Source-linked AI summary
A Network-based End-to-End Trainable Task-oriented Dialogue System
Tsung-Hsien Wen, David Vandyke, Nikola Mrksic, Milica Gasic, Lina M. Rojas-Barahona, Pei-Hao Su, Stefan Ultes, Steve Young
TL;DR
Task-oriented dialogue development traditionally requires multiple handcrafted or separately supervised components, with limited in-domain data and difficult reinforcement-learning design choices. The paper introduces an end-to-end trainable neural dialogue model with explicit database representations and a pipelined Wizard-of-Oz data-collection framework. The learned system interacts naturally with human subjects to complete an application-specific task, while the collection framework produces task-oriented data quickly at modest cost.
Problem
Task-oriented dialogue systems require application-specific components, training corpora, and carefully designed state, action, and reward structures.
Method
The paper combines an end-to-end trainable neural dialogue model with explicit database slot-value representations, distributed intent representations, and pipelined Wizard-of-Oz crowdsourcing.
Results
The learned model interacted efficiently and naturally with human subjects to complete an application-specific task.
Takeaways & Limitations
Pipelined parallel crowdsourcing enabled good-quality task-oriented dialogue data to be collected quickly at modest cost.
Takeaways & Limitations
The text-based model cannot directly handle noisy speech-recognition inputs or ask users for confirmation when uncertain, and scalability to larger domains remains open.
Abstract
from arXiv · showhide
Teaching machines to accomplish tasks by conversing naturally with humans is challenging. Currently, developing task-oriented dialogue systems requires creating multiple components and typically this involves either a large amount of handcrafting, or acquiring costly labelled datasets to solve a statistical learning problem for each component. In this work we introduce a neural network-based text-in, text-out end-to-end trainable goal-oriented dialogue system along with a new way of collecting dialogue data based on a novel pipe-lined Wizard-of-Oz framework. This approach allows us to develop dialogue systems easily and without making too many assumptions about the task at hand. The results show that the model can converse with human subjects naturally whilst helping them to accomplish tasks in a restaurant search domain.
1 Introduction
Task-oriented dialogue systems require application-specific components, training corpora, and carefully designed reinforcement-learning spaces and rewards. The paper proposes an end-to-end trainable neural model and a pipelined Wizard-of-Oz crowdsourcing process to reduce these development burdens.
- Existing systems still require supervised corpora for language understanding and generation despite reinforcement learning for dialogue policies.
- Designed state and action spaces can restrict reinforcement-learning models, while reward functions are difficult to design and measure at run-time.
- The proposed model is end-to-end trainable but modularly connected, combining explicit database slot-value representations with distributed user-intent representations.
- Delexicalisation and weight tying reduce the data required for training while preserving flexibility for larger datasets.
- A pipelined Wizard-of-Oz crowdsourcing mechanism collects human-human dialogue corpora online with fast, low-cost development.
2 Model
The model maps user utterances to intent and belief-state representations, queries a database, selects an action, and generates delexicalised responses that are later lexicalised. Its modular components include neural encoders, slot-specific belief trackers, a database operator, a policy network, and response generation.
- User tokens are converted into a distributed intent representation and a probability distribution over slot-value pairs called the belief state.
- The intent network encodes each user utterance into a distributed vector, using either an LSTM or a CNN encoder.
- Delexicalisation removes slot-value-specific information from encoded inputs, allowing the vector to represent distributed intent rather than hand-coded dialogue acts.
- 2.2 Belief Trackers: Each ontology slot has a specialised Jordan-type RNN tracker with a CNN feature extractor that maintains distributions over informable and requestable slot values.
- 2.3 Policy Network and Database Operator: The database operator turns belief-tracker outputs into a query and returns a binary vector indicating which database entities match it, while maintaining an entity pointer.
- 2.3 Policy Network and Database Operator: The policy network combines intent, belief-state, and database information into a system-action vector that conditions response generation.
- 2.4 Generation Network: The generation network replaces generic slots and values with surface forms and selected database attributes after producing the skeletal response.
- 2.4 Generation Network: Attention can dynamically combine tracker belief states at each output step instead of decoding from a static action vector.
3 Wizard-of-Oz Data Collection
The paper addresses the training-data bottleneck in task-oriented dialogue by introducing a crowdsourced, pipelined Wizard-of-Oz process for collecting domain-specific dialogues. The procedure supports parallel collection with low development costs and is demonstrated in a Cambridge restaurant-search domain.
- In-domain dialogue data is essential for task-oriented systems, making appropriate training-corpus collection a major development bottleneck.
- The proposed method is a crowdsourcing version of the Wizard-of-Oz paradigm for collecting domain-specific dialogue corpora.
- Users receive entity-finding tasks and submit natural-language turns, while wizards record conveyed slot-value information through dedicated webpages.
- Single-turn contributions and review of previous turns create pipelined dialogues that support large-scale parallel collection without response-waiting delays.
- 3000 HITs were run over roughly 3 days for a Cambridge restaurant-search domain containing 99 restaurants.
4 Empirical Experiments
The experiments evaluate tracker quality, corpus-based response generation, and human-facing dialogue performance. The proposed architecture benefits especially from explicit request tracking, attention, and weighted decoding for task success.
- Tracker performance: CNN-based trackers achieve high precision, while their recall advantage over N-gram trackers suggests better generalization to long-distance dependencies.The comparison attributes this pattern to CNNs handling more complex syntactic structures.
- Architecture variants: Without requestable trackers, task completion falls to ∼30% despite ∼90% correct-entity offers.The variant leaves user-request modeling to the intent network alone.
- Architecture variants: The CNN intent variant reaches a competitive BLEU score but only ∼58% task success, indicating weaker task completion than the full architecture.The authors associate this with the CNN’s local-feature encoding and lack of a global sentence view.
- Decoding and attention: Attention improves BLEU by ∼0.01 and task success by ∼5%, while weighted decoding adds about 3% task success improvement without significant BLEU gains.The weighted strategy contributes additional task-specific information during decoding; combining it with attention improves performance further.
- Human evaluation: Across 245 human-evaluated dialogues, the full attention-and-weighted-decoding model achieves a 98% average subjective success rate.Human evaluation also assessed perceived comprehension and response naturalness on five-point scales.
- Human evaluation: Against the handcrafted HDC baseline, the NN system is rated better on all compared metrics while achieving similar success rates, fewer turns, and higher preference.The NN system also receives higher comprehension and naturalness ratings.
5 Conclusions and Future Work
The paper presents an end-to-end neural framework and a pipelined crowdsourced Wizard-of-Oz data-collection method for task-oriented dialogue. Experiments show natural, efficient task completion, while broader-domain scaling and speech handling remain open challenges.
- The model is end-to-end trainable using two supervision signals and a modest corpus of training data.
- The pipelined crowdsourcing framework enables good-quality task-oriented dialogue data to be collected quickly at modest cost.
- The learned model interacts efficiently and naturally with human subjects to complete an application-specific task.
- The authors describe this as the first end-to-end neural-network model known to conduct meaningful dialogues in a task-oriented application.
- The text-based system cannot directly handle noisy speech-recognition inputs or ask users for confirmation when uncertain.
- Scaling the model to much larger and wider domains remains an open question.
Scoring Table
Table 5 defines an additional reward term for delexicalised tokens used with weighted decoding.
- The additional reward term encourages or discourages generating delexicalised tokens based on tracker observations.A positive score encourages the token; a negative score discourages it.
Appendix: Sample dialogues
Table 6 presents sample conversational logs between online judges and the end-to-end system.
- The table contains real conversational logs between online judges and the end-to-end system.