Scalable Multi-Domain Dialogue State Tracking

1. INTRODUCTION

• Ontology-based dialogue-state methods [9] [10] [11] represent each slot as a distribution over ontology-predefined values; large or unbounded value sets create scalability issues, and unseen training entities hinder dynamic databases.Source: §1 · p. 1
• The tracker generates candidate slot values from local dialogue context and possibly external knowledge, then scores only those candidates.A multi-domain language-understanding module tags slots in user utterances and delexicalizes them for downstream feature extraction.Source: §1 · p. 1
• GRU-based utterance representations provide tag-aligned slot-value features, alongside history-presence features independent of slot and value.The paper investigates bidirectional GRUs as an alternative to earlier deep feed-forward or convolutional representations.Source: §1 · p. 2
• Sharing parameters across slots and transferring them to unseen datasets or domains reduces per-slot training and facilitates adding new slots and domains.The paper attributes this adaptability to features independent of slot type or value.Source: §1 · p. 2

2. THE DIALOGUE STATE

• Candidate sets restrict tracking to values of interest, addressing the scalability limits of representing all possible values, especially for large, dynamic, or difficult-to-enumerate slots.The paper retains the assumption that the dialogue state factors into a distribution for each slot.Source: §2 · p. 2, §2.2 · p. 2

2.1. Candidate Set

• Each turn adds slot values detected in user and system utterances and caps the candidate set at K, ensuring newly mentioned values are included.If capacity is exceeded, less-preferred values can be removed; the initialization may also be extended with values from external sources.Source: §2.1 · p. 2
• The procedure sorts prior-turn candidates in decreasing score order before extending the set; experiments found candidate sets rarely reached maximum capacity when K = 7.Source: §2.1 · p. 2

2.2. State Representation

• For each slot, the state distribution covers the candidate set plus dontcare and null values, with padding to a fixed size of K + 2.The paper treats never-mentioned values as having probability close to zero, so they need not be explicitly stored; the resulting distribution is compact relative to the full value set.Source: §2.2 · pp. 2–3
• Figure 2 depicts shared neural-network parameters for time and restaurant candidate scorers, with candidate values derived from system and user utterances.When a candidate set has fewer than K entries, the remaining entries are padded with PAD.Source: Figure 2 · p. 3

3.1. Model Description

• For each slot, the discriminative tracker updates candidate scores and identifies the special dontcare and null states.Its inputs include the candidate set, latest user and system utterances, their dialogue acts, and the previous candidate set.Source: §3.1 · p. 3

3.2. Feature Extraction

• A two-layer bidirectional GRU processes delexicalized utterances, using final states for utterance representation and token-position states for candidate-related context.Delexicalization substitutes slot values recognized by the language-understanding model; recurrent features aim to capture long-term dependencies.Source: §3.2 · pp. 3–4, Figure 3 · p. 4

3.2.1. Utterance related features

• Utterance-related features combine user and preceding system utterance representations with their dialogue-act features, and are shared across candidate scorers.The user representation concatenates the GRU’s final forward and backward states; dialogue-act vectors mark acts without slot or value arguments.Source: §3.2.1 · p. 4

3.2.2. Slot related features

• Slot-related features are shared across that slot’s candidate scorers and include previous-turn scores for dontcare and null plus slot-specific dialogue acts.The acts cover both user and system turns.Source: §3.2.2 · p. 4

3.2.3. Candidate related features

• Candidate-related features include system dialogue acts associated with a candidate, its previous-turn score, and hidden-state context at its utterance positions.When the candidate is absent from the user utterance, its user-side hidden-state feature is a zero vector.Source: §3.2.3 · p. 5

3.3. Parameter Sharing and Transfer Learning

• The model shares parameters across slots within a domain and transfers parameters to a new domain without in-domain data; it also tests joint training across domains.The parameter dimensions do not depend on the slot or domain, respectively, supporting sharing or transfer.Source: §3.3 · p. 5

4. EVALUATION

• Joint goal accuracy counts a turn as correct only when every predicted slot value exactly matches the ground truth.Slot values are selected when their probabilities exceed a threshold tuned on the development set.Source: §4 · p. 5

4.1. Datasets

• Simulated restaurant and movie test sets include slot values unseen during training, with higher unseen-value rates than DSTC2.The datasets were generated with an agenda-based simulator and rule-based policy, then paraphrased by crowd-workers.Source: §4.1 · p. 5

4.2. Experiments

• GRU models with slot-level weight sharing perform better on DSTC2 and simulated restaurant conversations; results are mixed on the smaller movie dataset.The experiments compare CNN and GRU feature extraction with and without parameter sharing across slots.Source: Table 2 · p. 6, §4.2 · p. 6
• Combining datasets beats the best single-dataset model on both movie and restaurant datasets, while adding in-domain data further improves Sim-M performance.The transfer experiments use transcript utterances from DSTC2 training because the other datasets lack ASR noise.Source: Table 3 · p. 6, §4.2 · p. 6

4.3. Discussion

• On DSTC2, more elaborate NBT-CNN models can outperform the proposed approach, whose advantages are unbounded-value handling and transfer to new slots and domains.The proposed architecture is described as comparable to the NBT-CNN baseline.Source: Table 4 · p. 6, §4.3 · p. 6

• The paper identifies unseen values, large or unbounded value sets, and costly annotated datasets for new domains as three DST challenges.Source: §5 · p. 6
• The approach derives a bounded candidate set from dialogue history and possibly external knowledge, then tracks values within that set.Source: §5 · p. 6
• Experiments show benefits from sharing parameters across slots and transferring them to new domains for bootstrapping high-performance models.Source: §5 · p. 6