Knowledge-Grounded Dialogue Generation with Term-level De-noising

Wen Zheng, Natasa Milic-Frayling, Ke ZhouACL 2021

TL;DR

This paper introduces a novel Knowledge Term Weighting Model (KTWM) that addresses the problem of "noise" in knowledge-grounded dialogue generation. By learning to assign specific weights to individual knowledge terms, KTWM effectively filters out irrelevant information, leading to more coherent and accurate responses. This approach significantly improves dialogue generation quality, especially on datasets with unseen topics.

What Problem Does This Paper Address?

Modern dialogue systems often leverage external knowledge to generate more informative and engaging responses. However, simply injecting entire sentences or paragraphs of knowledge can introduce irrelevant or even contradictory information, a phenomenon known as "noise." This noise can degrade the quality of the generated dialogue. Existing methods primarily focus on selecting knowledge at the sentence or paragraph level, making it difficult to control for the quality of individual terms within that knowledge. This paper tackles the challenge of precisely controlling the contribution of each knowledge term to the final response, aiming to maximize the impact of relevant information while minimizing the detrimental effects of noise.

Methods

The core contribution of this paper is the Knowledge Term Weighting Model (KTWM). KTWM operates in two main phases:

  • Training Phase:
    • Simulated Response Vectors (SRVs): The model learns to generate SRVs from post embeddings. These SRVs act as a proxy for response embeddings when actual response information isn't available (e.g., during testing).
  • Knowledge Term Weighting: KTWM uses attention mechanisms to compute two key attention distributions:
  • Ark: Between response embeddings and knowledge embeddings.
  • Ask: Between SRVs and knowledge embeddings.
  • Loss Functions: Three loss functions are employed:
  • Lfilter: A Binary Cross-Entropy loss that encourages relevant knowledge terms to be promoted, using ground truth response information.
  • Lapprox: A Mean Squared Error loss that aims to make the SRV attention (Ask) similar to the response attention (Ark), helping SRVs approximate response behavior.
  • LNLL: The standard Negative Log Likelihood loss for training the decoder.
  • The weighted knowledge terms (derived from Ask in the test phase) are then fed into a Transformer decoder.
  • Test Phase:
    • KTWM uses the learned SRVs to compute attention scores with knowledge terms. These scores are then used to weight the knowledge terms before they are passed to the decoder. This allows for term-level de-noising without direct access to the ground truth response.

The architecture utilizes standard Transformer encoders and decoders, with the novelty lying in the SRV generation and the term-level weighting mechanism.

Datasets

The authors evaluated their model on two publicly available, human-curated dialogue datasets:

  • Wizard of Wikipedia (Wiz): This dataset features dialogues between a "wizard" (teacher) and an "apprentice" (student). The wizard uses Wikipedia information to answer the apprentice's questions. It includes a "seen test set" (topics present in training) and an "unseen test set" (topics not seen in training), crucial for evaluating robustness.
  • Holl-E: This dataset focuses on movie-related conversations, where background knowledge from movie plots, reviews, and facts is available.

No new datasets were introduced in this paper.

Key Insights and Ideas

  • Term-level de-noising is crucial: Simply injecting knowledge at the sentence level is insufficient. The paper demonstrates that focusing on individual term importance significantly boosts performance.
  • SRVs as a clever proxy: Using Simulated Response Vectors (SRVs) during training to mimic response behavior allows the model to learn effective knowledge term weighting even when the actual response is unknown during inference.
  • KTWM is robust to unseen topics: The significant improvements on the Wiz "unseen test set" highlight KTWM's ability to generalize and handle novel information effectively.
  • Noise detection is effective: The model shows high precision in identifying noisy terms, contributing to better dialogue quality.

Key Results

  • KTWM consistently outperformed baseline methods (Post-KS, SKT, TED) across various metrics (BLEU, METEOR, BOW Embedding) on both Wiz and Holl-E datasets, with statistically significant improvements.
  • The largest relative improvement was observed when KTWM was combined with Post-KS knowledge, showing a METEOR score increase of up to 54.5%.
  • The model demonstrated a remarkable performance on the Wiz unseen test set, indicating strong generalization capabilities.
  • An ablation study showed that removing the Lfilter loss function led to a decrease in BLEU-1 and Average scores, while removing Lapprox had a more substantial negative impact, underscoring the importance of both loss components.

Limitations and Future Work

  • Limitations:
    • The paper notes that while KTWM effectively detects noisy terms, it only achieves a 50% precision in predicting useful terms.
  • The effectiveness of KTWM can be diminished if the injected knowledge itself lacks useful terms.
  • The model sometimes assigns high weights to punctuation and common structural terms, which might be an artifact of their widespread distribution.
  • Future Work:
    • Extend KTWM to incorporate multiple sources of evidence, balancing selected knowledge with dialogue context (previous turns).
  • Investigate inter-dependencies among terms when weighting selected knowledge.

Summary updated

BibTeX
@inproceedings{zheng2021knowledge,
  author = {Wen Zheng and Natasa Milic-Frayling and Ke Zhou},
  title = {{Knowledge-Grounded Dialogue Generation with Term-level De-noising}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure