PaperlayerView exact revision on arXiv

Source-linked AI summary

A Conditional Variational Framework for Dialog Generation

Xiaoyu Shen, Hui Su, Yanran Li, Wenjie Li, Shuzi Niu, Yang Zhao, Akiko Aizawa, Guoping Long

arXiv:1705.00316v4cs.CL

TL;DR

Dialog generation with latent variables can produce responses that are difficult to control, while unsupervised latent representations do not provide explicit control. This paper conditions generation on labels and models the two speakers’ dialogue states separately; tests show label-conforming responses in genericness and sentiment scenarios without significant quality decline.

  • Problem

    Seq2seq models often produce generic responses, and unsupervised latent variables in VHRED do not explicitly control response generation.

    Source: §1 · p. 1

  • Method

    The framework conditions response generation on dialogue context, a latent variable, and an external label, while modeling each speaker’s dialogue state separately.

    Source: §1 · p. 1

  • Results

    In both tested scenarios, the framework generated responses in accordance with predefined labels without significantly decreasing response quality.

    Source: §4 · p. 5

  • Takeaways & Limitations

    The results indicate that generation can be constrained to a specific class while maintaining response quality in the tested scenarios.

    Source: §4 · p. 5

    The sentiment rules cover two manually defined, common sentiment-shift conditions; handling more complex sentiments is left for future work.

    Source: §4 · p. 5

Abstract

from arXiv · show

Deep latent variable models have been shown to facilitate the response generation for open-domain dialog systems. However, these latent variables are highly randomized, leading to uncontrollable generated responses. In this paper, we propose a framework allowing conditional response generation based on specific attributes. These attributes can be either manually assigned or automatically detected. Moreover, the dialog states for both speakers are modeled separately in order to reflect personal features. We validate this framework on two different scenarios, where the attribute refers to genericness and sentiment states respectively. The experiment result testified the potential of our model, where meaningful responses can be generated in accordance with the specified attributes.

1 Introduction

Seq2seq dialogue models tend to produce safe, commonplace responses, while latent-variable approaches add richer utterances but leave generation difficult to control. This paper proposes conditioning responses on labels and modeling each speaker’s dialogue state separately, with genericness and sentiment as example attributes.

1 Introduction

  • Seq2seq models often generate generic, safe, commonplace responses such as “I don’t know.”The paper attributes this partly to token-level stochastic variation favoring immediate short rewards over long-term structure.Source: §1 · p. 1
  • VHRED uses utterance-level latent variables to support coherent global structure, but its unsupervised variables do not explicitly control generation.Those variables are only vaguely interpretable as higher-level decisions such as topic or sentiment.Source: §1 · p. 1
  • The proposed conditional variational framework generates responses from dialogue context, a latent variable, and an external label, with separate speaker context states.The paper tests labels for genericness and imitated sentiment; sentiment tags are predicted when needed to guide generation.Source: §1 · p. 1
  • The framework is intended to extend to other specific areas by changing the label definition.The introduction describes it as expressive and extendable, with responses intended to match predefined labels while remaining meaningful.Source: §1 · pp. 1–2
  • SPHRED represents Speaker A and Speaker B with separate status RNNs, then concatenates their vectors as dialogue context.This structure is presented as the context representation used by the framework.Source: Figure 1 · p. 2

2 Models

The models use separate speaker states to form dialog context, then condition response generation on that context, a latent variable, and a label. The framework covers manually assigned genericness labels and sentiment labels predicted from context.

2 Models

  • The section presents SPHRED, the conditional variational framework, and two application scenarios as the paper’s model components.The framework uses a hierarchical encoder-decoder with separated context models.Source: §2 · p. 2

2.1 SPHRED

  • SPHRED gives each speaker a separate status RNN and concatenates their states into the higher-level context vector.Each status RNN takes the encoder state from two turns earlier as input.Source: §2.1 · p. 2

2.2 Conditional Variational Framework

  • The conditional variational framework conditions each utterance’s latent variable and generation on dialog context and its label.When the next label is unknown, a classifier predicts it from the context vector.Source: §2.2 · p. 2

2.3 Scenario 1

  • For genericness control, y=1 marks generic responses and y=0 non-generic responses; the manual phrase-list rule labels about 2% of the corpus as generic.At test time, fixing y=0 is intended to produce mostly non-generic responses.Source: §2.3 · p. 3

2.4 Scenario 2

  • For sentiment control, utterances receive positive, negative, or neutral tags, with next-turn tags either held constant or set to the average of the preceding two.The tag is predicted from context, and the framework is designed to generate responses conforming to it.Source: §2.4 · p. 3

3 Experiments

The experiments evaluate the framework on Ubuntu dialogs using embedding-based metrics, label accuracy, and human judgments. SPHRED outperforms HRED and LM on embedding metrics, while the framework generates responses aligned with labels without significant quality decline.

3 Experiments

  • Models were trained and evaluated on the Ubuntu Dialog Corpus, which contains about 500,000 multi-turn dialogs.The vocabulary comprised the 20,000 most frequent words.Source: §3 · p. 3

3.1 Training Procedures

  • Training followed VHRED hyperparameters except that the context RNN dimension was halved; decoding used beam search with size 5.The framework used a 0.0001 learning rate, while other models used 0.0002.Source: §3.1 · pp. 3–4

3.2 Evaluation

  • Evaluation used embedding-based semantic similarity, label accuracy for the two scenarios, and human ratings of grammaticality, coherence, and diversity.One hundred test examples were rated by five volunteers, with each response evaluated three times.Source: §3.2 · p. 4

3.3 Results of Metric-based Evaluation

  • 90.9% of SCENE1-A responses were generic and 86.9% of SCENE1-B responses were non-generic; sentiment predictions had 0.2% and 0.8% mismatches.The reported results support the effect of the genericness labels and successful sentiment prediction.Source: §3.3 · p. 4
  • SPHRED outperforms HRED and LM on all three embedding-based metrics, while framework scores remain comparable with SPHRED and can exceed VHRED.Imposing an external label did not bring significant quality decline.Source: Table 1 · p. 4, §3.3 · p. 4

3.4 Results of Human Evaluation

  • Human evaluations found similar sentence-quality scores for VHRED and the framework, except that SCENE1-A scored lower on coherence.Scenario 1 received extreme diversity scores, while VHRED and Scenario 2 performed close to each other.Source: §3.4 · p. 4

4 Discussion and future work

The paper reports that SPHRED provides better context representation than HRED and that the framework generates responses in accordance with predefined labels without significant quality decline. Its sentiment rules are primitive, and adapting a classifier for more complex sentiments is left to future work.

4 Discussion and future work

  • SPHRED provides better context representation than HRED and helps generate higher-quality responses.The framework generates responses in accordance with predefined labels in both scenarios without significant quality decline.Source: §4 · p. 5
  • The manually defined rules are primitive and cover two common sentiment-shift conditions; real-world use would require adapting a classifier for more complex sentiments.The authors leave this adaptation for future research.Source: §4 · p. 5
  • External knowledge is a proposed direction for improving the overall quality of generated responses.Source: §4 · p. 5

Related papers

Loading 1705.00316v4…