arXiv:1903.10245v3 [cs.AI] 19 Apr 2019
Knowledge Aware Conversation Generation with Explainable Reasoning
on Augmented Graphs
Zhibin Liu Zheng-Yu Niu Hua Wu Haifeng Wang
Baidu Inc., Beijing, China
{liuzhibin05,niuzhengyu,wu
hua,wanghaifeng}@baidu.com
Abstract
Two types of knowledge, triples from knowl-
edge gr aphs and texts from unstructured doc-
uments, have been studied for knowledge
aware open-domain conversation generation,
in which triple attributes or graph paths can
narrow down vertex candidates for knowledge
selection decision, and texts can provide rich
informa tion for response generation. Fusion of
a knowledge graph and texts might yield mu-
tually reinfor c ing advantages for co nversation
generation, but there is less study on that. To
address this challenge, we propose a knowl-
edge aware cha tting machine with th ree com-
ponen ts, an augmented knowledge graph con-
taining both triples and texts, knowledge selec-
tor, and response generator. For knowledge se-
lection on the graph, we formulate it as a prob-
lem of multi-hop graph reasoning that is more
explainable and flexible in compariso n with
previous works. To fully leverage lon g text
informa tion that differentiates our graph from
others, we improve a state of the art reason-
ing algorithm with machine reading compre-
hension technology. We demonstrate that sup-
ported by such unified knowledge and explain-
able knowledge selection method, our system
can generate more appropriate and informative
responses than baselines.
1 Introduction
One of the key goals of AI is to build a ma-
chine that can talk with humans when given an
initial topic. To achieve this goal, the machine
should be able to understand language with back-
ground knowledge, recall knowledge from mem-
ory or external resource, reason about these con-
cepts together, and finally output appropriate and
informative responses. Lots of research efforts
have been devoted to chitchat oriented conversa-
tion generation (
Ritter et al., 2011; Shang et al.,
2015). However, these models tend to produce
generic responses or incoherent responses for a
given topic, since it is quite challenging to learn
semantic interactions merely from dialogue data
(
Ghazvininejad et al., 2018; Zhou et al., 2018a)
without help of background knowledge.
Recently, some previous studies have been con-
ducted to introduce external knowledge in open-
domain chitchat generation (Ghazvininejad et al.,
2018; Liu et al., 2018; Vougiouklis et al., 2016;
Young et al., 2018; Zhou et al., 2018a). These
models usually recall background knowledge
from a source, either unstructured knowledge texts
(
Ghazvininejad et al., 2018; Vougiouklis et al.,
2016) or structured knowledge triples (Liu et al.,
2018; Young et al., 2018; Zhou et al., 2018a),
and then generate more informative responses
conditioned on selected knowledge.
For knowledge triples from a graph, e.g., facts
about movies, triple attributes or graph paths can
narrow down vertex candidates for knowledge se-
lection decision. Moreover, these prior infor-
mation can enhance generalization capability of
knowledge selection models. But it suffers from
information insufficiency for response generation
since there is simply a single word or entity to
facilitate generation. For knowledge sentences,
e.g., comments about movies, the text sentences
can provide rich information for generation, but
its unstructured (e.g., document based) representa-
tion scheme demands strong capability for models
to perform knowledge selection or attention from
the list of knowledge texts. Fusion of graph struc-
ture and knowledge sentences m ight yield mutu-
ally reinforcing advantages for knowledge selec-
tion in conversation systems, but there is less study
on that.
To bridge the gap between the two lines of
studies on knowledge aware conversation gener-
ation, we present an augmented knowledge graph
based open-domain chatting m achine (denoted as
AKGCM), which consists of knowledge selec-
tor and response generator. This two-stage ar-
chitecture and graph based knowledge selection
剩余9页未读，继续阅读
资源评论
