Adjusting Interpretable Dimensions in Embedding Space with Human Judgments

Katrin Erk, Marianna ApidianakiNAACL 2024 · DOI

TL;DR

This paper introduces a novel method to improve "interpretable dimensions" in word embeddings by combining traditional "seed-based" approaches with human judgments. Their FIT+S model significantly outperforms existing methods, especially when seed-based dimensions are weak, by better aligning predictions with human-perceived properties like object size or formality.

What Problem Does This Paper Address?

Word embeddings, the numerical representations of words in machine learning models, often contain "interpretable dimensions" that capture semantic properties like gender, formality, or object characteristics. These dimensions are valuable for understanding how LLMs represent knowledge and for applications in social science and neuroscience. However, the standard method for discovering these dimensions relies on pairs of "seed words" (e.g., "hot" vs. "cold"). This method is simple but can be unreliable, leading to inaccurate or noisy dimensions. The paper addresses the problem of "bad seeds" and the general difficulty in ensuring that these discovered dimensions truly reflect the intended property as understood by humans. They aim to create more robust and accurate interpretable dimensions by incorporating human intuition.

Methods

The paper proposes a new method called FIT+S (Fitted dimensions with seed words and seed dimensions) that augments traditional seed-based dimension extraction.

  • Seed-based (SEED) Model: This is the standard approach. It involves selecting pairs of words that represent opposite ends of a property (e.g., "safe" vs. "dangerous"). The dimension is then computed as the average of the difference vectors between these seed pairs.
  • Fitted (FIT) Model: This approach uses human-labeled data (gold ratings) for a specific property. It learns a dimension in the embedding space such that the scalar projection of a word's embedding onto this dimension is proportional to its gold rating. This model is trained to minimize the difference between predicted and gold ratings.
  • Fitted with Seed Words (FIT+SW) Model: This model extends the FIT model by including the seed words from the SEED method. Synthetic extreme ratings are assigned to these seed words to guide the fitting process.
  • Fitted with Seed Dimensions (FIT+SD) Model: This model combines the FIT approach with the direction of seed-based dimensions. It uses a loss function that penalizes deviations from the seed-based dimension direction.
  • Fitted with Seed Words and Seed Dimensions (FIT+S) Model: This is the proposed core contribution. It combines the FIT approach with both seed words (as in FIT+SW) and seed dimensions (as in FIT+SD) using a weighted loss function. This allows the model to leverage both human ratings and the guidance from seed-based dimensions.

Datasets

The paper utilizes several datasets:

  • Grand et al. (2022) Object Properties Dataset: This dataset contains human ratings for various object properties (e.g., danger, size, wealth) across nine categories like animals, clothing, and professions.
  • Pavlick and Nenkova (2015) Style Datasets: These datasets provide human ratings for word complexity and formality.
  • Pavlick and Nenkova (2015) Paraphrase Judgments: Used to extract seed words for formality and complexity dimensions.

No new benchmarks were introduced; the focus was on improving existing methods using established datasets.

Key Insights and Ideas

  • Human ratings are crucial for improving weak dimensions: The FIT+S model's biggest gains are seen in cases where the traditional seed-based dimensions perform poorly. This suggests that human intuition can effectively correct for the noise and inaccuracies inherent in purely seed-based approaches.
  • Combining multiple sources of information is key: Simply using human ratings (FIT model) can lead to overfitting. Incorporating seed words and seed dimensions provides essential guidance, leading to more robust and accurate interpretable dimensions.
  • FIT+S predictions align with human scales: Unlike seed-based methods that often produce predictions on a different scale than human ratings, the FIT+S model consistently predicts ratings that are on the same scale as the gold standard. This is a significant improvement for direct comparison and interpretation.
  • Contextualized embeddings (BERT/RoBERTa) don't always outperform static embeddings (GloVe) for this task: While contextualized embeddings are powerful, for this specific task of extracting interpretable dimensions, GloVe embeddings, when used with the FIT+S method, often yielded better results.

Key Results

  • Object Properties: The FIT+S model achieved the best performance in terms of extended pairwise rank accuracy (r+-acc), outperforming the SEED model significantly across various object properties. For example, in cases where SEED performed poorly, FIT+S showed an average improvement of 27.3 points in r+-acc.
  • Stylistic Properties (Formality/Complexity): While FIT+S still performed well, the improvement over SEED was less pronounced, likely because SEED dimensions already performed reasonably well on these tasks.
  • Mean Squared Error (MSE): FIT+S consistently produced much lower MSE values compared to SEED, indicating that its predictions were much closer to the gold ratings in scale and magnitude. For instance, 97% of FIT+S runs had MSE values below 2 for object properties, while SEED had many runs with very high MSE.
  • Embedding Spaces: GLoVe embeddings generally performed best when combined with the FIT+S method for both object properties and stylistic features. BERT and RoBERTa models, when used with FIT+SW, sometimes performed at the level of random guessing for object properties.

Limitations and Future Work

  • Language Dependency: The current experiments are limited to English models and data. Extending the methodology to other languages is possible but requires translated human ratings.
  • Averaging Human Ratings: The use of averaged human ratings might obscure legitimate differences between annotators and the effects of context or polysemy.
  • Contextualized Embedding Usage: The paper averaged representations from BERT/RoBERTa across multiple sentences and layers. Exploring different selection strategies or layer combinations could yield different results.
  • Multifaceted Properties: The proposed method currently addresses one property at a time. Future work aims to explore how to construct dimensions for multifaceted properties and how to represent words with multiple interpretable dimensions. This would allow for richer comparisons between words that share some properties but differ in others.

Summary updated

BibTeX
@inproceedings{erk2024adjusting,
  author = {Katrin Erk and Marianna Apidianaki},
  title = {{Adjusting Interpretable Dimensions in Embedding Space with Human Judgments}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.naacl-long.146},
}
Figure