PaperlayerView exact revision on arXiv

Source-linked AI summary

Sequential Latent Knowledge Selection for Knowledge-Grounded Dialogue

Byeongchang Kim, Jaewoo Ahn, Gunhee Kim

arXiv:2002.07510v2cs.CL

TL;DR

Knowledge-grounded dialogue requires selecting external knowledge across multi-turn contexts, where diverse choices make selection difficult. The paper proposes SKT, a sequential latent variable model that tracks prior and posterior knowledge distributions and uses response information. SKT improves knowledge selection and utterance generation, reaching state-of-the-art performance on Wizard of Wikipedia and Holl-E.

  • Problem

    Knowledge-grounded dialogue requires selecting appropriate external knowledge in multi-turn conversations, where one dialogue context can correspond to diverse knowledge choices.

  • Method

    SKT uses a sequential latent variable model that tracks prior and posterior knowledge distributions and incorporates response information for knowledge selection.

  • Results

    The model improves knowledge selection accuracy and utterance generation, achieving new state-of-the-art performance on Wizard of Wikipedia and Holl-E.

  • Takeaways & Limitations

    Sequential latent modeling provides a way to represent diverse knowledge choices and track knowledge topic flow in multi-turn dialogue.

Abstract

from arXiv · show

Knowledge-grounded dialogue is a task of generating an informative response based on both discourse context and external knowledge. As we focus on better modeling the knowledge selection in the multi-turn knowledge-grounded dialogue, we propose a sequential latent variable model as the first approach to this matter. The model named sequential knowledge transformer (SKT) can keep track of the prior and posterior distribution over knowledge; as a result, it can not only reduce the ambiguity caused from the diversity in knowledge selection of conversation but also better leverage the response information for proper choice of knowledge. Our experimental results show that the proposed model improves the knowledge selection accuracy and subsequently the performance of utterance generation. We achieve the new state-of-the-art performance on Wizard of Wikipedia (Dinan et al., 2019) as one of the most large-scale and challenging benchmarks. We further validate the effectiveness of our model over existing conversation methods in another knowledge-based dialogue Holl-E dataset (Moghe et al., 2018).

1 INTRODUCTION

The paper targets knowledge selection in multi-turn knowledge-grounded dialogue with a sequential latent variable model, SKT. It reports improvements in knowledge selection and utterance generation, achieving state-of-the-art performance on Wizard of Wikipedia and Holl-E.

  • Knowledge-grounded dialogue combines discourse context with selected external knowledge to generate informative responses.
  • The paper focuses on knowledge selection because dialogue contexts can correspond to multiple valid knowledge choices.
  • The model tracks prior and posterior knowledge distributions while using response information to predict knowledge.
  • SKT is presented as the first sequential latent variable model for knowledge selection in knowledge-grounded dialogue.
  • The proposed model improves knowledge selection accuracy and utterance generation, achieving new state-of-the-art performance on Wizard of Wikipedia and Holl-E.

2 PROBLEM STATEMENT AND MOTIVATION

The task requires selecting one relevant knowledge sentence and generating a response in a multi-turn conversation. Because dialogue contexts can support diverse knowledge choices, the paper motivates sequential latent modeling that tracks knowledge topic flow and uses response information.

  • Wizard of Wikipedia evaluates knowledge selection and response generation in large-scale open-domain multi-turn knowledge-based dialogue.
  • The Flow of Conversation: In the Wizard setting, a wizard uses retrieved Wikipedia knowledge to select one relevant sentence and construct an utterance.
  • The Flow of Conversation: The conversation involves shared topics, retrieved candidate knowledge, wizard responses, and repeated turns between the wizard and apprentice.
  • The Motivation of Sequential Latent Models: The task pipeline first selects a single relevant knowledge sentence, then generates an utterance from that knowledge and dialogue context.
  • The Motivation of Sequential Latent Models: Dialogue permits one-to-many context-to-knowledge relations, making knowledge selection diverse and motivating latent variables that track sequential topic flow.
  • The Motivation of Sequential Latent Models: Knowing the response improves knowledge sentence selection accuracy, supporting joint modeling of knowledge selection and response generation.

3 APPROACH

SKT models multi-turn knowledge selection as a sequential latent process, jointly inferring knowledge and generating responses from dialogue context and prior selections. It uses sequential distributions over knowledge, attention-based selection, and a Transformer decoder with copying.

  • Sentence Encoding: The model represents dialogue utterances and knowledge sentences, then conditions knowledge distributions on previous turns and previously selected knowledge.Utterance pairs are encoded with GRU states, while prior and posterior distributions are updated sequentially.
  • Sequential Knowledge Selection: SKT treats knowledge selection as a sequential decision process and models selections as latent variables for joint multi-turn inference with response generation.This replaces separate turn-by-turn inference with a sequential latent formulation.
  • Sequential Knowledge Selection: SKT uses a prior distribution from dialogue context and a posterior inference network that also uses the current response to approximate posterior knowledge selection.The decoder generates responses from selected knowledge, while the posterior supplies response-informed knowledge inference.
  • Knowledge Selection: Knowledge distributions are computed with attention over the current knowledge pool, sequentially updated with sampled knowledge and utterance embeddings.A GRU conditions the prior and posterior on previously selected knowledge; test-time selection uses the highest-probability knowledge.
  • Decoding with Copy Mechanism: The selected knowledge and current context are passed to a Transformer decoder using a copy mechanism to generate the wizard’s response word by word.At test time, the model selects the highest-probability knowledge before decoding the response.
  • Training: Training combines a variational lower-bound objective with an auxiliary cross-entropy knowledge loss, using posterior sampling with Gumbel-Softmax.The auxiliary loss uses true knowledge labels when available, while the variational objective trains the latent selection model.

4 EXPERIMENTS

The experiments evaluate SKT on Wizard of Wikipedia and Holl-E using automatic metrics and human judgments. SKT outperforms the baselines across reported knowledge-selection, response-generation, and human-evaluation comparisons.

  • Datasets and baselines: The evaluation covers Wizard of Wikipedia and Holl-E, with comparisons against state-of-the-art knowledge-grounded dialogue models.Wizard of Wikipedia is the main testbed, while Holl-E provides an additional knowledge-grounded chit-chat dataset.
  • Datasets and baselines: Wizard of Wikipedia contains separate Test Seen and Test Unseen subsets, enabling evaluation on overlapping and unseen topics.The dataset includes 18,430 training, 1,948 validation, and 1,933 test dialogues.
  • Evaluation protocol: The experiments measure knowledge-selection accuracy, unigram and bigram F1, and perplexity for response generation.Higher R-1 and R-2 and lower perplexity indicate better performance.
  • Quantitative results: SKT outperforms the compared models on all reported Wizard of Wikipedia knowledge-selection and utterance-generation metrics, with larger gaps on Test Unseen.The results also indicate that sequential latent modeling improves over models using prior and posterior knowledge distributions without the full sequential model.
  • Quantitative results: SKT likewise outperforms all baselines on every reported Holl-E metric, while BERT substantially reduces perplexity across models.The authors associate the perplexity reduction with Holl-E’s smaller dataset size and reduced overfitting.
  • Human evaluation: Human evaluators prefer SKT’s responses in both single-turn and multi-turn Wizard of Wikipedia evaluations, especially on Test Unseen.The single-turn study rates engagingness and knowledgeability, while the multi-turn study rates overall conversation quality.

5 RELATED WORK

Prior work studied knowledge-grounded dialogue through new datasets and models, but generally investigated knowledge selection less than knowledge-context integration. SKT differs from the closest related model by treating selection as a sequential latent decision process.

  • Knowledge-grounded dialogue: Knowledge-grounded dialogue research includes both collecting new datasets and developing models for combining knowledge with dialogue context.The related work covers multiple knowledge-based dialogue datasets and model families.
  • Knowledge selection: Most prior models focus more on combining given knowledge and dialogue context than on the knowledge-selection problem itself.This motivates the paper’s emphasis on selecting pertinent knowledge in multi-turn dialogue.
  • Closest related model: Lian et al. also study knowledge selection, but SKT models it as a sequential decision process with latent variables and adds a knowledge loss.The paper reports significantly better performance after these changes.
  • Sequential latent models: Sequential latent variable models have been applied to sequential data, document summarization, image captioning, and text generation.The paper places SKT within this broader line of sequential latent-model research.

6 CONCLUSION

The paper introduces SKT, a sequential latent variable model for knowledge selection in multi-turn knowledge-grounded dialogue. It reports state-of-the-art performance on Wizard of Wikipedia and a knowledge-annotated Holl-E dataset, while identifying further work on inference and interpretability.

  • Contribution: The paper proposes SKT as the first sequential latent variable model for knowledge selection in multi-turn knowledge-grounded dialogue.The conclusion identifies sequential latent modeling as the paper’s central contribution.
  • Results: SKT achieves new state-of-the-art performance on Wizard of Wikipedia and a knowledge-annotated version of Holl-E.The conclusion presents these results across both evaluated datasets.
  • Future directions: Future work includes sequential Monte Carlo inference with filtering variational objectives and methods for interpreting knowledge-selection uncertainty.The proposed directions target inference procedures and uncertainty in attention.

A DERIVATION OF CONDITIONAL PROBABILITY

The paper rewrites the conditional probability of the wizard’s response given dialogue context and previous responses. The derivation distinguishes approximated and true posterior distributions over knowledge.

  • The conditional probability of response y_t is re-written from Eq. (2) to Eq. (4) using the dialogue context and previous responses.The derivation is presented as a direct reformulation of the response probability.
  • The derivation therefore relates response generation to posterior-based knowledge modeling in the sequential dialogue setting.
  • q_φ(k_i) denotes an approximated posterior distribution, while p_θ(k_i) denotes the true posterior distribution.

B TRAINING DETAILS

Training uses Xavier initialization and Adam optimization, with different learning rates and pretrained components for models with and without BERT.

  • All parameters except pretrained components are initialized with the Xavier method.
  • The models are trained with Adam using β1 = 0.9, β2 = 0.999, and ϵ = 1e−07.
  • Models without BERT use learning rate 0.001 and fastText embeddings trained on the Common Crawl corpus.
  • Models with BERT use learning rate 0.00002 and BERT-Base, Uncased pretrained encoder weights.

C KNOWLEDGE SELECTION ACCURACY OVER TURNS

On Wizard of Wikipedia, the sequential latent-variable model consistently achieves higher knowledge-selection accuracy across dialogue turns. Accuracy drops after the first turn for all models, reflecting diversity in later knowledge selection.

  • The sequential latent-variable model consistently outperforms other methods in knowledge-selection accuracy across all dialogue turns.
  • Knowledge-selection accuracy significantly drops after the first turn across all models.The first turn is often predictable because it commonly contains a topic-definition sentence.
  • The post-first-turn decline illustrates the diverse nature of knowledge selection in multi-turn dialogue.

D QUANTITATIVE RESULTS ON SEMI-SUPERVISED SETTING

With partial knowledge labels on Wizard of Wikipedia, the model performs better with more labeled training data while remaining competitive with fewer labels.

  • Performance improves as the amount of labeled knowledge data used for training increases.
  • Using only 1/4 labeled training data produces performance comparable to E2E Transformer MemNet.
  • With 1/4 labeled training data, the model performs better than E2E Transformer MemNet on Test Unseen.

E EXAMPLES WITH SELECTED KNOWLEDGE

The section presents examples comparing the model and baselines on selected knowledge and generated responses, alongside human ground truths, for seen and unseen Wizard of Wikipedia test settings.

  • Seen Test: Figure 4 compares selected knowledge and generated utterances from the model and baselines with human ground truths on the WoW Seen Test set.The example topic is Italian Cuisine.
  • Unseen Test: Figure 5 compares selected knowledge and generated utterances from the model and baselines with human ground truths on the WoW Unseen Test set.The example topic is Hunting.
  • Seen Test: The Seen Test example uses an Italian Cuisine dialogue context and includes a response about chicken parmigiana and beef lasagna.
  • Unseen Test: The Unseen Test example discusses hunting, including government payments, excessive hunting, endangered species, and poaching.
Loading 2002.07510v2…