PaperlayerView exact revision on arXiv

Source-linked AI summary

A Sequence-to-Sequence Model for User Simulation in Spoken Dialogue Systems

Layla El Asri, Jing He, Kaheer Suleman

arXiv:1607.00070v1cs.CL

TL;DR

Statistical spoken dialogue systems need many dialogues to learn a good strategy, making data collection costly, while prior simulators have limitations in history use and multi-act generation. This paper presents an encoder-decoder RNN that maps dialogue contexts and user-goal status to user-act sequences. It reports improved F-score performance over prior simulators and demonstrates finer-grained modeling with original actions.

  • Problem

    Collecting the thousands of dialogues needed to train statistical spoken dialogue systems is costly, and prior simulators have limitations including history use and multi-act output.

    Source: §1 · p. 1, §2 · p. 2

  • Method

    An encoder-decoder RNN reads dialogue-context sequences and generates sequences of user-intention dialogue acts.

    Source: §1 · pp. 1, 3

  • Results

    The simulator outperforms a state-of-the-art simulator on F-score and is efficiently transferred to a new information-seeking domain.

    Source: §5 · p. 4

  • Takeaways & Limitations

    Training on the original action space removes post-processing and allows user behavior to be modeled at finer granularity.

    Source: §1 · p. 3, §4.4 · p. 4

    The model assumes a user goal is uniformly drawn at the beginning of each dialogue, represented by constraints and requests.

    Source: §3 · p. 2

Abstract

from arXiv · show

User simulation is essential for generating enough data to train a statistical spoken dialogue system. Previous models for user simulation suffer from several drawbacks, such as the inability to take dialogue history into account, the need of rigid structure to ensure coherent user behaviour, heavy dependence on a specific domain, the inability to output several user intentions during one dialogue turn, or the requirement of a summarized action space for tractability. This paper introduces a data-driven user simulator based on an encoder-decoder recurrent neural network. The model takes as input a sequence of dialogue contexts and outputs a sequence of dialogue acts corresponding to user intentions. The dialogue contexts include information about the machine acts and the status of the user goal. We show on the Dialogue State Tracking Challenge 2 (DSTC2) dataset that the sequence-to-sequence model outperforms an agenda-based simulator and an n-gram simulator, according to F-score. Furthermore, we show how this model can be used on the original action space and thereby models user behaviour with finer granularity.

1. Introduction

Collecting the thousands of dialogues statistical spoken dialogue systems typically need is costly, motivating intention-level user simulation. The paper introduces an encoder-decoder RNN simulator and reports stronger F-score performance than two simulator baselines.

1. Introduction

  • Statistical spoken dialogue systems typically need several thousand dialogues to learn a good strategy, but collecting them is costly.The paper focuses on intention-level simulation, which generates dialogue acts.Source: §1 · p. 1
  • Earlier simulators are described as limited in history use, structural flexibility, domain dependence, multi-act output, or action-space granularity.The cited drawbacks include inability to output several intentions in one turn and reliance on summarized action spaces for tractability.Source: §1 · p. 1
  • The proposed simulator uses an encoder RNN to process dialogue history and a decoder RNN to generate user-intention dialogue acts.The encoder reads a sequence of dialogue contexts and passes its internal representation to the decoder.Source: §1 · p. 1
  • RNN-based models outperform the agenda-based and n-gram simulators on F-score, and generalize best to new domains on DSTC3.The evaluation uses DSTC2 for training and comparison, with DSTC3 for new-domain generalization.Source: §1 · p. 1

2. Background

Earlier user simulators range from n-gram and graph-based models to Bayesian, information-state, HMM, agenda-based, and reinforcement-learning approaches. The background emphasizes long dialogue history and multiple user acts per turn as unresolved modeling challenges.

2. Background

  • Prior approaches include bi-gram, graph-based, Bayesian, information-state, HMM, agenda-based, and reinforcement-learning user simulators.The bi-gram model reacts only to the latest system act; graph-based modeling requires extensive domain knowledge and engineering.Source: §2 · pp. 1–2
  • Table 1 gives context examples for a restaurant-seeking dialogue with cheap and downtown constraints and an address request.Source: Table 1 · p. 2
  • For tractability, previous models often omit long dialogue histories, while multi-act turns can make the output space inefficient to compute.The agenda-based approach uses a stack-like structure to produce coherent sets of acts together.Source: §2 · p. 2

3. The Sequence-to-sequence Model

The simulator maps dialogue-context sequences to action sequences, with contexts incorporating machine behavior and user-goal status. Each dialogue begins with a sampled goal comprising constraints and requests.

3. The Sequence-to-sequence Model

  • The model takes a sequence of dialogue contexts as input and outputs a sequence of actions.Source: §3 · p. 2
  • At the start of each dialogue, the model uniformly draws a user goal G = (C, R), where C contains constraints and R contains requests.Restaurant-seeking constraints can concern food type, price range, and area; requests can include those slots and restaurant information.Source: §3 · p. 2
  • Each context includes the latest machine acts, their inconsistency with the user goal, and the informed-or-not status of constraints and requests.The machine-act component is encoded as a vector whose size is the number of possible machine acts.Source: §3 · p. 2

1. The vector am,t has ones for the current machine acts

The encoder represents the full sequence of observed contexts, and the decoder generates user-act sequences autoregressively. Outputs can be mapped to goal-specific actions, or trained directly on finer-grained original actions.

1. The vector am,t has ones for the current machine acts

  • The inconsistency representation tracks goal constraints omitted by a machine proposition and constraints misunderstood when mentioned by the machine.The constraint-status vector tracks what the user has told the system; request status tracks whether requested information has been supplied.Source: §1 · p. 2
  • The LSTM encoder summarizes the observed context sequence into a vector that initializes the decoder, which generates actions step by step.Training feeds observed user-act sequences to the decoder; runtime starts from a null action and generation continues until null actions end the sequence.Source: §1 · p. 3
  • The passage defines ⊙ as concatenation and states that the LSTM is implemented following equations, including an input-gate expression involving c_t and h_{t−1}.Source: §1 · p. 3
  • Dialogue-act sequences are mapped to slot-specific actions using the current goal, remaining constraints, and remaining requests.Training directly on original actions such as request-area removes this mapping step and models behavior at finer granularity.Source: §1 · p. 3

4. Experiments

The experiments compare sequence-to-sequence user simulation with agenda-based, n-gram, and sequence-to-one models, using DSTC2 and DSTC3 data. Results cover F-score, transfer to a larger domain, and slot-level behavior in the original action space.

4.1. User Simulation Models

  • The compared models include a bi-gram predicting from the last machine compound act, an agenda-based simulator, and a sequence-to-one model predicting over 54 predefined DSTC2 compound acts.A compound user act is defined as a sequence of dialogue acts; the agenda model samples one act and may add acts popped from its stack.Source: §4.1 · p. 3, §4 · p. 3

4.2. F-score

  • The four models are compared using F-score, described as the geometric mean of precision and recall.The supplied passage states that the models are trained on the DSTC2 training set.Source: §4.2 · p. 3, §4 · p. 3

4.3. The DSTC2 dataset

  • The experiments train the models on DSTC2; its 1,612 training, 506 validation, and 1,117 test dialogues were collected with real users seeking restaurants.Training and validation dialogues used handcrafted policy managers, while the test set used a statistical policy manager.Source: §4.3 · p. 3, §4 · p. 3

4.4. Results

  • On DSTC2 validation and test, both RNN models significantly outperform the agenda-based simulator in F-score; sequence-to-one is slightly better, while sequence-to-sequence remains close.Table 2 reports average F-scores over 50 runs.Source: Table 2 · p. 4, §4.4 · p. 4
  • On DSTC3, sequence-to-one and sequence-to-sequence significantly outperform the agenda-based model, though F-score degrades, mostly in recall, with 40 compound acts unseen in DSTC2.The reported DSTC3 F-scores are similar to the agenda-based model’s DSTC2 test-set F-score.Source: Table 2 · p. 4, §4.4 · p. 4
  • The sequence-to-sequence model can use the original act space, while its output space may scale better than a predefined compound-act inventory as slot sets or dialogue-act counts grow.The paper notes that the sequence-to-one output space could become too large to train on a small dataset.Source: §4.4 · p. 4
  • For food, area, and price-range goals, first-slot rates in the corpus versus simulator are 72% vs. 48% for food, 16% vs. 31% for price range, and 12% vs. 21% for area.The experiment uses separate inform-food, inform-pricerange, and inform-area acts, avoiding a mapper and modeling slot order at finer granularity.Source: Table 3 · p. 4, §4.4 · p. 4

5. Conclusions

The proposed sequence-to-sequence simulator uses dialogue history without external structures for coherence or a summarized action space. The paper reports strong F-score performance and transfer to a new information-seeking domain.

  • The simulator accounts for the entire dialogue history without relying on an external structure to ensure coherent behavior or mapping to a summarized action space.The authors state that avoiding summarized action-space mapping allows finer-grained modeling of user behavior.Source: §5 · p. 4
  • The model outperforms a state-of-the-art simulator on F-score and transfers efficiently to a new information-seeking domain.Source: §5 · p. 4
  • Future work will use the model to train a statistical spoken dialogue system and further explore the architecture.Source: §5 · p. 4
Loading 1607.00070v1…