ERICA and WikiTalk

Divesh Lala, Graham Wilcock, Kristiina Jokinen, Tatsuya KawaharaIJCAI 2019 · DOI

TL;DR

This paper introduces ERICA, a highly realistic female android, integrated with WikiTalk, a system that enables robots to converse about Wikipedia topics. The combination aims to create more natural and engaging human-robot interactions by leveraging ERICA's human-like appearance and speech with WikiTalk's extensive knowledge base.

What Problem Does This Paper Address?

Creating truly natural and engaging human-robot conversations is a significant challenge. While robots can process information, their ability to discuss a wide range of topics fluently and naturally, mimicking human conversational patterns, remains limited. Existing systems often struggle with smooth topic transitions, understanding nuanced user requests, and maintaining a coherent dialogue that aligns with a robot's persona. This paper addresses the gap in creating a robot that can not only access vast amounts of information (like Wikipedia) but also present it in a way that feels natural and human-like, enhancing the overall user experience.

Methods

The core of this work is the integration of two key components:

  • ERICA: A highly realistic female android robot designed for conversational roles. ERICA features a synthesized voice trained on a real actress, enabling realistic backchannels, laughs, and fillers. She uses a microphone array for natural speech capture and a Kinect and camera for speaker identification and behavioral sensing. Previous work on ERICA has focused on backchanneling, response generation, engagement recognition, and turn-taking.
  • WikiTalk: An interactive spoken information access system that allows robots to talk about thousands of topics using Wikipedia. The system has been previously implemented on NAO robots, incorporating features like dynamic vocabulary updates for speech recognition to predict likely next topics and enable smooth topic shifts.

For this paper, the integration focuses on how ERICA uses WikiTalk:

  • Natural Language Understanding for Topic Extraction: Instead of relying on simple keyword spotting, ERICA uses a more flexible approach. After speech recognition, the text is truecased, and then named entities and nouns are extracted using the spaCy Python module. A heuristic prioritizes named entities over compound nouns, and then single nouns, to identify the most likely topic. If multiple candidates exist, the last one mentioned is chosen.
  • Disambiguation: When a topic has multiple meanings (e.g., "John Smith"), ERICA presents a disambiguation page to the user, allowing them to confirm the intended topic.
  • Turn-Taking and Interaction: ERICA's turn-taking is more natural than previous NAO implementations. The user can simply start speaking, and ERICA recognizes the end of their turn after a period of silence. ERICA signals when she's heard the user and is searching. Users can also initiate topic shifts by simply stating the new topic, and ERICA can be interrupted with a "Stop" command.
  • Topic Shift Marking (Optional): While NAO used beat gestures, ERICA can optionally mark potential topic shifts by subtly changing her voice. This feature was not used in the demo video.

Datasets

Not explicitly discussed in the paper. The system relies on Wikipedia as its knowledge source.

Key Insights and Ideas

  • The Power of Realistic Embodiment: The paper highlights that a highly realistic robot like ERICA can significantly enhance the perceived naturalness of conversations, making the interaction more engaging. The human-like appearance sets expectations for human-like behavior.
  • Flexible Topic Extraction is Crucial: Moving beyond simple keyword matching to sophisticated NLP techniques for entity and noun extraction allows for more robust and natural understanding of user requests for information.
  • Seamless Topic Shifting Enhances Flow: The ability to smoothly transition between topics, whether related or unrelated, is key to mimicking human conversation and preventing abrupt or jarring interactions. ERICA's approach to this, including the optional voice-based marking, is a notable advancement.
  • User-Centric Interaction Design: ERICA's design prioritizes natural user input, such as allowing users to speak freely without needing to hold a microphone and recognizing turns based on silence, which lowers the cognitive load on the user.

Key Results

The paper focuses on a demonstration rather than quantitative results. The key qualitative findings are:

  • The combination of ERICA and WikiTalk results in more natural and engaging human-robot conversations.
  • ERICA can successfully extract topics from user utterances using NLP techniques, even when the requests are not explicitly phrased as commands.
  • The system demonstrates smooth topic shifts, both to related and unrelated topics, based on user prompts.
  • ERICA can handle disambiguation for ambiguous terms, presenting options to the user.
  • The interaction is designed to be natural, allowing users to speak freely and interrupt the robot.

Limitations and Future Work

  • Limitations:
    • The paper notes that the speech recognition system (Julius) uses a static, albeit large, vocabulary and is not dynamically updated with changing topics in the same way as the NAO implementation. This might limit the recognition of very new or rare topics compared to the NAO version.
  • The optional voice-based topic shift marking was not used in the demo video, so its effectiveness in that specific context is not demonstrated.
  • Future Work:
    • The authors mention ongoing work aims at developing long-term relations with human partners and adding Wikipedia-based listening to Wikipedia-based talking.
  • Further development of more natural interaction styles and flexibility in dialogue systems is implied.

Summary updated

BibTeX
@inproceedings{lala2019erica,
  author = {Divesh Lala and Graham Wilcock and Kristiina Jokinen and Tatsuya Kawahara},
  title = {{ERICA and WikiTalk}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/947},
}
Figure