgments of phrase similarity and meaning shift.
• Created controlled versions of datasets that remove cues from word overlap to isolate compositional effects from lexical effects.
• Applied classic sense selection tests (e.g., Kintsch, 2001) to evaluate phrasal composition at the phrase level.
• Conducted layerwise analysis across multiple models: BERT, RoBERTa, DistilBERT, XLNet, and XLM-RoBERTa.
• Evaluated various representation types: [CLS], [SEP], [CLS] token, and average of wordpiece embeddings (Avg-Phrase).
• Compared results between uncontrolled and controlled settings to isolate the contribution of lexical content from compositional meaning.

Experimental results

Research questions

• RQ1 To what extent do transformer representations reflect true compositional meaning of two-word phrases, independent of lexical overlap?
• RQ2 How does the quality of phrasal representation vary across different models (e.g., BERT, RoBERTa, DistilBERT) and layers?
• RQ3 Which representation type (e.g., CLS, Avg-Phrase, head word) best captures phrase meaning or word content?
• RQ4 Do landmark composition tests (e.g., sense selection) reliably measure compositional understanding, or are they confounded by lexical cues?
• RQ5 Can we identify specific model configurations (model, layer, representation type) that best support faithful phrasal representation for downstream tasks?

Key findings

• When word overlap is not controlled, models show strong correlation with human judgments of phrase similarity and high performance on paraphrase classification.
• After controlling for word overlap, performance on similarity and classification tasks drops substantially, indicating that model success is largely driven by lexical content rather than compositional meaning.
• RoBERTa’s Avg-Phrase representation in later layers and XLM-RoBERTa’s Avg-Phrase in later layers (when phrase is in sentence context) show the most promising signs of compositional sensitivity.
• DistilBERT’s CLS token in the final layer shows the strongest performance among CLS representations, though still limited in compositional fidelity.
• CLS tokens are generally poor for phrase representation, even in the final layer, except for DistilBERT.
• The landmark sense selection test shows results that parallel uncontrolled similarity tasks, suggesting it may also be influenced by lexical effects rather than true composition.