Polite Task-oriented Dialog Agents: To Generate or to Rewrite?

Diogo Silva, David Semedo, João MagalhãesACL 2022

TL;DR

This paper investigates how to make task-oriented dialog agents more polite, focusing on the fashion domain. They compare two main approaches: generating polite responses from scratch versus rewriting existing responses to be more polite. The study finds that rewriting is generally more effective, especially when adapted to the specific domain.

What Problem Does This Paper Address?

In conversational AI, especially for task-oriented agents (like those helping you shop or book a reservation), the tone of the agent's voice is crucial for user engagement and natural interaction. Simply providing correct information isn't enough; the agent needs to sound polite and align with the domain's specific language and culture. This is particularly challenging in domains like fashion, which have a rich and specialized vocabulary. Existing politeness strategies often work well in general contexts but struggle when transferred to these specialized domains due to vocabulary mismatches and domain-specific nuances. The paper aims to bridge this gap by exploring how to effectively adapt politeness strategies to task-specific domains, specifically fashion, and comparing two primary methods for achieving this: polite answer generation and politeness rewriting.

Methods

The authors explore two main strategies for imbuing politeness into task-oriented dialog agents:

  • Politeness through Utterance Generation (Gen): This approach integrates politeness directly into the answer generation process. They adapt the "Polite-RL" model, which uses a Seq2Seq architecture trained with Reinforcement Learning. A politeness classifier guides the generation towards more polite outputs. A key innovation here is the use of "Domain-Extended embeddings" (DE), which are trained on a combination of general politeness data and fashion domain data to better handle the specialized vocabulary.
  • Politeness through Utterance Rewriting (RW): This method separates politeness from answer generation, treating it as a style transfer task. They adapt the "Tag-and-Generate" model. A "Tagger" identifies style markers in an utterance, and a "Generator" then rewrites the utterance to incorporate polite style markers. To adapt this to the fashion domain, they propose four variants:
  • RW-Enron: Trained on the original Enron email dataset.
  • RW-Fashion: Trained on polite and rude fashion utterances.
  • RW-Fashion-Clean: Similar to RW-Fashion but excludes style markers associated with product nouns.
  • RW-Mixed: Learns style markers from the Enron domain and is then trained on the fashion dataset, aiming to transfer knowledge.

Datasets

The paper utilizes three datasets:

  • Stanford Politeness Corpus (SPC): Used for training the politeness classifier.
  • Enron Dataset: A collection of emails, used as the "original domain" for transfer learning experiments.
  • Multimodal Dialogs Dataset (MMD): A multi-turn dialog dataset for the fashion domain, serving as the "target domain." They create two subsets:
  • MMD-R: Focuses on system utterances recommending products.
  • MMD-A: Includes all neutral and polite system utterances with more than 5 tokens, generalizing beyond just recommendations.

Key Insights and Ideas

  • Rewriting Outperforms Generation for Politeness: The study consistently shows that rewriting approaches (RW) generally achieve better results than generative approaches (Gen) in terms of both politeness and content preservation, especially in a domain-transfer scenario. This is attributed to rewriting's ability to focus solely on politeness without the dual burden of understanding context and generating content.
  • Domain Adaptation is Crucial: Simply applying general politeness models to a new domain like fashion is insufficient. Adapting models, particularly through domain-specific embeddings or by training on domain data, is essential for effective politeness transfer.
  • Style Marker Curation Matters: For rewriting models, carefully curating the style markers is important. Excluding domain-specific product nouns (as in RW-Fashion-Clean) can lead to better content preservation.
  • Out-of-Domain Knowledge Can Help: The RW-Mixed model, which learns style markers from the Enron domain and applies them to fashion, shows promising results, suggesting that leveraging general linguistic patterns can be beneficial when domain-specific markers are scarce or noisy.
  • Politeness vs. Content Trade-off: While models aim to improve politeness, there's an inherent trade-off with preserving the original content. Rewriting models generally strike a better balance.
  • Neutral vs. Rude Utterances: Models tend to perform better on neutral utterances, but this is partly because they are less adept at identifying what needs to be changed to truly improve politeness. They often make fewer, less impactful changes on neutral text.

Key Results

  • Automatic Evaluation: Rewriting variations (RW) consistently outperformed generation-based ones (Gen) across politeness and content preservation metrics on the MMD dataset. For example, RW-Fashion-C and RW-Mixed achieved high politeness scores while maintaining good content preservation.
  • Human Evaluation: Human evaluation corroborated the automatic findings. The Gen+DE model scored lower on politeness than the reference, while most RW setups matched or improved it. Grammar scores were generally lower for all models compared to the reference.
  • Impact of Utterance Tone and Length:
    • Content preservation scores were generally lower for rude utterances compared to neutral ones.
  • Models showed a tendency to leave longer utterances untouched, making more changes on shorter ones.
  • RW-Fashion-C achieved the highest content preservation on short and medium utterance scenarios.
  • Qualitative Analysis: The RW-Mixed model sometimes produced slightly improved, more fashion-specific forms, while RW-Enron and RW-Fashion-C could produce good results if semantics were applied correctly.

Limitations and Future Work

  • Limitations:
    • The paper notes that generative models, even with domain-extended embeddings, can struggle with content preservation when heavily optimizing for politeness, sometimes leading to repetitive or nonsensical outputs.
  • The study acknowledges that automatic metrics for politeness have limitations and that human evaluation is crucial.
  • The performance on neutral utterances is noted as being less about true improvement and more about making minimal changes.
  • Future Work:
    • Extend the work to research methods that select the best politeness strategies while accounting for the specificity of different conversation phases (e.g., greetings vs. product descriptions).
  • Further investigate how to overcome domain shift issues, particularly for neutral text.
  • Explore more sophisticated ways to integrate domain knowledge and style markers.

Summary updated

BibTeX
@inproceedings{silva2022polite,
  author = {Diogo Silva and David Semedo and João Magalhães},
  title = {{Polite Task-oriented Dialog Agents: To Generate or to Rewrite?}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
}
Figure