Logician and Orator: Learning from the Duality between Language and Knowledge in Open Domain
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Logician and Orator: Learning from the Duality between Language and
Knowledge in Open Domain
Mingming Sun1,2 , Xu Li1,2 , Ping Li1
1
Big Data Lab (BDL-US), Baidu Research
2
National Engineering Laboratory of Deep Learning Technology and Application, China
{sunmingming01,lixu13,liping11}@baidu.com
Abstract traction” or “relation classification”, which iden-
We propose the task of Open-Domain Infor- tifies instances of a fixed and finite set of rela-
mation Narration (OIN) as the reverse task of tions from natural language corpus, using super-
Open Information Extraction (OIE), to imple- vised methods (Kambhatla, 2004; Zelenko et al.,
ment the dual structure between language and 2003; Miwa and Bansal, 2016; Zheng et al., 2017)
knowledge in the open domain. We then de- or weakly supervised methods (Mintz et al., 2009;
velop an agent, called Orator, to accomplish
Lin et al., 2016). In the meantime, the close-
the OIN task, and assemble the Orator and the
recently proposed OIE agent – Logician (Sun domain IN (CIN) task (Wiseman et al., 2017;
et al., 2018) into a dual system to utilize the Chisholm et al., 2017; Agarwal and Dymetman,
duality structure with a reinforcement learning 2017; Vougiouklis et al., 2017; Yin et al., 2016)
paradigm. Experimental results reveal the dual transforms a set of facts with a pre-defined schema
structure between OIE and OIN tasks helps to or relation types (such as facts from Freebase (Bol-
build better both OIE agents and OIN agents. lacker et al., 2008), DBpedia (Auer et al., 2007),
or database tables), into natural language sen-
1 Introduction
tences/documents. Furthermore, the dual structure
The duality between language and knowledge is between CIE and CIN tasks has been noticed and
natural for human intelligence. The human can ex- utilized in (Chisholm et al., 2017).
tract knowledge from natural language to learn or For the open-domain problem, the open-domain
remember, and then narrate the knowledge back IE (OIE) task is to investigate how the natural
to natural language to communicate. Information language sentences express the facts, and then
extraction (IE) is a task to simulate the first part use the learned knowledge to extract entity and
of the duality, which is a long-term hot spot for relation level intermediate structures from open-
NLP research. Recently, the task that fulfills the domain sentences (Christensen et al., 2011; Et-
last part of the duality, that is, assembling a set zioni et al., 2011; Schmitz et al., 2012; Pal and
of relation instances/facts or database records into Mausam, 2016). Although the OIE task has at-
natural language sentences/documents, has also tracted much interests and obtained many appli-
attracted many interests (Wiseman et al., 2017; cations (Christensen et al., 2013, 2014; Mausam,
Chisholm et al., 2017; Agarwal and Dymetman, 2016; Stanovsky et al., 2015; Khot et al., 2017;
2017; Vougiouklis et al., 2017; Yin et al., 2016). In Fader et al., 2014), the OIN task has not been
the literature, this task has been referred to as “data stated, neither the duality between the language
to document generation” (Wiseman et al., 2017) and knowledge in the open domain.
or “knowledge-to-text” (Chisholm et al., 2017). In
this paper, we name the task as information narra- Open-Domain Closed-Domain
tion (IN), to emphasize the reverse relationship to Extraction OIE CIE
the information extraction (IE) task. Narration OIN CIN
The duality between language and knowledge
Table 1: Taxonomy: Tasks between knowledge and
(and thus between the IE and IN tasks) can be natural language.
examined in closed-domain or open-domain. For
the closed-domain problem, the closed-domain IE The tasks involved in the duality between lan-
(CIE) task is often referred to as “relation ex- guage and knowledge is shown in Table 1, where
2119
Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2119–2130
Brussels, Belgium, October 31 - November 4, 2018. c 2018 Association for Computational Linguisticsthe OIN task has not been stated. In this paper, CIN agents face problems from different prob-
we focus on the OIN task and the duality between lem domains, from people biographies to basket-
the OIE and OIN tasks, for following reasons: 1) ball game records, but most of them follow the
the OIN task is an essential component for open- same sequence-to-sequence pattern. First, the al-
domain information processing pipeline. For ex- gorithm encodes a sequence of facts into a set
ample, it is helpful for building natural and in- of annotations and then decodes the annotations
formative response for open-domain KBQA sys- into a natural language text. Mechanisms such
tems (Khot et al., 2017; Fader et al., 2014). 2) (as as attention (Bahdanau et al., 2014) and copy-
the results in this paper will illustrate) the duality ing (Gu et al., 2016) are employed into the de-
between tasks can be valuable for building better coder to improve the performance. Then, the mod-
agents for both tasks (Xia et al., 2017, 2016). els are trained on a supervised dataset with back-
A major historical obstacle for investigating propagation.
the duality between OIE and OIN tasks is the In this work, we adopt a similar sequence-to-
absence of parallel corpus between natural lan- sequence architecture to build our baseline Ora-
guage sentence and open-domain facts. Recently, tor agent, but with following differences: 1) the
the SAOKE dataset (Sun et al., 2018) was re- Orator is proposed to narrate open domain facts,
leased, which contains more than forty thou- where the encoder must encode words rather than
sand of human-labeled open-domain sentence- the entities and relations in the closed domain; 2)
facts pairs, and thus essentially eliminates the ob- the baseline Orator will be fine tuned using the
stacle for our investigation. dual learning algorithm proposed in this paper.
The contribution of this paper lies in following
aspects: 2.2 Dual Learning Systems
For many natural language processing tasks, there
• We propose the concept of OIN task, which is exist corresponding reverse/dual tasks. One ex-
potentially an important component for open- ample of a pair of dual problems is the question
domain information pipeline. We develop the answering (QA) and question generation (QG). In
Orator agent to fulfill the task; (Tang et al., 2017), the duality between QA prob-
lem and QG problem was considered as a con-
• We build a multi-agent system with Logician
straint that both problems must share the same
and Orator to exploit the dual structure be-
joint probability. Then, a loss function that im-
tween language and knowledge in open do-
plemented the constraint was involved in the su-
main. Experimental results reveal that the
pervised learning procedures for both agents. Fur-
dual information is beneficial for improving
thermore, researchers (Tang et al., 2017; Duan
the performance of both agents.
et al., 2017; Sachan and Xing, 2018) use both
the question-answering agent and the question-
The paper is organized as follows: Section 2
generation agent to identify extra high-confident
discusses the related work. Section 3 explains the
question-answering pairs, which are further used
Orator agent for OIN task. Section 4 describes the
to fine tune the pre-trained agents.
multi-agent system with Logician and Oration and
its algorithm to learn from the duality between lan- Back-and-forth translation (or round-trip trans-
guage and sentence. The experimental results of lation) 1 is another example of duality, in the field
the fine tuned agents are shown and discussed in of machine translation. It has been employed to
Section 5. We conclude our work and discuss the evaluate the quality of machine translation sys-
future direction in Section 6. tems (Van Zaanen and Zwarts, 2006), or to test
the suitability of text for machine translation (Gas-
2 Related Work pari, 2006; Shigenobu, 2007). Recently, (Xia
et al., 2016) implemented the duality in a neural-
2.1 Information Narration based dual learning system, in which the quality
of each translation agent was improved on the un-
The closed-domain information narration (CIN)
labeled dataset using the rewards provided by its
task has been studied in (Wiseman et al., 2017;
Chisholm et al., 2017; Agarwal and Dymet- 1
https://en.wikipedia.org/wiki/
man, 2017; Vougiouklis et al., 2017). These Round-trip_translation
2120Chinese English Translation
Sentence 他发明了有两个固定点（水的沸 He invented a 100 degree centimeter temperature
点和冰点）的100摄氏度计温尺， scale with two fixed points (the boiling point of wa-
是世界大多数地区通用的摄氏温 ter and the freezing point), which is the precursor of
度计的前身。 the Celsius thermometer in most parts of the world.
Facts (他,发明,100摄氏度计温尺)(100摄 (He, invented, the 100 degree Celsius temperature
氏 度 计 温 尺,有,两 个 固 定 点)(水 scale) (100 Celsius temperature scale, has, two fixed
的[沸 点|冰 点],ISA,两 个 固 定 points) (water [boiling point | freezing point], ISA,
点)(100摄 氏 度 计 温 尺,是X的 前 two fixed points) (100 degrees Celsius temperature
身,世界大多数地区通用的摄氏温 scale, is the predecessor of X, most Celsius ther-
度计) mometers in most parts of the world)
Table 2: An example sentence and the corresponding facts in the SAOKE dataset, where “ISA” is a symbol denoting
the “is-a” relationship in the SAOKE format.
corresponding dual agent, using the reinforcement related rewards. Furthermore, our approach is
learning technique. more adaptable than the weight sharing approach
adopted in (Chisholm et al., 2017).
Parsing-reconstruction is also a pattern of du-
ality. (Konstas et al., 2017) considered the AMR 3 Orator
(Abstract Meaning Representation)(Banarescu
et al., 2013) parsing problem (text to AMR) 3.1 SAOKE Dataset
and AMR generation problem (AMR to text) in Symbolic Aided Open Knowledge Expression
one system, in which the AMR parser generated (SAOKE) is proposed in (Sun et al., 2018) as the
extra text-AMR pair data to fine tune the AMR form to honestly record the facts that humans can
generator. The AMR generator, however, does not extract from sentences when humans read them.
contribute to the performance improvement of the SAOKE uses a unified form - an n-ary tuple:
AMR parser. The CIE agent and CIN agent in
(Chisholm et al., 2017) also follow this pattern, (subject, predicate, object1 , · · · , objectN ),
where both agents help each other to improve
by sharing weights. Nevertheless, the sharing to express four categories of facts: 1) Rela-
weight strategy cannot be applied to agents with tion: Verb/preposition-based n-ary relations be-
different architecture, which is a typical situation tween entity mentions; 2) Attribute: Nominal at-
in practice. tributes for entity mentions; 3) Description: De-
scriptive phrases of entity mentions; 4) Concept:
From these practices, it can be seen that the
Hyponymy and synonymy relations among con-
duality can be implemented with two major ap-
cepts and instances.
proaches: 1) by providing additional labeled sam-
Using this SAOKE format, Sun et al. (2018)
ples via bootstrapping, and 2) by adding losses or
manually labeled the SAOKE dataset DSAOKE by
rewards to the training procedure of the agents. In
crowdsourcing, which includes more than forty
this paper, we follow the second approach. We
thousand sentence-facts pairs < S, F >.2 The la-
design a set of rewards, among which some are
beling procedure is under the supervision of the
related to OIE and OIN tasks respectively, and
“Completeness” criterion (Sun et al., 2018), so the
some are related to the duality of the problems.
facts recorded information in the sentence as much
Then we optimize both agents using the reinforce-
as possible (only auxiliary information and rela-
ment learning technique. The learning algorithm
tion between facts are omitted (Sun et al., 2018)).
is similar to the dual-NMT algorithm described
As a result, the SAOKE dataset is a valid open-
in (Xia et al., 2016), but with adaption for the OIE
domain sentence-facts parallel dataset for both
and OIN tasks, especially on the task related re-
OIE and OIN tasks. Table 2 is an example from
wards. Compared to the approach of applying the
the SAOKE dataset for an easy understanding of
regularization about sharing the same joint prob-
the dual relationship between sentence and facts.
ability (Tang et al., 2017), our approach directly
2
optimizes the task objective by introducing task http://ai.baidu.com/broad
21213.2 Model where pF is the probability of copying from F and
pV is the probability of selecting from V . The de-
The Orator is an agent O that assembles a set of
tails can be found in (Gu et al., 2016).
open-domain facts F into a sentence S with prob-
ability PO (S|F, ΘO ), where ΘO is the set of pa- 3.2.3 Coverage Mechanism
rameters of O : To cope with the problem of information lost or
redundancy in the generated sentence, the copied
Orator O: F → PO (S|F, ΘO ). histories of previous generated words should be
remembered to guide future generation. This
For each pair < S, F >∈ DSAOKE , the set could be done through the coverage mecha-
of facts F is actually expressed as a sequence of nism (Tu et al., 2016), in which a coverage vec-
facts, in the order of the labeler wrote them. So, tor mtj is introduced for each word wjF in F
the deep sequence to sequence paradigm is suit- and updated at each step t as a gated function of
able to model the Orator. In this work, we build hF t−1
j , αtj , st−1 , mj . By this means, the coverage
the base Orator model with the attention-based se- vectors remember the historical attentions over
quence to sequence model, together with copy and source sequence and can be incorporated in the
coverage mechanism, in a similar way of the im- alignment model to generate complete and non-
plementation of the Logician in (Sun et al., 2018). redundant sentences. Detailed formulations can be
found in (Tu et al., 2016) and (Sun et al., 2018).
3.2.1 Attention based Sequence-to-sequence
Learning 4 Learning the Dual Structure between
The attention-based sequence-to-sequence learn- Knowledge and Natural Language
ing (Bahdanau et al., 2014) first encodes the in-
4.1 Dual Structure between Orator and
put fact sequence F (actually the sequence of Ne -
Logician
dimensional word embedding vectors) into a Nh -
dimensional hidden states H F = [hF F In (Sun et al., 2018), an agent L, called Logician,
1 , · · · , hNS ]
using bi-directional GRU (Gated Recurrent Units) was trained to convert a sentence S into a set of
network (Cho et al., 2014). Then, when gen- facts F with probability PL (F|S, ΘL ), where ΘL
erating word wt of the target sentence, the de- is the set of parameters of L:
coder computes the probability of generating wt
Logician L: S → PL (F|S, ΘL ).
by p(wt |{w1 , · · · , wt−1 }, ct ) = g(ht−1 , st , ct ),
where st is the hidden state of the GRU decoder, Obviously, the Logician and Orator can coop-
g is the word generation model, and ct is the dy- erate to supervise each other. Given < S, F >∈
namic context vector which focuses attention on DSAOKE , the Logician produces a predicted set of
specific location l in the input hidden states H F . facts F ∗ for the sentence S, and the Orator can
For the Orator, we use the copy mechanism to calculate the probability PO (S|F ∗ , ΘO ) of recon-
implement the word generation model g and use struction S from F ∗ . Intuitively, if F ∗ loses major
the coverage mechanism to compute the dynamic information of S, honestly reconstructing S from
context vector ct . F ∗ would be impossible, and thus the probabil-
ity PO (S|F ∗ , ΘO ) would be small. Thus, it is a
3.2.2 Copy Mechanism
strong signal to evaluate the quality of F ∗ . Simi-
In the SAOKE dataset, the words in the set of facts larly, when the Orator produces a sentence S ∗ for
(excluding the external symbols) must be in the the set of facts F, the probability PL (F|S ∗ , ΘL )
corresponding sentence, so the problem is suitable provided by the Logician is a strong signal for
to be modeled via the copy mechanism (Gu et al., evaluating the quality of S ∗ . These signals are
2016). In the copy mechanism, when the decoder helpful to conquer several problems of the origi-
is considering generating a word wt , it can either nal agents, including information lost, information
be copied from the source fact sequence F or se- redundancy, and non-fluency.
lect from a vocabulary V : Note that the supervision signals
PO (S|F ∗ , ΘO ) and PL (F|S ∗ , ΘL ) do not rely on
p(wt |wt−1 , st , ct ) = pF (wt |wt−1 , st , ct ) + any supervised parallel corpus. Thus, similar to
pV (wt |wt−1 , st , ct ), the application of dual learning paradigm on NMT
2122task (Xia et al., 2016), it is theoretically possible 4.2.2 Similarity Rewards
to use unparalleled sentences and sets of facts to Since the SAOKE dataset has label information,
compute these signals. However, unsupervised the similarities between the predicted results and
collections of fact-groups that can be reasonably the ground truths can be used as rewards.
narrated in a sentence are not naturally available. For the Orator, since the S ∗ can be viewed as
Currently, the only available collection is the the summarization of S, we use the widely used
sets of facts provided by the SAOKE dataset, ROUGE-L (Lin, 2004) measure in the text sum-
where the supervised information is available. marization field to evaluate the quality of S ∗ :
As a result, we implement our dual learning
system in a supervised approach, which uses the SO = ROUGEL (S, S ∗ ).
reinforcement learning algorithm to optimize the
Orator and the Logician. The involved rewards For the Logician, we use following procedure to
are described in the next subsection, and then the calculate the similarity between F and F ∗ . First,
algorithm is detailed in the last subsection. we compute the similarity between each predicted
fact f ∗ ∈ F ∗ and each ground truth fact f ∈ F
4.2 Rewards with following measure:
Given < S, F >∈ DSAOKE , we sample a set Pmin(|f ∗ |,|f |)
of facts F ∗ from distribution PL (·|S, ΘL ) and a ∗ i=1 SimStr(fi∗ , fi )
SimF act(f , f ) = ,
sentence S ∗ from distribution PO (·|F, ΘO ). Fol- max(|f ∗ |, |f |)
lowing rewards are introduced into the proposed
dual learning system, and the relationships be- where fi∗ and fi denote the i-th element of tu-
tween them are shown in Figure 1. ples of fact f ∗ and f , SimStr(·, ·) denotes the
gestalt pattern matching (Ratcliff and Metzener,
VL (Fi ) 1988) measure for two strings, and | · | is the car-
dinality function. Then, each predicted fact in
F ∗ is aligned to its corresponding ground-truth
Logician fact in F by solving a linear assignment prob-
S {Fi } lem (Wikipedia, 2017) to maximize the sum of
RO (S, Fi ) similarities between the aligned facts. Finally, the
similarity reward for the Logician is calculated by:
SO (Sj , S) SL (F, Fi ) P
∗ SimF act(f ∗, f )
SL (F , F) = ,
RL (F, Sj ) max(|F ∗ |, |F|)
{Sj } F where f ∗ ∈ F ∗ , f ∈ F are aligned pair of facts.
Orator
4.2.3 Validity Rewards
VO (Sj ) For the Orator, the output is expected as a valid
natural language sentence, so the validity reward
Figure 1: Illustration of the dual learning system of Lo-
can be defined as:
gician and Orator.
VO (S ∗ ) = LM (S ∗ ),
where the LM (·) is a language model.
4.2.1 Reconstruction Rewards For the Logician, the output should represent a
Following the idea described in above subsection, valid collection of facts, which means: 1) the out-
we design the reconstruction reward for the Orator put can be parsed into a collection of facts; 2) there
as: is no duplicated fact (identified by the SimF act
RO (S ∗ , F) = log PL (F|S ∗ , ΘL ), value larger than 0.85) in the parsed collection.
The validity reward for Logician is defined as:
and that for the Logician as:
(
0 if F ∗ is valid;
RL (F ∗ , S) = log PO (S|F ∗ , ΘO ). ∗
VL (F ) =
−1 otherwise.
2123Algorithm 1 A simple dual-learning algorithm for facts extraction and expression
Require:
A set of sentence-facts pairs {< S, F >};
An initial Logician L and an initial Orator O;
Beam size K;
repeat
1: Sample a sentence-facts pair < S, F >; 1: Sample a sentence-facts pair < S, F >;
2: Logician produces K sets of facts 2: Orator produces K sentences S1 , · · · , SK
F1 , · · · , FK from S via beam search; from F via beam search;
3: for each set of facts Fi do 3: for each sentence Si do
4: Compute the reward for Fi as: 4: Compute the reward for Si as:
riF = α1 RO (S, Fi )+α2 VF (F)+α3 SF (F, Fi ). riS = β1 RL (F, Si )+β2 VS (Si )+β3 SS (S, Si ).
5: end for 5: end for
Compute the total reward r = K1 K S
P
1 PK
i=1 ri ;
6: Compute the total reward r = K F 6:
i=1 ri ;
7: Compute the stochastic gradient of ΘL : 7: Compute the stochastic gradient of ΘO ,
K K
1 X F 1 X S
∇ΘL Ê[r] = ri DΘL (S, Fi ) ∇ΘO Ê[r] = ri DΘO (Si , F)
K K
i=1 i=1
8: Compute the stochastic gradient of ΘO : 8: Compute the stochastic gradient of ΘL :
K K
α1 X β1 X
∇ΘO Ê[r] = DΘO (Fi , S) ∇ΘL Ê[r] = DΘL (Fi , S)
K K
i=1 i=1
9: Model updates: 9: Model updates:
ΘL ← ΘL + ηL · ∇ΘL Ê[r], ΘO ← ΘO + ηO · ∇ΘO Ê[r],
ΘO ← ΘO + ηO · ∇ΘO Ê[r]. ΘL ← ΘL + ηL · ∇ΘL Ê[r].
until convergence
4.3 Algorithm 4.3.2 Learning from Facts to Sentence
For each pair < S, F >∈ DSAOKE , the following We sample S ∗ from the Logician PO (·|F, ΘO )
procedures are performed respectively (details are and define the total reward for S ∗ by:
shown in Algorithm 1): rO = β1 RO (S ∗ , F) + β2 VO (S ∗ ) +
4.3.1 Learning from Sentence to Facts β3 SO (S ∗ , S),
We sample F ∗ from the Logician PL (·|S, ΘL ) and where
P
βi = 1. The gradients can be computed
calculate the total reward for F ∗ by as follows:
rL = α1 · RL (F ∗ , S) + α2 · VL (F ∗ ) + ∇ΘL E[rO ] = E[β1 DΘL (F, S ∗ )],
α3 · SL (F ∗ , F), ∇ΘO E[rO ] = E[rO DΘO (S ∗ , F)].
In practice, we use beam search (Sutskever
P
where αi = 1. The gradients of the expected
reward E[rL ] to the parameters of agents can be et al., 2014) to obtain high-quality samples as F ∗
computed as follows, according to the policy gra- and S ∗ , and estimate the true gradient with the em-
dient theorem (Sutton et al., 1999): pirical average of gradients over these samples.
∇ΘL E[rL ] = E[rL DΘL (F ∗ , S)], 5 Experimental Results
∗
∇ΘO E[rL ] = E[α1 DΘO (S, F )]. 5.1 Experimental Design
where DΘL (F, S) = ∇ΘL log PL (F|S, ΘL ) and First, we evaluate the performance of each agent
DΘO (S, F) = ∇ΘO log PO (S|F, ΘO ). fine-tuned by the dual learning procedure on the
2124SAOKE dataset. Then we evaluate the Orator on strings are concatenated with commas to form the
noisy facts, which accords with real OIN applica- final sentence.
tion scenarios. Last, we investigate the behavior
5.1.3 Reward Implementation
of agents in the dual system.
In the experiments, the SAOKE dataset is split For the validity reward of the Orator, the lan-
into the training set, validating set and testing guage model is trained using an RNN based
set with ratios of 80%, 10%, 10%, respectively. method (Mikolov et al., 2010) with the same vo-
For each algorithm involved in the experiments, cabulary V and the web pages from Baidu Baike
we perform grid search to find the optimal super- website.
parameters, and the model with the best perfor- For the reconstruction reward of the Orator,
mance on the validating set is chosen as the learnt since the Logician needs the shallow tag and de-
model to be evaluated on the testing set. pendency information of S ∗ as inputs, the infor-
mation is extracted using the LTP tool-set (Che
5.1.1 Evaluation Metric et al., 2010) and then fed to the Logician.
For the Orator, BLEU-4 and ROUGE-L are used
5.1.4 Training
to measure how well the output matches the
ground truth sentence. When training the base model for each agent, the
For the Logician, based on the fact-equivalence batch size is set to 20. When training two agents
judgment proposed in (Sun et al., 2018), we com- in dual learning, the batch size is set to 12, and the
pute the Precision(P), Recall (R) and F1-score beam size is set to 3. Both agents are trained using
over the testing set of the SAOKE dataset as the the stochastic gradient descent (SGD) with RM-
evaluation metric. SPROP strategy (Hinton et al., 2012) and early-
stop strategy on the validating set. In dual learn-
5.1.2 Agent Implementation ing, the super-parameters, including αi , βi , is de-
For the Orator, we make a vocabulary V with termined by grid-search.
size 72,591 by collecting all web pages from
Baidu Baike website3 (a Chinese alternative to 5.2 Evaluation of Agents on the SAOKE
Wikipedia) and identifying the words occurred in dataset
more than 100 web pages. For the Orator, the di- First, we evaluate the performance of the agents
mension of embedding vectors is set to Ne = 256, optimized by the dual learning method. To iden-
and the dimension of hidden states is set to Nh = tify the contribution of the dual structure, we train
256. We use a three-layer bi-directional GRU with another pair of agents with α1 = 0 and β1 = 0
dimension 128 as the encoder. All dimensions of in Algorithm 1 to exclude the dual information.
hidden states in the decoder are set to 256. Without the dual information, these two agents are
For the Logician, we implement the model de- trained independently to each other with reinforce-
scribed in (Sun et al., 2018), including the shallow ment learning on their own supervised informa-
tag information and the gated dependency atten- tion. We name these two agents as R-Logician and
tion mechanism. R-Orator, where “R” means “Reinforced”. In the
Furthermore, to provide an intuitive compre- experimental results of this paper, the symbol at
hension of the OIN task, we implement a rule- the top mark means that the marked result is sig-
based method for OIN task. For each sequence of nificantly different (with p = 0.05) with the corre-
facts in the SAOKE dataset, the method first iden- sponding result of the agent with the specific mark.
tifies the subsequences in which the facts share the
same subject. Then it preserves the subject of the Methods Precision Recall F1
first fact in each subsequence and removes the sub- Logician∗ 0.449 0.400 0.423
jects of following facts (by replacing it with an R-Logician∓ 0.462∗ 0.432∗ 0.446∗
empty string). It is necessary since the SAOKE Logician@Dual 0.494∗∓ 0.426∗ 0.457∗∓
dataset requires the shared subject to be repeated
for completeness of the related facts. At last, each Table 3: Performance of the Logicians.
fact is formatted into a string by filling the objects
The experimental results for the Logician
into the placeholders of the predicate and these
agents are shown in Table 3, from which we can
3
http://baike.baidu.com observe a significant performance improvement
2125from Logician to R-Logician and also from R- 5.4 Evaluation of the Dual System
Logician to Logician@Dual. The experimental re- In this section, we investigate the behavior of
sults for the Orator agents are shown in Table 4. agents in the dual system. We first examine the
The neural based Orator agents significantly out- Orator Logician
procedure F −−−−→ S ∗ −−−−−→ F ∗∗ , that is, for
perform the rule-based agent. For both evalua-
each F in the testing set of the SAOKE dataset, let
tion metric, the R-Orator and Orator@Dual are
the Orator narrate it into a sentence S ∗ , and then
both significantly outperform the original Orator.
let the Logician to extract facts F ∗∗ from S ∗ . Then
The Orator@Dual significantly outperforms the
the quality of F ∗∗ is measured by comparing it
R-Orator on the BLEU-4 score, but is not signifi- Logician Orator
cantly different on the ROUGE-L score. with F. Then we examine S −−−−−→ F ∗ −−−−→
S ∗∗ , which is the reverse procedure. The compar-
Methods BLEU-4 ROUGE-L ison is made between the family of base agents
Rule? 0.257 0.434 and that of the dual-trained agents. The results are
Orator∗ 0.401? 0.556? shown in Table 6, and two instance of these two
R-Orator∓ 0.405?∗ 0.559?∗ experiments are shown in Table 7 and 8 respec-
Orator@Dual 0.419?∗∓ 0.559?∗ tively. From these results, we can observe large
improvements of reconstruction quality on both
Table 4: Performance of the Orators. directions.
Orator Logician
By comparing the performance of R-agents F −−−−→ S ∗ −−−−−→ F ∗∗
and the agents@Dual, we can observe that Methods Precision Recall F1
agents@Dual generally achieve better perfor- Base∗ 0.574 0.488 0.527
mance on precision, but may recall less informa- Dual 0.657∗ 0.565∗ 0.608∗
tion, resulting in smaller advances in the balanced Logician Orator
S −−−−−→ F ∗ −−−−→ S ∗∗
evaluation metric (F1 and ROUGE-L). This may
Methods BLEU-4 ROUGE-L
imply that the agents tend to provide easy input
Base∗ 0.428 0.565
for each other for higher accuracy, by neglecting ∗
Dual 0.635 0.635∗
some difficult part of the problem which they cur-
rently cannot handle properly. This interesting
Table 6: Reconstruction performance for the Logicians
phenomenon is the subject of our future research. and the Orators.
5.3 Evaluation of Orator on Noisy Facts 6 Conclusion
Experiments in the previous subsection show the
In this paper, we investigate the OIN task and its
performance of Orators to narrate a set of human-
duality to the OIE task. The proposed Orator has
labeled facts. In practice, however, the input to
shown its ability to fulfill the OIN task, that is, as-
the Orator might not be the human-labeled perfect
sembling open-domain facts into high quality sen-
facts, but some noisy facts automatically extracted
tences. Furthermore, our attempt to utilize the du-
by OIE algorithms. In this subsection, we make
ality between OIN and OIE tasks for improving
a collection of sets of noisy facts by feeding the
the performances for both OIN and OIE agents ac-
sentences in the testing set of the SAOKE dataset
complishes a preliminary success.
to the base Logician model and collecting the out-
Our work suggests at least three future research
puts. Then we evaluate the series of Orator models
topics: Firstly, one can enrich the theoretical study
on these noisy facts, and report their performance
of the duality between the OIE and OIN tasks.
at Table 5, from which we can see the performance
Secondly, one can investigate how to conquer the
improvement from the Orator to Orator@Dual.
barrier of the absence of an extensive collection of
Methods BLEU-4 ROUGE-L reasonable sets of open-domain facts and incorpo-
Orator∗ 0.428 0.565 rate unsupervised information into this Logician-
R-Orator∓ 0.431∗ 0.567∗ Orator dual learning structures for further im-
Orator@Dual 0.458∗∓ 0.572∗∓ provement. Lastly, one can also interested in de-
veloping task-oriented rewards for adapting the
Table 5: Performance of the Orators on noisy facts. agent to a specific task, for example, the answer
generation task for open-domain KBQA system.
2126S 大综货物吞吐量均保持两位数增长，其中铁矿石吞吐量突破4500万吨，木材吞吐量
突破600万立方，均创历史新高，成为全国进口木材第一大港。
S in English The throughput of all integrated cargoes kept a double-digit growth. Among them, the
throughput of iron ore exceeded 45 million tons and the throughput of timber exceeded 6
million cubic meters, all of which hit record highs, and became the country’s largest port of
timber imports.
Base Models Dual Models
Logician
S −−−−−−→ F ∗ (大综货物吞吐量, 保持, 两位数增长) (铁 (大综货物吞吐量, 均保持, 两位数增长) (铁
矿石吞吐量, 突破, 4500万吨) (木材吞吐量, 矿石吞吐量, 突破, 4500万吨) (木材吞吐量,
突破, 600万立方) (创历史, DESC, 新高) (_, 突破, 600万立方) (_, 均创, 历史新高) (_, 成
成为, 全国进口木材第一大港) 为, 全国进口木材第一大港)
Logician
S −−−−−−→ F ∗ (The throughput of all integrated cargoes, (The throughput of integrated cargoes, all
in English kept, double-digits growth) (Throughput kept, double-digit growth) (Throughput
of iron ore, exceeded, 45 million tons) of iron ore, exceeded, 45 million tons)
(Throughput of wood, breakthrough, 6 mil- (Throughput of timber, exceeded 6, million
lion cubic meters) (Hit historical, DESC, new cubic meters) (_, all hit, record highs) (_,
high) (_, become, the country’s largest port of become, the country’s largest port of timber
timber imports) imports)
Logician
S −−−−−−→ 大综货物吞吐量保持两位数增长，突 大综货物吞吐量均保持两位数增长，铁
∗ Orator
F −−−−→ S ∗∗ 破4500万吨，突破600万立方，新高，成 矿石吞吐量突破4500万吨，木材吞吐量突
为全国进口木材第一大港。 破600万立方，均创历史新高，成为全国
进口木材第一大港。
Logician
S −−−−−−→ The throughput of all integrated cargoes The throughput of all integrated cargoes kept
∗ Orator
F −−−−→ S ∗∗ kept double-digit growth, breaking 45 million double-digit growth. The throughput of iron
in English tons, breaking 6 million cubic meters, a new ore exceeded 45 million tons and the through-
high, became the country’s largest port of tim- put of timber exceeded 6 million cubic me-
ber imports. ters, all hit a record high and became the
country’s largest port of timber imports.
Table 7: Examples for sentence reconstruction.
F (营业法人, 具备, 中专（高中）以上学历) (营业法人, 有, 一定的[管理|经营]能力) (营
业法人, 具备, [较强的服务理念|团队合作精神])
F in English (Business legal person, is with, technical secondary school (high school) or above) (Busi-
ness legal person, has, certain [managerial | operational] ability) (Business legal person, is
with, [strong philosophy of service | teamwork spirit])
Base Models Dual Models
Orator
F −−−−→ S ∗ 营业法人具备中专（高中）以上学历，有 营业法人具备中专（高中）以上学历，营
一定的管理和经营能力，具备较强的服务 业法人有一定的管理和经营能力，具备较
理念和团队合作精神。 强的服务理念和团队合作精神。
Orator
F −−−−→ S ∗ in Business legal person is with secondary (high Business legal person is with secondary (high
English school) or above, has certain managerial and school) or above, business legal person has
operational capabilities, and is with strong certain managerial and operational capabili-
philosophy of service and teamwork spirit. ties, and is with strong philosophy of service
and teamwork spirit.
Orator
F −−−−→ (营业法人具备中专（高中）以上学历, 有, (营业法人, 具备, 中专（高中）以上学历)
∗ Logician
S −−−−−−→ F ∗∗ 一定的[管理|经营]能力) (营业法人具备中 (营业法人,有,[一定的管理|经营能力]) (营
专的[管理|经营]能力, 具备, 较强的[服务理 业法人, 具备, 较强的[服务理念|团队合作
念|团队合作精神]) 精神])
Orator
F −−−−→ (Business legal person with technical sec- (Business legal person, is with, technical sec-
∗ Logician
S −−−−−−→ F ∗∗ ondary school (high school) or above, has, ondary school (high school) or above) (Busi-
in English certain [managerial | operational] ability) ness legal person, has, [certain managerial |
(Business legal person with technical sec- operational ability]) (Business legal person,
ondary school (high school) or above, is is with, [strong philosophy of service | team-
with, [strong philosophy of service | team- work spirit])
work spirit])
Table 8: Examples for fact reconstruction.
2127References Janara Christensen, Stephen Soderland, and Gagan
Bansal. 2014. Hierarchical Summarization: Scaling
Shubham Agarwal and Marc Dymetman. 2017. A Up Multi-Document Summarization. In Proceed-
Surprisingly Effective Out-of-the-Box Char2char ings of the 52nd Annual Meeting of the Association
Model on the E2E NLG Challenge Dataset. In Pro- for Computational Linguistics, pages 902–912.
ceedings of the 18th Annual SIGdial Meeting on Dis-
course and Dialogue, August, pages 158–163. Nan Duan, Duyu Tang, Peng Chen, and Ming Zhou.
2017. Question Generation for Question Answer-
Sören Auer, Christian Bizer, Georgi Kobilarov, Jens
ing. In Proceedings of the 2017 Conference on Em-
Lehmann, Richard Cyganiak, and Zachary Ives.
pirical Methods in Natural Language Processing,
2007. DBpedia: A Nucleus for a Web of Open Data.
pages 866–874.
The Semantic Web, 4825 LNCS:722–735.
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- Oren Etzioni, Anthony Fader, Janara Christensen,
gio. 2014. Neural Machine Translation by Jointly Stephen Soderland, and Mausam. 2011. Open In-
Learning to Align and Translate. In International formation Extraction: The Second Generation. In
Conference on Learning Representations. Proceeding of International Joint Conference on Ar-
tificial Intelligence, pages 3–10.
Laura Banarescu, Claire Bonial, Shu Cai, Madalina
Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Anthony Fader, Luke S Zettlemoyer, and Oren Et-
Knight, Philipp Koehn, Martha Palmer, and Nathan zioni. 2014. Open Question Answering Over Cu-
Schneider. 2013. Abstract Meaning Representation rated and Extracted Knowledge Bases. In Proceed-
for Sembanking. Proceedings of the 7th Linguistic ings of the 20th ACM SIGKDD International Con-
Annotation Workshop and Interoperability with Dis- ference on Knowledge Discovery and Data Mining,
course, pages 178–186. pages 1156–1165.
Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Federico Gaspari. 2006. Look Who’s Translating: Im-
Sturge, and Jamie Taylor. 2008. Freebase: a Col- personations, Chinese Whispers and Fun with Ma-
laboratively Created Graph database for Structuring chine Translation on the Internet. In EAMT-2006:
Human Knowledge. In Proceedings of the 2008 11th Annual Conference of the European Associa-
ACM SIGMOD International Conference on Man- tion for Machine Translation, pages 19–20.
agement of Data, pages 1247–1250. ACM.
Jiatao Gu, Zhengdong Lu, Hang Li, and Victor O.K.
Wanxiang Che, Zhenghua Li, and Ting Liu. 2010. LTP: Li. 2016. Incorporating Copying Mechanism in
A Chinese Language Technology Platform. In Pro- Sequence-to-Sequence Learning. In Proceedings of
ceedings of the Coling, pages 13–16. the 54th Annual Meeting of the Association for Com-
putational Linguistics, pages 1631–1640.
Andrew Chisholm, Will Radford, and Ben Hachey.
2017. Learning to Generate One-sentence Biogra- Geoffrey Hinton, Nitish Srivastava, and Kevin Swer-
phies from Wikidata. In Proceedings of the 15th sky. 2012. Overview of Mini-batch Gradient De-
Conference of the European Chapter of the Associa- scent. Technical report.
tion for Computational Linguistics, volume 1, pages
633–642. Nanda Kambhatla. 2004. Combining Lexical, Syntac-
tic, and Semantic Features with Maximum Entropy
Kyunghyun Cho, Bart van Merrienboer, Caglar Gul- Models for Extracting Relations. In Proceedings of
cehre, Dzmitry Bahdanau, Fethi Bougares, Hol- the ACL 2004 on Interactive Poster and Demonstra-
ger Schwenk, and Yoshua Bengio. 2014. Learn- tion Sessions.
ing Phrase Representations using RNN Encoder-
Decoder for Statistical Machine Translation. In Pro- Tushar Khot, Ashish Sabharwal, and Peter Clark. 2017.
ceedings of the 2014 Conference on Empirical Meth- Answering Complex Questions Using Open Infor-
ods in Natural Language Processing, pages 1724– mation Extraction. In Proceedings of the 55th An-
1734. nual Meeting of the Association for Computational
Linguistics, pages 311—-316.
Janara Christensen, Mausam, Stephen Soderland, and
Oren Etzioni. 2011. An Analysis of Open Informa- Ioannis Konstas, Srinivasan Iyer, Mark Yatskar, Yejin
tion Extraction based on Semantic Role Labeling. Choi, and Luke Zettlemoyer. 2017. Neural amr:
In Proceedings of the Sixth International Conference Sequence-to-sequence models for parsing and gen-
on Knowledge Capture, pages 113–120. ACM Press. eration. In Proceedings of the 55th Annual Meet-
ing of the Association for Computational Linguis-
Janara Christensen, Mausam, Stephen Soderland, Oren tics, pages 146–157. Association for Computational
Etzioni, Mausam, Stephen Soderland, and Oren Et- Linguistics.
zioni. 2013. Towards Coherent Multi-Document
Summarization. In Proceedings of the 2013 Con- C Y Lin. 2004. Rouge: A Package for Automatic Eval-
ference of the North American Chapter of the Asso- uation of Summaries. In Proceedings of the 2004
ciation for Computational Linguistics: Human Lan- ACL Workshop on Text Summarization Branches
guage Technologies, Section 3, pages 1163–1173. Out, 1, pages 25–26.
2128Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, Mingming Sun, Xu Li, Xin Wang, Miao Fan, Yue Feng,
and Maosong Sun. 2016. Neural Relation Extrac- and Ping Li. 2018. Logician: A Unified End-to-End
tion with Selective Attention over Instances. In Pro- Neural Approach for Open-Domain Information Ex-
ceedings of the 54th Annual Meeting of the Asso- traction. In Proceedings of the Eleventh ACM In-
ciation for Computational Linguistics, pages 2124– ternational Conference on Web Search and Data
2133. Mining, February, pages 556–564, New York, New
York, USA. ACM Press.
Mausam. 2016. Open Information Extraction Systems
and Downstream Applications. In Proceedings of Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014.
the 25th International Joint Conference on Artificial Sequence to Sequence Learning with Neural Net-
Intelligence, pages 4074–4077. works. In Advances in Neural Information Process-
ing Systems 27, volume 155, page 9.
T Mikolov, M Karafiat, L Burget, J Cernocky, and
S Khudanpur. 2010. Recurrent Neural Network Richard S. Sutton, David Mcallester, Satinder Singh,
based Language Model. In Proceedings of Inter- and Yishay Mansour. 1999. Policy Gradient Meth-
speech, September, pages 1045–1048. ods for Reinforcement Learning with Function Ap-
proximation. In Advances in Neural Information
Mike Mintz, Steven Bills, Rion Snow, and Dan Juraf- Processing Systems 12, pages 1057–1063.
sky. 2009. Distant Supervision for Relation Extrac-
tion without Labeled Data. In Proceedings of the Duyu Tang, Nan Duan, Tao Qin, Zhao Yan, and
Joint Conference of the 47th Annual Meeting of the Ming Zhou. 2017. Question answering and ques-
ACL and the 4th International Joint Conference on tion generation as dual tasks. arXiv preprint
Natural Language, volume 2, page 1003, Morris- arXiv:1706.02027.
town, NJ, USA. Association for Computational Lin-
guistics. Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu,
and Hang Li. 2016. Modeling Coverage for Neural
Makoto Miwa and Mohit Bansal. 2016. End-to-End Machine Translation. In Proceedings of the Annual
Relation Extraction using LSTMs on Sequences and Meeting of the Association for Computational Lin-
Tree Structures. In Proceedings of the 54th Annual guistics, pages 76–85.
Meeting of the Association for Computational Lin-
guistics, pages 1105–1116, Stroudsburg, PA, USA. Menno Van Zaanen and Simon Zwarts. 2006. Unsu-
Association for Computational Linguistics. pervised Measurement of Translation Quality using
Multi-engine, Bi-directional Translation. In Aus-
Harinder Pal and Mausam. 2016. Demonyms and tralasian Joint Conference on Artificial Intelligence,
Compound Relational Nouns in Nominal Open IE. pages 1208–1214. Springer.
In Proceedings of the 5th Workshop on Automated
Knowledge Base Construction, pages 35–39. Pavlos Vougiouklis, Hady Elsahar, Lucie-Aimée
Kaffee, Christoph Gravier, Frederique Laforest,
John W Ratcliff and David E Metzener. 1988. Pattern Jonathon Hare, and Elena Simperl. 2017. Neural
Matching: The Gestalt Approach. Dr Dobb’s, 13(7). Wikipedian: Generating Textual Summaries from
Knowledge Base Triples. Journal of Web Seman-
Mrinmaya Sachan and Eric Xing. 2018. Self-Training tics: Science, Services and Agents on the World Wide
for Jointly Learning to Ask and Answer Questions. Web.
In Proceedings of the 2018 Conference of the North
American Chapter of the Association for Computa- Wikipedia. 2017. Assignment problem— Wikipedia,
tional Linguistics: Human Language Technologies, The Free Encyclopedia.
volume 1, pages 629–640.
Sam Wiseman, Stuart M. Shieber, and Alexander M.
Michael Schmitz, Robert Bart, Stephen Soderland, and Rush. 2017. Challenges in Data-to-Document Gen-
Oren Etzioni. 2012. Open language learning for in- eration. In Proceedings of the 2017 Conference on
formation extraction. In Proceedings of the 2012 Empirical Methods in Natural Language Process-
Joint Conference on Empirical Methods in Natural ing, pages 2243–2253.
Language Processing and Computational Natural
Language Learning, pages 523–534. Yingce Xia, Di He, Tao Qin, Liwei Wang, Nenghai Yu,
Tie-Yan Liu, and Wei-Ying Ma. 2016. Dual Learn-
Tomohiro Shigenobu. 2007. Evaluation and Usability ing for Machine Translation. In Advances in Neural
of Back Translation for Intercultural Communica- Information Processing Systems 29, pages 1–9.
tion. In International Conference on Usability and
Internationalization, pages 259–265. Springer. Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai
Yu, and Tie-Yan Liu. 2017. Dual Supervised Learn-
Gabriel Stanovsky, Ido Dagan, and Mausam. 2015. ing. In Proceedings of 34th International Confer-
Open IE as an Intermediate Structure for Semantic ence on Machine Learning.
Tasks. In Proceedings of the 53rd Annual Meeting of
the Association for Computational Linguistics and Jun Yin, Xin Jiang, Zhengdong Lu, Lifeng Shang,
the 7th International Joint Conference on Natural Hang Li, and Xiaoming Li. 2016. Neural Gen-
Language Processing, pages 303–308. erative Question Answering. In Proceedings of
21292016 NAACL Human-Computer Question Answer-
ing Workshop, pages 36–42.
Dmitry Zelenko, Chinatsu Aone, Anthony Richardella,
Jaz Kandola, Thomas Hofmann, Tomaso Poggio,
and John Shawe-Taylor. 2003. Kernel Methods for
Relation Extraction. Journal of Machine Learning
Research, 3:1083–1106.
Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing
Hao, Peng Zhou, and Bo Xu. 2017. Joint Extraction
of Entities and Relations Based on a Novel Tagging
Scheme. In Proceedings of the 55th Annual Meet-
ing of the Association for Computational Linguis-
tics, pages 1227–1236.
2130You can also read
