[Paper Review] Probing Causal Common Sense in Dialogue Response Generation.

Pei Zhou, Pegah Jandaghi+4 more|arXiv (Cornell University)|Apr 19, 2021

Topic Modeling4 citations

TL;DR

This paper introduces CEDAR, a task that formalizes conversational common sense as causal explanations for responses, and proposes a probing framework to evaluate whether response generation models can reason logically about these explanations. Results show models struggle with logical validity but easily detect grammatical naturalness.

ABSTRACT

Communication is a cooperative effort that requires reaching mutual understanding among the participants. Humans use commonsense reasoning implicitly to produce natural and logically-coherent responses. As a step towards fluid human-AI communication, we study if response generation (RG) models can emulate human reasoning process and use common sense to help produce better-quality responses. We aim to tackle two research questions: how to formalize conversational common sense and how to examine RG models capability to use common sense? We first propose a task, CEDAR: Causal common sEnse in DiAlogue Response generation, that concretizes common sense as textual explanations for what might lead to the response and evaluates RG models behavior by comparing the modeling loss given a valid explanation with an invalid one. Then we introduce a process that automatically generates such explanations and ask humans to verify them. Finally, we design two probing settings for RG models targeting two reasoning capabilities using verified explanations. We find that RG models have a hard time determining the logical validity of explanations but can identify grammatical naturalness of the explanation easily.

Motivation & Objective

• To formalize conversational common sense as causal textual explanations for dialogue responses.
• To evaluate whether response generation models can reason logically about these explanations.
• To develop an automated method for generating and verifying explanations using human validation.
• To design probing settings that test two distinct reasoning capabilities in RG models: logical validity and grammatical naturalness.
• To assess the extent to which current RG models emulate human-like commonsense reasoning in dialogue.

Proposed method

• Propose CEDAR, a task that frames common sense in dialogue as causal explanations linking prior dialogue context to a response.
• Design an automatic pipeline to generate candidate explanations for given dialogue-response pairs.
• Use human annotation to verify the quality and validity of generated explanations.
• Construct two probing settings: one testing model sensitivity to logical validity of explanations, another testing sensitivity to grammatical naturalness.
• Compare modeling loss under valid vs. invalid explanations to measure model behavior.
• Use the difference in loss between valid and invalid explanations as a proxy for logical reasoning capability.

Experimental results

Research questions

• RQ1 How can conversational common sense be formally represented as causal explanations in dialogue response generation?
• RQ2 To what extent can response generation models detect the logical validity of causal explanations?
• RQ3 Can RG models distinguish between grammatically natural and unnatural explanations?
• RQ4 How do model losses differ when given valid versus invalid causal explanations in a probing setup?
• RQ5 What are the limitations of current RG models in emulating human-like commonsense reasoning?

Key findings

• RG models exhibit significant difficulty in determining the logical validity of causal explanations, indicating a lack of deep commonsense reasoning.
• Models can easily identify grammatical naturalness of explanations, suggesting sensitivity to surface-level fluency.
• The loss difference between valid and invalid explanations is small, implying weak discrimination on logical grounds.
• Human-verified explanations are essential for constructing reliable probing benchmarks in commonsense reasoning.
• Current RG models prioritize fluency and coherence over logical consistency when generating responses.
• The gap between human reasoning and model behavior highlights a key challenge in achieving human-like dialogue understanding.