Proactive Human-Machine Conversation with Explicit Conversation Goals (1906.05572v2)

Abstract: Though great progress has been made for human-machine conversation, current dialogue system is still in its infancy: it usually converses passively and utters words more as a matter of response, rather than on its own initiatives. In this paper, we take a radical step towards building a human-like conversational agent: endowing it with the ability of proactively leading the conversation (introducing a new topic or maintaining the current topic). To facilitate the development of such conversation systems, we create a new dataset named DuConv where one acts as a conversation leader and the other acts as the follower. The leader is provided with a knowledge graph and asked to sequentially change the discussion topics, following the given conversation goal, and meanwhile keep the dialogue as natural and engaging as possible. DuConv enables a very challenging task as the model needs to both understand dialogue and plan over the given knowledge graph. We establish baseline results on this dataset (about 270K utterances and 30k dialogues) using several state-of-the-art models. Experimental results show that dialogue models that plan over the knowledge graph can make full use of related knowledge to generate more diverse multi-turn conversations. The baseline systems along with the dataset are publicly available

Proactive Human-Machine Conversation with Explicit Conversation Goals

The paper "Proactive Human-Machine Conversation with Explicit Conversation Goals" presents a novel approach in the domain of dialogue systems by introducing the concept of proactive conversation agents. Unlike conventional systems that respond passively, the proposed methodology endows conversational agents with the ability to actively lead discussions by maintaining or transitioning topics, thus approximating human-like interactions more closely.

Key Contributions and Methodology

The paper introduces a new dataset named DuConv, which is designed to facilitate the development of dialogue systems that plan over knowledge graphs to achieve explicit conversational goals. This leads to more natural and engaging interactions. DuConv consists of approximately 270,000 utterances and 30,000 dialogues, making it a substantial resource for training and evaluating proactive dialogue models.

In DuConv, one participant takes on the role of a conversation leader equipped with a knowledge graph, while the other participant follows. The leader strategically introduces or shifts topics according to a predefined goal. This goal is represented as a knowledge path structured as "[start] $\to$ topic_a $\to$ topic_b," where topics correspond to entities within the knowledge graph. The design challenges the dialogue model to exploit both dialogic understanding and knowledge graph planning capabilities.

Two primary methodologies are proposed to test the DuConv dataset:

1. Retrieval-Based Model: This model searches for the best response by considering dialogue context, goals, and related knowledge. Knowledge is stored in an external memory module, allowing the model to incorporate relevant data effectively.
2. Generation-Based Model: Built on a sequence-to-sequence framework, this model generates responses while selecting relevant knowledge. It employs both a prior and posterior distribution to align machine reasoning with human-mediated knowledge extraction effectively.

Experimental Outcomes

The experiments demonstrate that models trained on the DuConv dataset and equipped with knowledge-graph planning capabilities outperform traditional passive dialogue systems in generating diverse and knowledge-rich conversations. This is measured through automated metrics such as BLEU, F1 score, and knowledge precision, in addition to qualitative human evaluations. The results indicate that models implementing knowledge graphs better achieve conversation goals and maintain coherence throughout interactions.

Implications and Future Directions

This research marks a significant step towards more advanced dialogue systems capable of proactive conversations. Such systems have practical applications in customer service, personal assistants, and educational tools, where engaging and informative interactions are crucial.

Theoretically, the integration of explicit conversation goals and reasoning over knowledge graphs opens avenues for improved AI reasoning capabilities and adaptive learning in dynamic real-world environments. Future research could focus on scaling these models to more complex knowledge domains and enhancing the mechanisms for real-time reasoning, further bridging the gap between machine-generated and human conversation nuances.

In conclusion, this paper provides a robust framework for proactive dialogue systems, backed by comprehensive experimentation and a rich dataset, contributing significantly to the advancement of conversational AI.