HyST: A Hybrid Approach for Flexible and Accurate Dialogue State Tracking

TL;DR

HyST is a hybrid dialogue state tracking method that combines the strengths of probability distribution and candidate-generation approaches, improving accuracy and scalability in multi-domain dialogue systems.

Contribution

This paper introduces HyST, a novel hybrid approach that learns to select the best method for each slot type, enhancing performance on large, multi-domain datasets.

Findings

HyST outperforms previous state-of-the-art methods by 24%

HyST demonstrates scalability to multi-domain applications

The hybrid approach improves accuracy over individual methods

Abstract

Recent works on end-to-end trainable neural network based approaches have demonstrated state-of-the-art results on dialogue state tracking. The best performing approaches estimate a probability distribution over all possible slot values. However, these approaches do not scale for large value sets commonly present in real-life applications and are not ideal for tracking slot values that were not observed in the training set. To tackle these issues, candidate-generation-based approaches have been proposed. These approaches estimate a set of values that are possible at each turn based on the conversation history and/or language understanding outputs, and hence enable state tracking over unseen values and large value sets however, they fall short in terms of performance in comparison to the first group. In this work, we analyze the performance of these two alternative dialogue state tracking methods, and present a hybrid approach (HyST) which learns the appropriate method for each slot type. To demonstrate the effectiveness of HyST on a rich-set of slot types, we experiment with the recently released MultiWOZ-2.0 multi-domain, task-oriented dialogue-dataset. Our experiments show that HyST scales to multi-domain applications. Our best performing model results in a relative improvement of 24% and 10% over the previous SOTA and our best baseline respectively.

Tables

Table 1. Table 1: An example dialogue with dialogue states after each turn. Agent Turns are followed by their dialogue acts in the brackets.

User:	I need to book a hotel in the east that has 4 stars.
Hotel	area=east, stars=4
Agent:	I can help you with that. What is your price range?
	(Hotel-Request(Price))
User:	That doesn’t matter if it has free wifi and parking.
Hotel	parking=yes, internet=yes
	price=dontcare, stars=4, area=east
Agent:	If you’d like something cheap, I recommend Allenbell.
	(Hotel-Recommend(Price), Hotel-Recommend(Name))
User:	That sounds good, I would also like a taxi to the hotel
	from cambridge
Hotel	parking=yes, internet=yes
	price=dontcare, area=east, stars=4
Taxi	departure=Cambridge
	destination=Allenbell

Table 2. Table 2: Some dataset statistics. Numerical values refer to things like ‘time’ and ‘people’ which are open ended.

Data property	Train	Dev	Test
# Dialogues	8,483	1,000	1000
# User turns	56,781	7,374	7372
# Uservocab (with num. values)	4311	1875	1840
# Uservocab (without num. values)	3805	1709	1646
Median user sent length (tokens)	11	11	11

Table 3. Table 3: Slot breakdown for MultiWOZ-2.0. The vocab size is the number of unique values seen for a slot in the training set. %None refers to the percentage of turns where the slot was not set. Out-of-vocabulary (OOV) rate refers to percentage of slots in the development set which are not present in the training set for each method. Slots marked with * are the ones which had better performance using the open-vocabulary (OV) approach.

Slot Name	Vocab	%None	OV	JST
	Size		OOV rate	OOV rate
taxi.leaveAt*	119	96.19%	1.25%	0.19%
taxi.destination*	277	92.76%	0.80%	0.14%
taxi.departure*	261	92.87%	0.94%	0.34%
taxi.arriveBy*	101	96.84%	0.39%	0.08%
restaurant.people	9	84.10%	0.72%	0.00%
restaurant.day	10	84.11%	0.18%	0.00%
restaurant.time*	61	84.22%	0.27%	0.05%
restaurant.food	104	71.65%	0.60%	0.00%
restaurant.pricerange	11	74.62%	0.58%	0.00%
restaurant.name*	183	87.14%	1.48%	0.26%
restaurant.area	19	74.02%	0.47%	0.01%
bus.people	1	100%	0.00%	0.00%
bus.leaveAt	2	99.99%	0.00%	0.00%
bus.destination	5	99.94%	0.00%	0.00%
bus.day	2	99.99%	0.00%	0.00%
bus.arriveBy	1	100%	0.00%	0.00%
bus.departure	2	99.94%	0.00%	0.00%
hospital.department	52	99.30%	0.09%	0.00%
hotel.people	9	84.61%	0.83%	0.00%
hotel.day	11	84.59%	0.18%	0.00%
hotel.stay	9	84.64%	0.66%	0.00%
hotel.name	89	84.80%	1.68%	0.24%
hotel.area	24	80.82%	0.18%	0.00%
hotel.parking	8	85.58%	0.45%	0.00%
hotel.pricerange	9	82.74%	0.72%	0.00%
hotel.stars	13	83.59%	1.83%	0.01%
hotel.internet	8	85.88%	0.49%	0.00%
hotel.type	18	82.18%	1.49%	0.07%
attraction.type	37	81.45%	3.95%	0.01%
attraction.name*	137	89.71%	1.37%	0.38%
attraction.area	16	82.80%	0.33%	0.03%
train.people	13	89.11%	2.54%	0.00%
train.leaveAt*	134	86.68%	3.09%	0.72%
train.destination	29	71.89%	1.00%	0.05%
train.day	11	72.90%	0.20%	0.04%
train.arriveBy*	107	86.81%	1.82%	0.18%
train.departure	35	72.38%	0.94%	0.07%
All values			25.60%	2.48%

Table 4. Table 4: Comparison of various apporaches on different domains. The numbers presented are joint goal accuracy.

Domain	JST	OV ST	HyST
Taxi	91.48%	92.30%	92.30%
Restaurant	78.55%	75.05%	79.67%
Bus	100%	99.95%	100%
Hospital	100%	100%	100%
Hotel	79.38%	74.04%	79.14%
Attraction	83.42%	82.80%	85.63%
Train	82.13%	71.19%	84.03%

Table 5. Table 5: Joint goal Accuracy on MultiWOZ-2.0. We present the ensemble (we do three independent model runs and take the majority vote per slot) results for our methods (presented in boldface ) with the single model results in parentheses.

Method	Accuracy
Majority Baseline	1.5%
MultiWOZ-2.0 Benchmark	25.83%
OV ST	31.11% (29.73%)
GLAD [20, 19]	35.57%
Previous SOTA (GCE) [19]	35.58%
JST	40.74% (38.42%)
HyST	44.24% (42.33%)

Equations

$$L(D)=\sum _{i=1}^N\sum _{k=1}^T\sum _{j=1}^{C_i}\log P\left(y_i^{jk}|c_i^j,s_i^k,D_i\right)$$

$$E_i=\overleftarrow{LST{M}^{sent}}(u_i)\oplus \overrightarrow{LST{M}^{sent}}(u_i)$$

$$Z_i=LST{M}^{dialogue}(E_1,...E_i)$$

$$LST{M}^{dialogueAct}(s_1,...s_k)$$

$$F_{context}=[E_i;Z_i;A_i]$$

$$\widehat{y_j}=\text{mbox}sigmoid(F{F}_k(c_i^j,F_{context})).$$

$$P_i(s_k\in S_k|Z_i)=\text{mbox}{softmax}_k(F{F}_k(Z_i))$$

$$A^k(M)=\frac{{\sum }_{i=1}^{N}1\{y_i^k={\hat{y}}_i^{kM}\}}{N}$$

$$M_{opt}=\underset{JST,OVST}{\mathrm{argmax}}(A^k(JST),A^k(OVST)).$$

Full text

Index Terms: multi-domain dialogue systems, dialogue state tracking, scaling to previously unseen data

1 Introduction

Task-oriented dialogue systems aim to enable users to accomplish tasks through spoken interactions. Dialogue state tracking in task-oriented dialogue systems has been proposed as a part of dialogue management and aims to estimate the belief of the dialogue system on the state of a conversation given the entire previous conversation context [1]. In the past decade, dialogue state tracking challenges (DSTC) [2] provided datasets and a framework for comparing a variety of methods.

In DSTC-2 [3], many systems that rely on delexicalization, where slot values from a semantic dictionary are replaced by slot labels, outperformed systems that rely on spoken language understanding outputs. More recently, an end-to-end approach that directly estimates the states from natural language input using hierarchical recurrent neural networks (RNNs) with LSTM cells has achieved the state-of-the art results [4]. However, these approaches do not scale to real applications, where one can observe natural language utterances that can include previously unseen slot value mentions and a large, possibly unlimited space of dialogue states. To deal with this scaling issue, [5] proposed the neural belief tracker approach that also eliminates the need for language understanding by directly operating on the user utterance and integrating pre-trained word embeddings to deal with the richness in natural language. However, this joint approach does not scale to a large dialogue state space as it iterates over all slot value pairs in the ontology to make a decision. More recently, Rastogi et al [6] proposed an open vocabulary candidate-set ranking approach, where the set of candidates are generated from a language understanding system’s hypotheses to deal with the scalability issue. However, this approach does not consider multi-valued slots due to the softmax layer over all the values. Other work relied on all possible n-grams from the conversation context as possible values for a candidate-set and estimated probabilities for multiple possible values [7]. While these methods were shown to handle previously unseen values, their performance is lower in comparison to the previous approaches. There are multiple possible explanations for this. For example, the first group of generative methods can deal with values that were not observed in user utterances by learning to make inferences, i.e., “fancy restaurant” could map to the backend value of “expensive” for the price range slot, whereas for the second group, the candidate-set may fail to capture “expensive” in the candidate-set, as it was not observed in the conversation context explicitly. Furthermore, some slot types may naturally resolve to few slot values, such as days of a week, which may have many instances observed in the training set. The first group of generative approaches may be more appropriate for tracking such slots.

In this paper, we analyze both the hierarchical RNN-based and the open-vocabulary candidate-generation approaches and propose hybrid state tracking, HyST, a hybrid approach for flexible and accurate dialogue state tracking, which aims to learn what method to rely on for each slot type. To investigate the appropriateness of HyST for a rich set of domains, we experiment with the recently released MultiWOZ-2.0 corpus [8] which includes single as well as multi-domain interactions. These conversations include task completion across multiple domains and allow for transfer of values between slots of different domains, as demonstrated with an example hotel-reservation and taxi-booking dialogue in Table 1. When tracking dialogue state over the 7 domains included in this corpus, our baselines outperform the previous benchmark for joint-goal accuracy (which requires estimating the correct values for all slots of all the 7 domains). Our best hybrid approach achieves a joint-goal accuracy of 44.22%, which is 4.1% (absolute) higher than our best baseline, resulting in an 24% relative improvement over the previous SOTA.

2 Related Work

Dialogue state tracking (or belief tracking) aims to maintain a distribution over possible dialogue states [9, 10], which are often represented as a set of key-value pairs. The dialogue states are then used when interacting with the external back-end knowledge base or action sources in determining what the next system action should be. Previous work on dialogue state tracking include rule-based approaches [11], Bayesian networks [12], conditional random fields (CRF) [13], recurrent neural networks [14], end-to-end memory networks [15], pointer networks [16] and embedding-based approaches [5, 17].

Previous work that investigated joint language understanding and dialogue state tracking include work by [4, 18]. Our hierarchical RNN approach is inspired by [4] and uses a hierarchical recurrent neural network to represent utterances and the dialogue flow. This approach estimates a probability for all possible values, and hence suffers from the scalability issues. Our hybrid approach aims to tackle this issue by learning to switch to a candidate-generation-based approach.

For multi-domain dialogues, previous work [19] presented results for two approaches, global-locally self-attentive dialogue state tracker (GLAD) [20] and globally conditioned encoder (GCE) [19]. GLAD is composed of an encoder and scoring modules, where the encoder uses global biLSTM modules to share parameters between estimators for all the slots and local biLSTM modules to learn slot-specific features. GCE is based on GLAD, but it simplifies the GLAD encoder by removing slot-specific recurrent and self attention layers. In our experiments, we use these two approches as baselines.

3 Methodology

A dialogue $D$ with $N$ turns is denoted as a series of agent ($a_{i}$) and user ($u_{i}$) turns i.e. $a_{1},u_{1},a_{2},u_{2},...,a_{N},u_{N}$. The task of state tracking is to predict the state ($S_{i}$) after each user turn, $u_{i}$, of the conversation. The conversation state ($S_{i}$) is commonly defined as a set of slot values, $s_{i}^{k}$, for slot types $s^{k}$, where $k\in\{1,...,T\}$ which are predefined. We define and implement the two following prevailing approaches to dialogue state tracking.

3.1 Approach 1: Open Vocabulary State tracking (OV ST)

Similar to  [7], we adopt an open-vocabulary scoring model for each slot type that needs to be tracked. The input to the model after user turn $u_{i}$ is a set of candidates $\{1,\dots,|C_{i}|\}$, where $|C_{i}|$ is number of candidates in turn $i$ that could be a value for each slot type $s^{k}$, and the conversation context $D_{i}$ ($a_{1},u_{1},...,a_{i},u_{i}$). For each candidate $c_{i}^{j}$ and for each slot type $s^{k}$, the model makes a binary decision $\hat{y}^{jk}_{i}\in[0,1]$, that denotes $c_{i}^{j}$ to be one of the values of that slot type.

If $\hat{y}^{jk}_{i}=1$, we update the value of slot $s_{i}^{k}$ with candidate $c_{i}^{j}$. For a given dialogue, we maximize the following objective:

[TABLE]

Given a user turn $u_{i}$ , we construct a candidate set for that turn. A candidate set is an open set consisting of possible slot values for each slot type. In a typical dialogue system this could be constructed from the output of a SLU system augmented with additional values obtained using simple rules (such as business logic or entity resolvers). For our experiments, we formed the candidate sets to include all word $n$-grams in user and agent utterances up-to turn $i$ in that dialogue. To reduce the size of the total candidate set, we only include those n-grams that were seen as possible slot values in the training set. We extend the candidate set with {yes, no, dontcare} as they are implied values which do not appear explicitly in the conversation. In practice, to increase coverage, such a system could also include additional values like synonyms, ASR corrections, values from a knowledge graph and resolved entities from a entity resolution system.

The system starts with a default state for every slot. After each user utterance, we update our dialogue state with candidates that are predicted as positive. Based on the system design, various update strategies or constraints can be incorporated in the dialogue state update step. For example, if we want to enforce the constraint that one slot can have only one value, we can select the candidate with the highest score from the pool of positive candidates. The dialogue history context features are flexible and we can easily add new context features by appending them to the existing context vector. For our experiments we use the following context features at each user turn $u_{i}$.

User utterance encoder ($E_{i}$): We use a biLSTM to encode each utterance, $u_{i}=w_{1}^{i},...,w_{n_{i}}^{i}$, where $n_{i}$ denotes the number of tokens in $u_{i}$ and the final utterance representation for utterance $e_{i}$ is obtained by concatenating the last hidden state of the forward lstm, $\overrightarrow{LSTM}$ and the first hidden state of the backward lstm, $\overleftarrow{LSTM}$.

[TABLE] 2. 2.

Hierarchical LSTM ($Z_{i}$): We use a unidirectional LSTM over past user utterances to encode the dialogue context.

[TABLE] 3. 3.

Dialogue Act LSTM ($A_{i}$): We use a unidirectional LSTM over agent dialogue acts to encode agent dialogue acts.

[TABLE]

We concatenate all of these features into a context feature vector $F_{context}$. The context encoders are shared for all slots. For every slot type, we have:

[TABLE]

[TABLE]

The final layer $FF_{k}$ is a feed forward layer, which estimates the probability of $c_{i}^{j}$ filling slot $k$.

3.2 Approach 2: Joint State tracking (JST)

The joint state tracking approach builds a hierarchical RNN modeling words and turns of each dialogue [4]. Similar to the open vocabulary state tracking approach, we obtain the dialogue representation $Z_{i}$ (Equation 2).The final layer is a feed forward network for each slot type $k$, $FF_{k}$, which estimates a probability distribution over all the possible values for that slot type, $S_{k}=s_{k}^{1},...,s_{k}^{|S_{k}|}$:

[TABLE]

The vocabulary of possible values, $V_{k}$, is formed of the values observed in the training set, including none, and dontcare. The hierarchical RNN layers are shared for all the slot types.

3.3 Hybrid state tracking (HyST)

We combine the two aforementioned approaches into a hybrid approach. For each slot we choose between OV ST and JST. Let

[TABLE]

be the accuracy of slot $k$ over given a approach $M$. For each slot we pick the optimal approach $M_{opt}$ as

[TABLE]

We learn the approach to pick using our development set. The slots on which the open vocabulary approach performed better on the development set are marked with ‘*’ in Table 3.

4 Data

For our state tracking experiments we use the MultiWOZ-2.0 dataset [8].The MultiWOZ-2.0 dataset consists of multi-domain conversations from 7 domains with a total of 37 slots across domains. Some of the slots, for example, day and people, occur in multiple domains. An example conversation is shown in Table 1. For our experiments, we treat each slot independently and do not share slots between domains. So, the same slot type is present in several domains and is represented by appending domain and slot names as domain.slot in Table 3. If a user turn does not have a slot value assigned to it, we mark it as None. Some dataset statistics are shown in Table 2. We also present a detailed breakdown of different slot types in MultiWOZ-2.0 in Table 3. The table includes out-of-vocabulary (OOV) slot value rates in the development set for each slot type. This is computed as the percentage of values of each slot type in the development set that was not observed in the training partition.

To showcase the complexity for different slot types, we also include percentage of turns with a “None” value for each slot type and the number of unique values (i.e., Vocabulary size) for each slot. The final row of Table 3 presents the percentage of turns whose complete state has never been observed in the training set for JST and in the candidate set generated by our OV ST approach.

5 Experimental Setup

In all experiments, we clip each turn to 30 tokens and each dialogue to past 30 turns. We use ADAM [21] for optimization with a learning rate of 0.001 and default parameters. We use a batch size of 128 while training. We initialize our embedding matrices randomly and learn them during training. We use manual search to tune all our parameters using our development set.

Open vocabulary state tracking: The model consists of four encoders: the sentence encoder, hierarchical dialogue encoder, dialogue act encoder and the candidate encoder. Our candidate encoder is an embedding lookup of dimension 300. We use the same embedding layer as input to the sentence encoder. Our sentence encoder is a biLSTM with hidden size of 256. Our sentence representation is the final state of the biLSTM. The hierarchical dialogue encoder is a LSTM with hidden size of 512 which takes the sentence representation as input. We use an embedding size of 50 for system dialog acts and encode them using a LSTM with an hidden size of 64. We concatenate these representations and pass it through a feed-forward network with an output of 256. The final 256-dimensional vector is used for a binary decision per slot type.

Joint state tracking: The joint model represents words and system actions with 300-dimensional vectors, with a hidden layer size of 200 for the utterance LSTM and 150 for the dialogue level LSTM. The agent actions were found to be not useful in the early experiments and are excluded from the final results.

6 Results

We present per domain results in Table 4. As in previous work, we report joint goal accuracy as our metric. For each user turn, we get the joint goal correct if our predicted state exactly matches the ground truth state for all the slots in that domain. As our candidate set generation is based on n-grams OOV rate for the OV oracle (Table 3) is high. This implies that the performance ceiling for this approach is around 74.4% which is much lower than the ceiling for the JST (97.5%). Still, we observe that for the slots with large vocabulary sizes (ones marked with * in Table  3), the OV approach outperforms the joint model. All slots with over 100 possible values with the exception of one, restaurant.food with a vocab size of 104, were better tracked with the OV approach. Combining the two approaches into a hybrid approach leads to the best performance on most domains except ’Hotel’ where there was no significant change.

Table 4 presents the joint goal accuracy for each domain with the three approaches. From this table, we observe a large difference in the 2 approaches for domains like Train and Hotels. One reason for this is the implicit entity resolution which is part of the dataset. For example, when a user says: ‘I want a taxi to petersbrough’, the state in MultiWOZ-2.0 gets updated as ‘To: Petersbrough, From: Cambridge’. The city Cambridge is never mentioned by the user or the agent but is implicit due to task location. JST is able to infer these relations due to similar data patterns but for OV ST, the word ‘Cambridge’ never occurs in the candidate set leading to lower performance.

We present overall joint accuracy on MultiWOZ-2.0 in Table 5. We also report results on a few baselines. The first is a dumb baseline which assigns none (the majority class) to all slots. We observe that this is a very weak baseline due to the large label space. Second, we present the benchmark baseline from the MultiWOZ-2.0 corpus [8] webpage111http://dialogue.mi.eng.cam.ac.uk/index.php/corpus/. We also report results on GLAD [20] and its extension GCE [19]. The hybrid approach is better than all baselines, and results in a joint goal accuracy of 44.24% when all the 7 domains and 37 slots are considered.

7 Conclusions

The joint tracking approach couples spoken language understanding and dialogue state tracking to achieve high accuracy on state tracking benchmarks, but this limits its performance on slots with large vocabulary as shown in our experiments. On the other hand the open-vocabulary approach is very flexible and shows better performance on large-vocabulary slots. In this work we presented HyST, a hybrid approach for dialogue state tracking by combining the aforementioned approaches. By learning to switch between the 2 approaches, our approach outperforms both of them on the challenging MultiWOZ-2.0 corpus. HyST achieves 44.2% joint goal accuracy on MultiWOZ-2.0 beating previous SOTA by over 24% (relative). Going forward we would like to experiment with better candidate representations for the OV ST approach. One exciting follow up would be enabling zero-shot state tracking by copying over values which have appeared in previous states to new domains.