Human-machine conversation is one of the most important topics in artificial intelligence (AI) and has received much attention across academia and industry in recent years. Currently dialogue system is still in its infancy, which usually converses passively and utters their words more as a matter of response rather than on their own initiatives, which is different from human-human conversation. Therefore, we set up a new conversation task, named proactive conversation which is a type of knowledge driven dialogue, where machines converse with humans based on a built knowledge graph. It aims at testing machines’ ability to conduct human-like conversations.
We crawled the related knowledge information from the website MTime.com, which records the information of most films, heroes, and heroines in China. We collect both structured knowledge (such as “Harry Potter” is “directed by” “Chris Columbus”) as well as unstructured knowledge including short comments and synopsis. After the raw data collection, we construct a knowledge graph. Our knowledge graph is comprised of multiple SPO (Subject, Predicate, Object) knowledge triplets, where objects can be factoid facts and non-factoid sentences such as comments and synopsis.
Given a dialogue goal g and a set of topic-related background knowledge M = f1 ,f2 ,..., fn , the system is expected to output an utterance "ut" for the current conversation H = u1, u2, ..., ut-1, which keeps the conversation coherent and informative under the guidance of the given goal. During the dialogue, the system is required to proactively lead the conversation from one topic to another. The dialog goal g is given like this: "Start->Topic_A->TOPIC_B", which means the machine should lead the conversation from any start state to topic A and then to topic B. The given background knowledge includes knowledge related to topic A and topic B, and the relations between these two topics.
Figure1.Proactive Conversation Case. Each utterance of "BOT" could be predicted by system, e.g., utterances with black words represent history H,and utterance with green words represent the response ut predicted by system.
We collected around 30k conversations containing 270k utterances named DuConv. Each conversation was created by two random selected crowdsourced workers. One worker was provided with dialogue goal and the associated knowledge to play the role of leader who proactively leads the conversation by sequentially change the discussion topics following the given goal, meanwhile keeping the conversation as natural and engaging as possible. Another worker was provided with nothing but conversation history and only has to respond to the leader.
We devide the collected conversations into training, development, test1 and test2 splits. The test1 part with reference response is used for local testing such as the automatic evaluation of our paper. The test2 part without reference response is used for online testing such as the competition we had held and the Leader Board which is opened forever in https://ai.baidu.com/broad/leaderboard?dataset=duconv. The dataset is available at https://ai.baidu.com/broad/subordinate?dataset=duconv.
We provide retrieval-based and generation-based baseline systems. Both systems were implemented by PaddlePaddle (the Baidu deeplearning framework). The performance of the two systems is as follows:
baseline system | F1/BLEU1/BLEU2 | DISTINCT1/DISTINCT2 |
---|---|---|
retrieval-based | 31.72/0.291/0.156 | 0.118/0.373 |
generation-based | 32.65/0.300/0.168 | 0.062/0.128 |
Rank | F1/BLEU1/BLEU2 | DISTINCT1/DISTINCT2 |
---|---|---|
1 | 49.22/0.449/0.318 | 0.118/0.299 |
2 | 47.76/0.430/0.296 | 0.110/0.275 |
3 | 46.40/0.422/0.289 | 0.118/0.303 |
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。