PaperlayerView exact revision on arXiv

Source-linked AI summary

Target-Guided Open-Domain Conversation

Jianheng Tang, Tiancheng Zhao, Chenyan Xiong, Xiaodan Liang, Eric P. Xing, Zhiting Hu

arXiv:1905.11553v2cs.CLcs.AIcs.LG

TL;DR

Open-domain chat applications may need to reach goals, but target-guided dialogue must balance natural conversation with proactively reaching a target, and the authors report no public learning dataset for this strategy. The paper combines coarse-grained keywords, turn-level learning, and discourse-level constraints, and reports improved, effective conversations compared with other approaches.

  • Problem

    Open-domain dialogue applications need to achieve goals during open-ended chats, but target-guided conversation must balance natural transitions with reaching the target, and the authors report no public dataset for learning that strategy.

    Source: §1 · pp. 1–2

  • Method

    The system uses coarse-grained keywords to control response content, turn-level supervised learning for smooth transitions, discourse-level constraints to approach the target, and a keyword-augmented conversation dataset.

    Source: §1 · p. 2

  • Results

    The authors report meaningful and effective conversations with a decent target-reaching success rate, improving over other approaches in different respects.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The evaluations support target-guided open-domain conversation as a promising direction for future research.

    Source: §1 · p. 2, §7 · p. 9

    Target achievement is operationalized as either speaker mentioning the target or a similar word, though the authors note that more meaningful measures may suit specific applications.

    Source: §3 · p. 3

Abstract

from arXiv · show

Many real-world open-domain conversation applications have specific goals to achieve during open-ended chats, such as recommendation, psychotherapy, education, etc. We study the problem of imposing conversational goals on open-domain chat agents. In particular, we want a conversational system to chat naturally with human and proactively guide the conversation to a designated target subject. The problem is challenging as no public data is available for learning such a target-guided strategy. We propose a structured approach that introduces coarse-grained keywords to control the intended content of system responses. We then attain smooth conversation transition through turn-level supervised learning, and drive the conversation towards the target with discourse-level constraints. We further derive a keyword-augmented conversation dataset for the study. Quantitative and human evaluations show our system can produce meaningful and effective conversations, significantly improving over other approaches.

1 Introduction

Open-domain chat agents often lack control over conversation outcomes, while practical applications need open-ended conversations to achieve specific goals. This paper proposes keyword-guided target steering, a dataset for learning it, and evaluations reporting meaningful, effective conversations that improve over other approaches.

1 Introduction

  • Practical applications such as therapeutic conversation, education, recommendation, and persuasion motivate adding goals to open-ended dialogue.The introduction identifies defining goals and encoding dialogue strategy as challenges, while also seeking a general method without specialized handcrafting and annotations.Source: §1 · p. 1
  • The system is intended to chat naturally while proactively guiding an open-domain conversation to a designated target subject.The target is unknown to the human, and the conversation may begin from an arbitrary topic.Source: Figure 1 · p. 1, §1 · p. 2
  • The system uses coarse-grained utterance keywords to control response content, discourse-level rules to move toward the target, and turn-level supervision for smooth transitions.The authors also derive a keyword-augmented dataset from an existing daily-life chat corpus to learn keyword transitions and utterance production.Source: §1 · p. 2
  • The authors report meaningful and effective conversations, a decent target-reaching success rate, and improvements over other approaches in different respects.They present target-guided open-domain conversation as a promising direction for future research.Source: §1 · p. 2

2 Related Work

Prior dialogue research addresses either goal-oriented tasks or open-domain chat, while discourse-level strategies have also been represented through latent variables or hand-crafted actions. This paper positions predicted keywords as a non-parametric representation for intended response content in open-domain conversation.

2 Related Work

  • Task-oriented dialogue systems pursue specific goals, including providing bus schedules, negotiations, and symmetric collaborations.The cited examples include bus schedules, negotiations [Lewis et al., 2017; He et al., 2018], and symmetric collaborations.Source: §2 · p. 2
  • Chat-oriented systems model open-domain conversation without specific goals and have pursued next-utterance generation or retrieval performance using large chit-chat datasets.Despite architectural improvements, prior systems can still produce dull responses or inconsistent personas [Li et al., 2016a].Source: §2 · p. 2
  • Prior discourse-level strategies use latent response representations or hand-crafted coarse-grained actions, with the latter typically focused on domains such as negotiation and movie recommendation.The cited latent-variable work includes Zhao et al. (2017), Yarats and Lewis (2018), and Zhao et al. (2019); hand-crafted action work includes Williams et al. (2017) and He et al. (2018).Source: §2 · p. 2
  • The paper frames its approach as dividing a task lacking full supervision into sub-objectives addressed through direct supervision or simple rules.The passage relates this divide-and-conquer approach to text style transfer and content manipulation [Lin et al., 2019].Source: §2 · pp. 2–3

3 Task Definition: Target-guided Open-domain Conversation

The task is to guide a conversation from an arbitrary initial topic to a target while keeping responses contextually smooth. The authors describe these aims as competing and note the lack of public large-scale data for learning the required discourse strategy end to end.

3 Task Definition: Target-guided Open-domain Conversation

  • A target is defined as a word, is known only to the agent, and is counted as achieved when either speaker mentions it or a similar word.The authors note that more complex targets and more meaningful achievement measures could be used for specific applications.Source: §3 · p. 3
  • The agent must balance transition smoothness with target achievement: an abrupt target mention may be unnatural, while only reactive smooth replies may never approach the target.The passage illustrates the first tension with a response that introduces “cat” after a human says they went to a movie.Source: §3 · p. 3
  • The authors report no public large dataset suited to this setting that permits end-to-end learning of discourse-level target-guiding strategy.Available open-ended conversation data generally comes from interlocutors chatting freely without a specified end target.Source: §3 · p. 3
  • The proposed response is to break down the task, use partial supervision, and introduce structure, then present an approach and an open-ended conversation dataset.Source: §3 · p. 3

4 The Proposed Approach

The approach separates turn-level smoothness from target guidance by predicting coarse-grained response keywords, applying a discourse-level rule, and conditioning response retrieval on the selected keyword.

4.1 Turn-level Keyword Transition

  • The paper explores PMI-based keyword-pair transitions, an end-to-end neural predictor, and a hybrid kernel method for next-turn keyword prediction.The PMI method sums pairwise scores for current and candidate keywords and normalizes them into a next-turn distribution.Source: §4.1 · p. 4
  • The hybrid method combines embedding-similarity kernel features with neural scoring, summing features from multiple current keywords to model multiple-to-one transitions.Its candidate scores are normalized into a probability distribution.Source: §4.1 · p. 4
  • The hybrid kernel-based approach provides the strongest performance among the studied keyword-transition approaches.The empirical study reports this as the strongest-performing approach.Source: §4.1 · p. 5

4.2 Discourse-level Target-Guided Strategy

  • The discourse-level strategy permits only next-turn keywords with greater cosine similarity to the target than the current keyword.The turn-level predictor selects or samples from this valid candidate set.Source: §4.2 · p. 5

4.3 Keyword-augmented Response Retrieval

  • The response retriever scores candidate responses using both conversation-history and predicted-keyword features.The two candidate-feature interactions are concatenated and passed to a sigmoid scoring layer.Source: §4.3 · p. 5

5 Dataset

The dataset adapts PersonaChat with automatically extracted utterance keywords for training turn-level transitions and response retrieval.

5 Dataset

  • PersonaChat conversations support learning smooth turn-level transitions through varied topics, but lack target-guided behavior for learning discourse-level strategies.The corpus was originally created without end targets.Source: §5 · p. 5
  • The adaptation removes persona information, extracts keywords using TF-IDF and part-of-speech features, and creates train, validation, and test splits.The test set contains 500 conversations with relatively frequent keywords.Source: §5 · p. 6
  • The resulting dataset trains both the turn-level transition and response-retrieval modules, using 19 randomly sampled negative responses per turn for retrieval training.Source: §5 · p. 6

6 Experiments

Experiments test turn-level keyword prediction and response selection, then evaluate target-guided conversation through simulation, human studies, and examples. The Kernel system leads the reported comparisons, though qualitative analysis notes occasional unnatural transitions.

6.1 Experimental Setup

  • The comparison includes unconditioned Retrieval, keyword-guided Retrieval-Stgy, and Kernel, PMI, Neural, and Random transition variants; shared training uses 200-dimensional embeddings and hidden states.Retrieval-Stgy retrieves responses containing a candidate keyword, while Random selects the next keyword randomly.Source: §6.1 · p. 6

6.2 Turn-level Evaluation

  • Kernel leads on all turn-level metrics across keyword prediction and response selection except R20@3, where PMI performs best; keyword-augmented systems significantly outperform base Retrieval on response selection.Random transition performs close to base Retrieval, suggesting erroneous keywords can be ignored after training.Source: §6.2 · pp. 6–7, Table 3 · p. 7

6.3.2 Human Evaluation

  • Kernel outperforms all comparison systems in human-rated target success and transition smoothness, and earns 30–50% wins in paired rankings.The paired rankings compare Kernel with other agents; the rating results are reported as consistent with simulation success rates.Source: Table 6 · p. 7, Table 5 · p. 7, §6.3.2 · pp. 8–9

6.3.1 Self-Play Simulation

  • In 500 simulations, Kernel has the highest target success rate; base Retrieval stays below 10%, while Retrieval-Stgy takes 6.56 turns to reach targets.Around 81% of Kernel’s predicted keywords occur in its produced utterances.Source: §6.3.1 · p. 7, Table 4 · p. 7

6.4 Qualitative Study

  • Examples show Kernel reaching dance via ride and music more quickly than Retrieval-Stgy, and reaching the relatively uncommon McDonald’s target by turn four; a failure example has an illogical transition despite reaching its target.The listen response is inappropriate after the human’s sports keyword.Source: §6.4 · p. 9

7 Conclusions & Discussions

The paper presents a modular approach to target-guided open-domain conversation and reports promising, improved results from quantitative and human evaluations. It frames the work as an initial bridge to task-oriented dialogue, while describing its keyword treatment as preliminary for complex real applications.

7 Conclusions & Discussions

  • Quantitative and human evaluations demonstrate promising, improved results for the proposed approach.Source: §7 · p. 9
  • The approach uses coarse-grained keywords as a logical backbone, partial supervision, and heuristic rules, and the paper derives a dataset for the task.Source: §7 · p. 9
  • The work is an initial bridge between open-domain chit-chat and task-oriented dialogue, but simple keywords may be preliminary for complex applications.The paper describes possible uses in guiding users toward task-oriented systems and calls for more fine-grained sentence- and discourse-level control.Source: §7 · p. 9
Loading 1905.11553v2…