Probing Pre-trained Language Models for Semantic Attributes and their Values

Meriem Beloucif, Chris BiemannEMNLP 2021 · DOI

Abstract

Pre-trained Language Models (PTLMs) like BERT and RoBERTa are central to modern NLP, yet their understanding of the relationship between semantic attributes (e.g., "net worth") and their values (e.g., "rich") remains under-explored. This paper presents a probing framework using Wikidata to evaluate whether these models can link descriptive adjectives to the underlying concepts they measure. Our findings reveal a significant performance gap between PTLMs and humans, highlighting the models' reliance on pattern matching rather than deep semantic reasoning.

The Problem: Attribute-Value Understanding

Humans intuitively link descriptive values to abstract attributes. For example, the term "heavy" is immediately associated with the attribute of "mass," and a "dense" city is understood in terms of its "population." This attribute-value relationship is fundamental to language comprehension.

While PTLMs are trained on massive datasets where these associations occur, it is unclear if they truly "understand" these links. A model that cannot reliably connect "rich" to "net worth" will struggle with complex tasks like comparative Question Answering or automated fact-checking. This study aims to probe the depth of this semantic knowledge.

Probing Methodology

The researchers employed a probing technique using Masked Language Modeling (MLM) with the following steps:

  1. Pattern Generation: 15 sentence templates were created to trigger attribute-value associations, including single-object prompts (e.g., "Gates is rich, thus he has a high [MASK]") and comparative prompts.
  2. Model Probing: BERT, RoBERTa, and XLNet were tasked with predicting the [MASK] token.
  3. Evaluation: Predictions were manually scored based on semantic relatedness to the target attribute.

Dataset and Domains

The authors constructed a dataset of 18,327 English statements extracted from Wikidata, covering five key domains: People, Chemistry, Geography, Politics, and Art.

Table 1: Examples of probing patterns and domain distribution across the dataset.

The dataset includes both standard commonsense statements and "blind" variations to test if the models' logic holds up when entities are swapped.

Key Insights: Semantic Gaps and Brittleness

The study identified several critical weaknesses in PTLMs:

  • The Semantic Gap: Models excel at identifying hypernyms (e.g., knowing gold is a metal) but fail to link those concepts to physical attributes like "weight" in a sentence.
  • Brittleness: Models are highly sensitive to surface-level changes. Changing a city name in a prompt about density can shift predictions from "population" to unrelated terms like "vegetation" or "moisture."
  • Model Performance: RoBERTa generally outperformed BERT, suggesting its training captures more semantic nuance, though it still trails human performance. XLNet performed poorly in this specific MLM setup.

Results: The Human-Model Performance Gap

Quantitative results show a massive gap between human intuition and model performance across all tested domains.

Figure 5: Performance comparison between humans, BERT, and RoBERTa across five domains.

In the Geography domain, RoBERTa achieved approximately 40% accuracy, while BERT scored around 25%. In the People domain, both models performed even lower. Qualitatively, BERT often failed to predict relevant attribute words, guessing unrelated terms depending on the specific entities mentioned in the prompt.

Limitations and Future Work

The results indicate that PTLMs have not internalized abstract attribute concepts and are vulnerable to minor input variations. To improve these models, the authors suggest:

  • Structured Fine-tuning: Training models on semantically rich resources like FrameNet or Knowledge Bases like Wikidata.
  • Contextual Integration: Enhancing how models handle context to ensure semantic logic remains robust across different proper nouns.

Summary updated

BibTeX
@inproceedings{beloucif2021probing,
  author = {Meriem Beloucif and Chris Biemann},
  title = {{Probing Pre-trained Language Models for Semantic Attributes and their Values}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.findings-emnlp.218},
}
Figure