Deep learning evaluation using deep linguistic processing

Alexander Kuhnle, Ann CopestakeNAACL 2018

TL;DR

This paper argues that current evaluation methods for visual question answering (VQA) models are insufficient, often allowing models to achieve high accuracy through superficial pattern matching rather than true understanding. The authors propose using artificially generated, linguistically rich datasets to probe deeper into model capabilities and address the "Clever Hans effect."

What Problem Does This Paper Address?

Deep learning models, particularly in Natural Language Processing (NLP) and Computer Vision, have achieved remarkable performance. However, evaluating their true understanding, especially in multimodal tasks like Visual Question Answering (VQA), remains a significant challenge. Standard evaluation practices often rely on large, real-world datasets that can be exploited by models through superficial correlations or biases, rather than genuine reasoning. This phenomenon, termed the "Clever Hans effect," means models might appear intelligent but lack true comprehension. The paper highlights that existing VQA datasets, like the popular VQA Dataset, suffer from issues that allow systems to achieve competitive performance without truly learning the underlying abilities. This makes it difficult to assess if models are genuinely understanding language and its grounding in visual information.

Methods

The core of the paper's methodology lies in the automatic generation of artificial VQA data. They leverage existing deep linguistic processing technology, specifically the DELPH-IN framework and its bidirectional, high-precision English Resource Grammar (ERG) built on Minimal Recursion Semantics (MRS). This approach allows them to:

  • Generate Abstract Microworlds: Randomly sample abstract world models specifying entities, their attributes (shape, color, position, etc.), which can be visualized.
  • Syntactically Rich Language Generation: Use a compositional semantic approach with a bidirectional grammar to generate natural language sentences from abstract DMRS graphs. This enables the creation of complex and non-trivial captions with arbitrary syntactic complexity.
  • Controlled Data Generation: The framework allows for precise control over data generation, enabling the creation of datasets that specifically target certain linguistic phenomena, like quantification. This contrasts with the uncontrolled nature of real-world data.

Datasets

The paper discusses several datasets:

  • VQA Dataset (Antol et al., 2015): A popular real-world VQA dataset based on MS COCO images, which the authors argue is inadequate due to issues allowing superficial learning.
  • C-VQA (Agrawal et al., 2017): A composition-focused split of VQA v1.0.
  • VQA 2.0 (Goyal et al., 2017): An extension of VQA based on minimal image pairs.
  • SHAPES dataset (Andreas et al., 2016): An artificial dataset with abstract scenes and colored objects.
  • CLEVR dataset (Johnson et al., 2017a): An artificial dataset for compositional language and elementary visual reasoning.
  • NLVR dataset (Suhr et al., 2017): A dataset of images with abstract scenes and colored objects, using human-written captions.
  • ShapeWorld framework (Kuhnle and Copestake, 2017): The authors' implementation of their proposed approach for generating artificial VQA data.

The paper's primary contribution is the ShapeWorld framework for generating artificial VQA data, designed to overcome the limitations of existing datasets.

Key Insights and Ideas

  • The "Clever Hans Effect" is Pervasive: Models can achieve high accuracy on real-world datasets by exploiting superficial cues and biases, rather than demonstrating genuine understanding. This is analogous to the story of "Clever Hans," a horse that appeared to do math but was actually responding to subtle cues from its owner.
  • Artificial Data as a Diagnostic Tool: Instead of just augmenting training data, artificial datasets are crucial for targeted investigations of specific model capabilities. They act as a necessary, though not sufficient, benchmark for true language understanding.
  • Compositionality is Key: True language understanding requires compositionality – the ability to combine the meanings of smaller units to form the meaning of larger ones. Current DNNs, particularly recurrent ones, may struggle with this.
  • Controlled Generation for Rigorous Evaluation: By generating abstract microworlds and using deep linguistic processing, researchers can create datasets with controlled complexity and specific linguistic phenomena (like quantification), allowing for more precise evaluation of model reasoning.
  • Modular Evaluation is Superior: Evaluating DNNs on atomic test datasets that specifically target instance types (e.g., counting, spatial relations) provides a more modular and trustworthy assessment of understanding than a single accuracy number on a monolithic dataset.

Key Results

The paper doesn't present specific quantitative results from running models on their generated data in this excerpt. Instead, it focuses on:

  • Demonstrating the potential of artificial data: They show how their approach can create challenging abstract datasets.
  • Illustrating complex linguistic phenomena: Figure 2 provides examples of quantification-focused captions that are difficult to resolve without careful visual and linguistic analysis, highlighting the limitations of existing datasets.
  • Proposing a principled evaluation methodology: The paper lays out the theoretical and technical groundwork for generating data that can reveal deeper insights into model understanding.

Limitations and Future Work

  • Current Scope: The authors acknowledge that their current approach does not yet consider scope ambiguity of nested quantifiers, although the underlying DMRS formalism supports it and can be extended.
  • Future Extensions:
    • Implementing paraphrase rules on the grammar level to increase linguistic variety.
  • Extending the framework to other languages by translating the internal mapping of DMRS components.
  • Potentially extending to more complex domains beyond abstract shapes, such as clip-arts.

Summary updated

BibTeX
@inproceedings{kuhnle2018deep,
  author = {Alexander Kuhnle and Ann Copestake},
  title = {{Deep learning evaluation using deep linguistic processing}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
}
Figure