Sequential Latent Knowledge Selection for Knowledge-Grounded Dialogue

• Because dialogue context can correspond to multiple knowledge choices, tracking prior selections can narrow the current turn’s candidate set, while response information can improve selection.Source: §1 · p. 1
• SKT tracks prior and posterior knowledge distributions, updating them sequentially with previous-turn responses to predict knowledge from the posterior.Source: §1 · pp. 1–2
• The latent model can infer which knowledge other speakers are likely to select when previous-turn knowledge-selection labels are unavailable.Source: §1 · p. 2
• The authors propose SKT, the first attempt to use a sequential latent-variable model for knowledge selection in knowledge-grounded chit-chat.Source: §1 · p. 2
• SKT improves knowledge-selection accuracy and utterance generation, achieving new state-of-the-art performance on Wizard of Wikipedia (Dinan et al., 2019) and knowledge-annotated Holl-E (Moghe et al., 2018).Source: §1 · p. 2

Testbeds

• WoW (Dinan et al., 2019) tests knowledge selection and response generation with ground-truth selection labels and explicit knowledge-response grounding; Holl-E (Moghe et al., 2018) is also evaluated with clearer sentence labels.WoW is described as a large-scale, challenging open-domain, multi-turn benchmark.Source: §2 · p. 2

The Flow of Conversation

• In WoW, a Wizard and an Apprentice discuss one of 1,431 topics; the Wizard retrieves candidate knowledge, selects one sentence, generates a response, and continues for at least five turns each.The pool averages 67.57 sentences, including retrieved article paragraphs and sentences from the topic’s Wikipedia page.Source: §2 · pp. 2–3

Problem Statement

• The task separates knowledge selection from response generation: choose one relevant candidate sentence, then generate an utterance grounded in that knowledge and dialogue context.The passage situates this pipeline in open-domain TextQA and cites Chen et al. (2017) and Min et al. (2018) on its origins and effectiveness for single-document TextQA.Source: §2 · p. 3

Motivation

• WoW’s one-to-many context-to-knowledge relations make dialogue knowledge selection diverse, motivating a sequential latent model that represents this diversity and tracks topic flow across turns.Unlike TextQA, dialogue can continue with diverse knowledge unless the context asks a direct question.Source: §2 · p. 3
• Knowing a response boosts knowledge-sentence selection accuracy for humans and different models on WoW, suggesting joint modeling of selection and generation over multi-turn dialogue.The passage attributes this observation to Table 1 and notes that responses are written based on selected knowledge.Source: §2 · p. 3

3 APPROACH

• SKT treats knowledge selection as a sequential latent-variable decision, conditioning on previously selected knowledge to jointly infer multi-turn knowledge selection and response generation.The model uses dialogue context and selected knowledge from prior turns to select knowledge for the current response.Source: §3 · pp. 3–4
• The knowledge prior conditions on dialogue context and prior selected knowledge, while the approximate posterior additionally uses the current response to infer knowledge.The prior is a categorical conditional distribution; the inference network approximates the posterior over knowledge.Source: §3 · pp. 4–5
• The decoder generates responses from the current context and selected knowledge using a Transformer decoder with a copy mechanism.At test time, the model selects the highest-probability knowledge sentence from the prior distribution.Source: §3 · p. 5

3.1 TRAINING

• Training combines the variational lower bound with an auxiliary cross-entropy loss between predicted and true knowledge sentences.The model approximates the posterior expectation with one Gumbel-Softmax sample.Source: §3.1 · p. 5

4.1 DATASETS

• Wizard of Wikipedia has 18,430 training dialogues and separate Test Seen and Test Unseen subsets; Holl-E’s inconsistent span labels were converted into sentence-level ground-truth knowledge annotations.Holl-E annotations use the minimum consecutive sentences containing a span, or a “no passages used” label when no span is given; the latter covers 5% of its ground-truth labels.Source: §4.1 · pp. 5–6

4.3 QUANTITATIVE RESULTS

• SKT outperforms state-of-the-art models on Wizard of Wikipedia in knowledge-selection accuracy and response-generation unigram and bigram F1, with larger gaps on Test Unseen.PostKS with copy is the strongest baseline, while the results support sequential latent modeling as critical to improving knowledge selection and, in turn, generation.Source: Table 2 · p. 6, §4.3 · p. 7
• On Holl-E, SKT outperforms all baselines in every reported metric; BERT considerably reduces perplexity across models.The authors suggest BERT may prevent overfitting on Holl-E’s smaller dataset.Source: Table 3 · p. 6, §4.3 · p. 7

4.2 EXPERIMENTAL SETTING

• Evaluation measures response unigram and bigram F1, perplexity, and knowledge-selection accuracy, against Transformer-memory and PostKS baselines and their four ablation variants.For Holl-E’s multiple-reference evaluation, the best score across references is used, and a knowledge prediction is correct if it matches any correct answer.Source: §4.2 · pp. 6–7

4.4 QUALITATIVE RESULTS

• Annotators preferred SKT’s single-turn responses to baseline responses, with larger performance gaps on Test Unseen.The study sampled 100 test examples, rated engagingness and knowledgeability, and applied Bayesian calibration to scores.Source: §4.4 · pp. 7–8, Table 4 · p. 8
• In multi-turn human evaluation, annotators preferred SKT’s conversations to baselines, with a larger gap on Test Unseen.Selected dialogue examples also show SKT capturing topic changes and generating more appropriate responses.Source: Table 5 · p. 8, §4.4 · p. 8

5 RELATED WORK

• Most knowledge-grounded dialogue models focus on combining given knowledge with dialogue context to improve informativeness, rather than investigating knowledge selection.Examples include Ghazvininejad et al. (2018), Li et al. (2019b), and Dinan et al. (2019), whose approach selects knowledge before generating a response.Source: §5 · p. 8
• Unlike the closely related selection-focused model of Lian et al. (2019), SKT treats knowledge selection as a sequential latent-variable decision process and introduces a knowledge loss.The passage attributes significantly better experimental performance to these updates.Source: §5 · p. 8
• Prior sequential latent models include VRNN, SRNN, and Z-Forcing, with applications in summarization, captioning, and text generation; SKT applies this approach to knowledge-grounded chit-chat.Related latent attention models use latent variables to model attention mappings between input and output sequences (Deng et al., 2018; Yang et al., 2017).Source: §5 · p. 9

• SKT achieved new state-of-the-art performance on Wizard of Wikipedia (Dinan et al., 2019) and a knowledge-annotated version of Holl-E (Moghe et al., 2018).Source: §6 · p. 9
• The work proposed SKT, the first sequential latent variable model for knowledge selection in multi-turn knowledge-grounded dialogue.Source: §6 · p. 9
• Future directions include sequential Monte Carlo methods using filtering variational objectives (Maddison et al., 2017a) and interpretability studies such as measuring attention uncertainty (Heo et al., 2018).Source: §6 · p. 9

A DERIVATION OF CONDITIONAL PROBABILITY

• The section rewrites the conditional probability of wizard response y_t given dialogue context x_≤t and prior responses y_<t from Eq. (2) to Eq. (4).It presents this as a derivation in Section 3.Source: A DERIVATION OF CONDITIONAL PROBABILITY · p. 12
• qφ(k_i) denotes an approximated posterior distribution, whereas pθ(k_i) denotes a true posterior distribution.Source: A DERIVATION OF CONDITIONAL PROBABILITY · p. 12

B TRAINING DETAILS

• Non-BERT models use a 0.001 learning rate and fastText embeddings, while BERT models use a 0.00002 learning rate and BERT-Base, Uncased weights.Other parameters are initialized with Xavier; Adam uses β1 = 0.9, β2 = 0.999, and ϵ = 1e −07.Source: B TRAINING DETAILS · p. 12
• Label smoothing is 0.1 for knowledge selection and 0.05 for response generation, with Gumbel-Softmax temperature τ = 0.1 and knowledge-loss weight λ = 1.0.Source: B TRAINING DETAILS · p. 12
• Training batches dialogues rather than individual turns and runs for up to 5 epochs on two NVIDIA TITAN Xp GPUs.Source: B TRAINING DETAILS · p. 12

C KNOWLEDGE SELECTION ACCURACY OVER TURNS

• Knowledge selection accuracy significantly drops after the first turn for all models.The passage attributes this pattern to the diversity of knowledge selection; the first turn is often an easily predictable topic-definition sentence.Source: C KNOWLEDGE SELECTION ACCURACY OVER… · p. 12
• The model consistently outperforms other methods in knowledge selection accuracy at every Wizard of Wikipedia turn.Table 6 compares accuracy by turn.Source: C KNOWLEDGE SELECTION ACCURACY OVER… · p. 12, Table 6 · p. 13

D QUANTITATIVE RESULTS ON SEMI-SUPERVISED SETTING

• With only 1/4 of labeled training data, the model matches E2E Transformer MemNet and performs better on Test Unseen, while avoiding a severe performance drop.Performance improves as more labeled knowledge data is used for training.Source: D QUANTITATIVE RESULTS ON… · p. 13

E EXAMPLES WITH SELECTED KNOWLEDGE

• Figures 4 and 5 present selected examples of knowledge selection and response generation; for each dialogue context, our method’s and baselines’ selected knowledge and generated utterances are compared with human ground truths.Source: E EXAMPLES WITH SELECTED KNOWLEDGE · p. 13
• Figure 4 presents examples of selected knowledge and generated responses on the WoW Seen Test set.Source: Figure 4 · p. 14
• Figure 5 presents examples of selected knowledge and generated responses on the WoW Unseen Test set.Source: Figure 5 · p. 14