TL;DR
This paper introduces new evaluation tasks to measure visual commonsense knowledge in language models, specifically investigating if multimodal training (text + images) improves this knowledge compared to text-only training. Surprisingly, they found little significant difference, suggesting that current multimodal models might not be fully leveraging visual information for commonsense reasoning.
What Problem Does This Paper Address?
Large language models (LLMs) trained solely on text have shown impressive capabilities, but they often lack true understanding and real-world grounding. This is partly because text alone can be ambiguous, biased, and doesn't capture the richness of sensory experience. Multimodal models, which are trained on both text and images, are seen as a promising solution to imbue LLMs with a deeper understanding of the world. However, a critical challenge is evaluating what visual commonsense knowledge these multimodal models actually acquire. Existing benchmarks often require images as input, making it difficult to isolate the knowledge gained purely from multimodal pre-training. This paper aims to fill this gap by developing text-only evaluation tasks to specifically measure visual commonsense knowledge.
Methods
The authors introduce two novel evaluation tasks designed to be used with standard masked language modeling (MLM) techniques, allowing for easy evaluation of various LLMs without task-specific fine-tuning.
- Memory Colors: This task, adapted from prior work, tests a model's knowledge of the typical colors of objects (e.g., "The usual color of a banana is [MASK]").
- Visual Property Norms: This is a new task created by the authors, leveraging the CSLB concept property norms dataset. It probes visual conceptual knowledge by asking models to fill in missing visual features of objects (e.g., "A banana has [MASK]").
Both tasks use English text templates with a [MASK] token, allowing models to predict the missing word. To account for potential prompt sensitivity, they evaluate models using multiple query templates for each task. The evaluation metric used is mean average precision (mAP).
Datasets
- Training Data: The paper mentions the datasets used to train the multimodal models, including MS COCO, SBU Captions, VG-QA, CC, VG, GQA, and VQA.
- Evaluation Data:
- Memory Colors: Uses a dataset of 109 typical objects with 13 query templates, created with input from human participants.
- Visual Property Norms: Based on the CSLB concept property norms dataset, which contains conceptual knowledge for 541 concrete objects from 30 human participants. The authors filtered this dataset to include only visual perceptual features and specific feature relations.
Key Insights and Ideas
- Multimodal Training Doesn't Guarantee Better Visual Commonsense (Yet): The most striking finding is that multimodal models (trained on text and images) did not significantly outperform unimodal models (trained only on text, sometimes text from image-caption pairs) on the visual commonsense tasks. This challenges the assumption that multimodal training automatically leads to superior visual commonsense.
- Text-Based Evaluation is Feasible: The paper successfully demonstrates that visual commonsense knowledge can be evaluated using purely text-based cloze tests, which is a significant contribution for benchmarking multimodal models.
- Prompt Sensitivity is Real: The results highlight that LLMs are sensitive to the phrasing of prompts, meaning the same underlying knowledge might be accessed differently depending on how the question is asked. This raises questions about what exactly is being measured in cloze-style evaluations.
- Fine-tuning on Visual Text Can Help: Unimodal models fine-tuned on text specifically curated from visual tasks (like image captions or VQA) showed improved performance, suggesting that such text might be a good proxy for visual information.
Key Results
- On the Memory Colors task, the multimodal CLIP-BERT-explicit model performed best, but a unimodal BERT model trained on LXMERT data was a close second, indicating a marginal multimodal advantage.
- On the Visual Property Norms task, the results were even less conclusive regarding a multimodal advantage. The multimodal CLIP-BERT-implicit model performed well, but unimodal models trained on visual text data were competitive.
- Performance generally increased with higher production frequency thresholds in Visual Property Norms, suggesting models are better at identifying more common or apparent visual features.
- The authors found that models' performance varied significantly across different query templates, underscoring the importance of robust evaluation strategies.
Limitations and Future Work
- Limited Model Scope: The study evaluated a subset of vision-and-language models, so the findings might not generalize to all types of multimodal models.
- Adaptation to Unimodal Input: The paper acknowledges that they adapted multimodal models to a unimodal (text-only) input by removing visual components. The effectiveness of this adaptation and how well these models would perform if truly designed for unimodal input is not fully explored.
- Prompt Sensitivity Concerns: The authors themselves raise concerns about prompt sensitivity, questioning the extent to which cloze tasks truly measure "knowledge" versus the model's ability to pattern-match based on specific phrasing.
- Future Work: The authors suggest further investigation into prompt engineering and exploring alternative evaluation methods beyond cloze tests to better understand what multimodal models learn. They also propose investigating how multimodal models adapt to unimodal inputs.
Summary updated