Abstract
The Automatic Dialogue Evaluation Model (ADEM) was introduced to address the limitations of word-overlap metrics in open-domain dialogue systems. However, this research provides a rigorous critique of ADEM, demonstrating its vulnerability to simple adversarial attacks and its lack of discriminative power. Through theoretical analysis of embedding conicity and empirical stress-testing, the authors reveal that ADEM's linear scoring function often fails to distinguish between high-quality ground-truth responses and nonsensical, syntactically broken ones.
1. Introduction
Evaluating open-domain chatbots is a significant challenge in NLP. Traditional metrics like BLEU and ROUGE rely on word overlap, which fails to account for the "one-to-many" nature of conversation—where many valid responses may share no words with a single reference. ADEM was proposed as a solution, using neural network embeddings to calculate semantic similarity. While initially promising, this paper investigates whether ADEM is a robust evaluator or if it merely provides "safe," average scores that are easily manipulated.
2. Methodology
The authors employ a three-pronged approach to evaluate ADEM:
- Adversarial Perturbations: Creating "broken" responses (e.g., jumbling word order, removing nouns) to test if ADEM's scores decrease appropriately.
- Theoretical Analysis: Using linear system theory to analyze the scoring function and the concept of conicity, which measures how vectors cluster in a specific direction.
- White-Box Attacks: Utilizing guided backpropagation and an Annoy index to find high-scoring but irrelevant responses from a massive Twitter database.
3. Theoretical Analysis: The Conicity Problem
A core finding of the paper is the "conicity problem." The authors prove that ADEM’s multiplicative embedding interactions force response vectors into a narrow cone. This high conicity makes different responses appear nearly identical to the model, severely limiting its ability to differentiate between distinct semantic meanings.
4. Experimental Results: Adversarial Vulnerabilities
Empirical tests show that ADEM is remarkably indifferent to grammar and syntax. For instance, reversing the word order of a response resulted in a score identical to the original ground truth. Furthermore, the model tends to assign a "safe" score of approximately 2.75 on a 5-point scale to almost any input, regardless of its quality.
As shown in the table above, even extreme corruptions like "Reverse the response" or "Jumble words in the sentence" maintain scores nearly identical to the reference response. In some cases, generic phrases like "fantastic! how are you?" actually outscored the ground truth in 87% of instances.
5. Discussion and Future Work
The authors conclude that ADEM’s primary weakness lies in its linear scoring formulation. Because the model relies on a linear combination of embeddings, it cannot capture the complex, non-linear nuances of human language.
Future Directions: * Non-linear Models: Moving toward non-linear scoring functions to handle dialogue diversity. * Adversarial Training: Training evaluation models specifically to resist syntactic attacks. * Syntactic Awareness: Incorporating structural and grammatical checks into evaluation metrics.
Summary updated