Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
KELM: K NOWLEDGE ENHANCED PRE-TRAINED LAN-
GUAGE REPRESENTATIONS WITH MESSAGE PASSING
ONHIERARCHICAL RELATIONAL GRAPHS
Yinquan Lu1,5Haonan Lu2yGuirong Fu3Qun Liu4
Huawei Technologies Co., Ltd.1OPPO Guangdong Mobile Telecommunications Co., Ltd.2
ByteDance3Huawei Noah’s Ark Lab4Shanghai AI Laboratory5
luyinquan@pjlab.org.cn, luhaonan@oppo.com
fuguirong@bytedance.com, qun.liu@huawei.com
ABSTRACT
Incorporating factual knowledge into pre-trained language models (PLM) such as
BERT is an emerging trend in recent NLP studies. However, most of the existing
methods combine the external knowledge integration module with a modiﬁed
pre-training loss and re-implement the pre-training process on the large-scale
corpus. Re-pretraining these models is usually resource-consuming, and difﬁcult
to adapt to another domain with a different knowledge graph (KG). Besides,
those works either cannot embed knowledge context dynamically according to
textual context or struggle with the knowledge ambiguity issue. In this paper, we
propose a novel knowledge-aware language model framework based on ﬁne-tuning
process, which equips PLM with a uniﬁed knowledge-enhanced text graph that
contains both text and multi-relational sub-graphs extracted from KG. We design
a hierarchical relational-graph-based message passing mechanism, which allows
the representations of injected KG and text to mutually update each other and can
dynamically select ambiguous mentioned entities that share the same text1. Our
empirical results show that our model can efﬁciently incorporate world knowledge
from KGs into existing language models such as BERT, and achieve signiﬁcant
improvement on the machine reading comprehension (MRC) tasks compared with
other knowledge-enhanced models.
1 I NTRODUCTION
Pre-trained language models beneﬁt from the large-scale corpus and can learn complex linguistic
representation (Devlin et al., 2019; Liu et al., 2019b; Yang et al., 2020). Although they have achieved
promising results in many NLP tasks, they neglect to incorporate structured knowledge for language
understanding. Limited by implicit knowledge representation, existing PLMs are still difﬁcult to
learn world knowledge efﬁciently (Poerner et al., 2019; Yu et al., 2020). For example, hundreds
of related training samples in the corpus are required to understand the fact “ banmeans an ofﬁcial
prohibition or edict against something” for PLMs.
By contrast, knowledge graphs (KGs) explicitly organize the above fact as a triplet “(ban, hypernyms,
prohibition)” . Although domain knowledge can be represented more efﬁciently in KG form, entities
with different meanings share the same text may happen in a KG (knowledge ambiguity issue). For
example, one can also ﬁnd “(ban, hypernyms, moldovan monetary unit)” in WordNet (Miller, 1995).
Recently, many efforts have been made on leveraging heterogeneous factual knowledge in KGs to
enhance PLM representations. These models generally adopt two methods: (1). Injecting pre-trained
entity embeddings into PLM explicitly, such as ERNIE (Zhang et al., 2019), which injects entity
embeddings pre-trained on a knowledge graph by using TransE (Bordes et al., 2013). (2). Implicitly
This work is done when Yinquan Lu, Haonan Lu and Guirong Fu work at Huawei Technologies Co., Ltd.
yCorresponding author
1Words or phrases in the text corresponding to certain entities in KGs are often named “entity mentions” .
While entities in KGs that correspond to entity mentions in the text are often named “mentioned entities”
1arXiv:2109.04223v2  [cs.CL]  5 May 2022Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
learning factual knowledge by adding extra pre-training tasks such as entity-level mask, entity-based
replacement prediction, etc. (Wang et al., 2020c; Sun et al., 2020). Some studies use both of the
above two methods such as CokeBERT (Su et al., 2020).
However, as summarized in Table 4 of Appendix, most of the existing knowledge-enhanced PLMs
need to re-pretrain the models based on an additional large-scale corpus, they mainly encounter
two problems below: (1) Incorporating external knowledge during pretraining is usually resource-
consuming and difﬁcult to adapt to other domains with different KGs. By checking the third column
of Table 4 in Appendix, one can see that most of the pretrain-based models use Wiki-related KG as
their injected knowledge source. These models also use English Wikipedia as pre-training corpus.
They either use an additional entity linking tool (e.g. TAGME (Ferragina & Scaiella, 2010)) to align
the entity mention in the text to a single mentioned entity in a Wiki-related KG uniquely or directly
treat hyperlinks in Wikipedia as entity annotations. These models depend heavily on the one-to-one
mapping relationship between Wikipedia corpus and Wiki-related KG, thus they never consider
handling knowledge ambiguity issue. (2) These models with explicit knowledge injection usually use
algorithms like BILINEAR (Yang et al., 2015) to obtain pre-trained KG embeddings, which contain
information about graph structure. Unfortunately, their knowledge context is usually static and cannot
be embedded dynamically according to textual context.
Several works (Qiu et al., 2019; Yang et al., 2019) concentrate on injecting external knowledge based
on ﬁne-tuning PLM on downstream tasks, which is much easier to change the injected KGs and adapt
to relevant domain tasks. They either cannot consider multi-hop relational information, or struggle
with knowledge ambiguity issue. How to fuse heterogeneous information dynamically based on the
ﬁne-tuning process on the downstream tasks and use the information of injected KGs more efﬁciently
remains a challenge.
Figure 1: Uniﬁed Knowledge-enhanced Text Graph (UKET)
consists of three parts corresponding to our model: (1) KG
only part, (2) Entity link to token graph, (3) Text only graph.To overcome the challenges mentioned
above, we propose a novel frame-
work named KELM , which injects
world knowledge from KGs during
the ﬁne-tuning phase by building a
Uniﬁed Knowledge-enhanced Text
Graph (UKET) that contains both in-
jected sub-graphs from external knowl-
edge and text. The method extends
the input sentence by extracting sub-
graphs centered on every mentioned
entity from KGs. In this way, we can
get a Uniﬁed Knowledge-enhanced
Text Graph as shown in Fig. 1, which is
made of three kinds of graph: (1) The
injected knowledge graphs, referred to
as the “ KG only ” part; (2) The graph
about entity mentions in the text and mentioned entities in KGs, referred to as the “ entity link to
token ” part. Entity mentions in the text are linked with mentioned entities in KGs by string matching,
so one entity mention may trigger several mentioned entities that share the same text in the injected
KGs (e.g. “Ford” in Fig. 1); (3) The “ text only ” part, where the input text sequence is treated as a
fully-connected word graph just like classical Transformer architecture (Vaswani et al., 2017).
Based on this uniﬁed graph, we design a novel Hierarchical relational-graph-based Message Passing
(HMP) mechanism to fuse heterogeneous information on the output layer of PLM. The implemen-
tation of HMP is via a Hierarchical Knowledge Enhancement Module as depicted in Fig. 2, which
also consists of three parts, and each part is designed for solving the different problems above: (1)
For reserving the structure information and dynamically embedding injected knowledge, we utilize
a relational GNN (e.g. rGCN (Schlichtkrull et al., 2017)) to aggregate and update representations
of extracted sub-graphs for each injected KG (corresponding to the “KG only” part of UKET). All
mentioned entities and their K-hop neighbors in sub-graphs are initialized by pre-trained vectors
obtained from the classical knowledge graph embedding (KGE) method (we adopt BILINEAR here).
In this way, knowledge context can be dynamically embedded, the structural information about the
graph is also kept; (2) For handling knowledge ambiguity issue and selecting relevant mentioned
entities according to the input context, we leverage a specially designed attention mechanism to
2Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
Figure 2: Framework of KELM (left) and illustrates how to generate knowledge-enriched token
embeddings (right).
weight these ambiguous mentioned entities by using the textual representations of words/tokens to
query the representations of their related mentioned entities in KGs (corresponding to the “entity
link to token” graph of UKET). The attention score can help to select knowledge according to the
input sentence dynamically. By concatenating the outputs of this step with the original outputs
of PLM, we can get a knowledge-enriched representation for each token; (3) For further interac-
tions between knowledge-enriched tokens, we employ a self-attention mechanism that operates on
the fully-connected word graph (corresponding to the “text only” graph of UKET) to allow the
knowledge-enriched representation of each token to further interact with others.
We conduct experiments on the MRC task, which requires a system to comprehend a given text
and answer questions about it. In this paper, to prove the generalization ability of our method,
we evaluate KELM on both the extractive-style MRC task (answers can be found in a span of the
given text) and the multiple-response-items-style MRC task (each question is associated with several
choices for answer-options, the number of correct answer-options is not pre-speciﬁed). MRC is a
challenging task and represents a valuable path towards natural language understanding (NLU). With
the rapid increment of knowledge, NLU becomes more difﬁcult since the system needs to absorb new
knowledge continuously. Pre-training models on large-scale corpus is inefﬁcient. Therefore, ﬁne-
tuning the knowledge-enhanced PLM on the downstream tasks directly is crucial in the application.2
2 R ELATED WORK
2.1 K NOWLEDGE GRAPH EMBEDDING
We denote a directed knowledge graph as G(E;R), whereEandRare sets of entities and relations,
respectively. We also deﬁne Fas a set of facts, a fact stored in a KG can be expressed as a triplet
(h;r;t )2F, which indicates a relation rpointing from the head entity hto tail entity t, where
h;t2E andr2R . KGE aims to extract topological information in KG and to learn a set of
low-dimensional representations of entities and relations by knowledge graph completion task (Yang
et al., 2015; Lu & Hu, 2020).
2.2 M ULTI -RELATIONAL GRAPH NEURAL NETWORK
Real-world KGs usually include several relations. However, traditional GNN models such as
GCN (Kipf & Welling, 2017), and GAT (Veli ˇckovi ´c et al., 2018) can only be used in the graph
2Code is available at here.
3Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
with one type of relation. (Schlichtkrull et al., 2017; Haonan et al., 2019) generalizes traditional
GNN models by performing relation-speciﬁc aggregation, making it possible to encode relational
graphs. The use of multi-relational GNN makes it possible to encode injected knowledge embeddings
dynamically in SKG and CokeBERT.
2.3 J OINT LANGUAGE AND KNOWLEDGE MODELS
Since BERT was published in 2018, many efforts have been made for further optimization, basically
focusing on the design of the pre-training process and the variation of the encoder. For studies of
knowledge-enhanced PLMs, they also fall into the above two categories or combine both of them
sometimes. Despite their success in leveraging external factual knowledge, the gains are limited by
computing resources, knowledge ambiguity issue, and the expressivity of their methods for the fusion
of heterogeneous information, as summarized in Table 4 of Appendix and the introduction part.
Recent studies notice that the architecture of Transformer treats input sequences as fully-connected
word graphs, thus some of them try to integrate injected KGs and textual context into a uniﬁed
data structure. Here we argue that UKET in our KELM is different from the WK graph proposed
in CoLAKE/K-BERT. These two studies heuristically convert textual context and entity-related
sub-graph into input sequences, both entities and relations are treated as input words of the PLM,
then they leverage a Transformer with a masked attention mechanism to encode those sequences
from the embedding layer and pre-train the model based on the large-scale corpus. Unfortunately, it
is not trivial for them to convert the second or higher order neighbors related to textual context (Su
et al., 2020), the structural information about the graph is lost. UKET differs from the WK graph
of CoLAKE/K-BERT in that, instead of converting mentioned entities, relations, and text into a
sequence of words and feeding them together into the input layer of PLM (they unify text and KG into
a sequence), UKET uniﬁes text and KG into a graph. Besides, by using our UKET framework, the
knowledge fusion process of KELM is based on the representation of the last hidden layer of PLM,
making it possible to directly ﬁne-tune the PLM on the downstream tasks without re-pretraining the
model. SKG also utilizes relational GNN to fuse information of KGs and text representation encoded
by PLM. However, SKG only uses GNN to dynamically encode the injected KGs, which corresponds
to part one of Fig. 1. Outputs of SKG are made by directly concatenating outputs of graph encoder
with the outputs of PLM. It cannot select ambiguous knowledge and forbids the interactions between
knowledge-enriched tokens corresponding to part two and part three of Fig. 1, respectively. KT-NET
uses a specially designed attention mechanism to select relevant knowledge from KGs. For example,
it treats all synsets of entity mentions within the WN183as candidate KB concepts. This limits the
ability of KT-NET to select the most relevant mentioned entities4. Moreover, the representations
of injected knowledge are static in KT-NET, they cannot dynamically change according to textual
context, the information about the original graph structure in KG is also lost.
3 M ETHODOLOGY
The architecture of KELM is shown in Fig. 2. It consists of three main modules: (1) PLM Encoding
Module; (2) Hierarchical Knowledge Enhancement Module; (3) Output Module.
3.1 PLM E NCODING MODULE
This module utilizes PLM (e.g.BERT) to encode text to get textual representations for pas-
sages and questions. An input example of the MRC task includes a paragraph and a ques-
tion with a candidate answer, represented as a single sequence of tokens of the length n:
T=f[CLS];Q;(A);[SEP ];P;[SEP ]g=ftign
i=1, whereQ,AandPrepresent all tokens for ques-
tion, candidate answer and paragraph, respectively5.[SEP ]and[CLS]are special tokens in BERT
and deﬁned as a sentence separator and a classiﬁcation token, respectively. i-th token in the sequence
is represented by ~ht
i2Rdt, wheredtis the last hidden layer size of used PLM.
3A subset of WordNet.
4Refer the example given in the case study of KT-NET, the most relevant concept for the word “ban” is
“forbidding_NN_1” (with the probability of 86.1%), but not “ban_NN_4”.
5Depending on the type of MRC task (extractive-style v.s. multiple-response-items-style), candidate answer
A is not required in the sequence of tokens for the extractive-style MRC task.
4Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
3.2 H IERARCHICAL KNOWLEDGE ENHANCEMENT MODULE
This module is the implementation of our proposed HMP mechanism to fuse information of textual
and graph context. We will formally introduce graph construction for UKET, and the three sub-
processes of HMP in detail in the following sections.
3.2.1 C ONSTRUCTION OF UKET G RAPH
(1) Given a set with jQjelements:fGq
k(Eq
k;Rq
k)gjQj
q=1and input text, where jQjis the total number of
injected KGs, and qindicates the q-th KG. We denote the set of entity mentions related to the q-th
KG asXq=fxq
igjXqj
i=1, wherejXqjis the number of entity mentions in the text. The corresponding
mentioned entities are shared by all tokens in the same entity mention. All mentioned entities
Mq=fmq
igjMqj
i=1are linked with their relevant entity mentions in the text, where jMqjis the number
of mentioned entities in the q-th KG. We deﬁne this "entity link to token graph" in Fig. 1 as
Gq
m(Eq
m;Rq
m), whereEq
m=Xq[Mqis the union of entity mentions and their relevant mentioned
entities,Rq
mis a set with only one element that links mentioned entities and their relevant entity
mentions. (2) For i-th mentioned entity mq
iinMq, we retrieve all its K-hop neighbors fNx
mq
igK
x=0
from theq-th knowledge graph, where Nx
mq
iis a set ofi-th mentioned entity’s x-hop neighbors, hence
we haveN0
mq
i=fmq
ig. We deﬁne "KG-only graph": Gq
s(Eq
s;Rq
s), whereEq
s=SjMqj
i=0SK
x=0Nx
mq
iis the
union of all mentioned entities and their neighbors within the K-hops sub-graph, and Rq
sis a set
of all relations in the extracted sub-graph of q-th KG. (3) The text sequence can be considered as
a fully-connected word graph as pointed out previously. This “text-only graph” can be denoted as
Gt(Et;Rt), whereEtis all tokens in text and Rtis a set with only one element that connects all
tokens. Finally, we deﬁne the full hierarchical graph consisting of all three parts fGq
sgjQj
q=1,fGq
mgjQj
q=1,
andGt, as Uniﬁed Knowledge-enhanced Text Graph (UKET).
3.2.2 D YNAMICALLY EMBEDDING KNOWLEDGE CONTEXT
We use pre-trained vectors obtained from the KGE method to initialize representations of entities in
Gq
s(Eq
s;Rq
s). Considering the structural information of injected knowledge graph forgotten during
training, we utilize jQjindependent GNN encoders (i.e. g1(:),g2(:)in Fig. 2, which is the case of
injecting two independent KGs in our experiment setting) to dynamically update entity embeddings
ofjQjinjected KGs. We use rGCN to model the multi-relational nature of the knowledge graph. To
updatei-th node ofq-th KG inl-th rGCN layer:
~ sq(l+1)
i =(X
r2Rq
sX
j2Nr
i1
jNr
ijWq(l)
r~ sq(l)
j)(1)
WhereNr
iis a set of neighbors of i-th node under relation r2Rq
s.Wq(l)
ris trainable weight matrix
atl-th layer and ~ sq(l+1)
i is the hidden state of i-th node at ( l+1)-th layer. After Lupdates,jQjsets
of node embeddings are obtained. The output of the q-th KG can be represented as Sq2RjEq
sjdq,
wherejEq
sjanddqare the numbers of nodes of extracted sub-graph and the dimension of pre-trained
KGE, respectively.
3.2.3 D YNAMICALLY SELECTING SEMANTICS -RELATED MENTIONED ENTITIES
To handle the knowledge ambiguity issue, we introduce an attention layer to weight these ambiguous
mentioned entities by using the textual representations of tokens (outputs of Section 3.1) to query
their semantics-related mentioned entities representations in KGs. Here, we follow the attention
mechanism of GAT to update each entity mention embedding in Gq
m:
~ xq
i=(X
j2Nq
iq
ijWq~ sq
j)(2)
Where~ sq
jis the output embeddings from the q-th rGCN in the previous step. ~ xq
iis the hidden state
ofi-th entity mention xq
iinXq, andNq
iis a set of neighbors of xq
iinGq
m.Wq2Rdoutdinis a
5Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
trainable weight matrix, we set din=dout=dq(thus~ xq
i2Rdq).is a nonlinear activation function.
q
ijis the attention score that weights ambiguous mentioned entities in the q-th KG:
q
ij=exp(LeakyReLU (~ T
q[Wq~ht0
ijjWq~ sq
j]))
P
k2Nq
iexp(LeakyReLU (~ Tq[Wq~ht0
ijjWq~ sq
k]))(3)
The representation ~ht
iwith a dimension of dtis projected to the dimension of dq, before using it
to query the related mentioned entity embeddings of Sq:~ht0
i=Wq
proj~ht
i, whereWq
proj2Rdqdt.
~ q2R2dqis a trainable weight vector. Tis the transposition operation and jjis the concatenation
operation.
Finally, we concatenate outputs of jQjKGs with textual context representation to get ﬁnal knowledge-
enriched representation:
~hk
i= [~ht
i;~ x1
i;:::;~ xjQj
i]2Rdt+d1++djQj (4)
If tokentican’t match any entity in q-th KG (say ti=2Xq), we ﬁll~ xq
iin Eq.4 with zeros. Note
that mentioned entities in KGs are not always useful, to prevent noise, we follow (Yang & Mitchell,
2017)’s work and add an extra sentinel node linked to each entity mention in Gq
m. The sentinel node
is initialized by zeros and not trainable, which is the same as the case of no retrieved entities in the
KG. In this way, according to the textual context, KELM can dynamically select mentioned entities
and avoid introducing knowledge noise.
3.2.4 I NTERACTION BETWEEN KNOWLEDGE -ENRICHED TOKEN EMBEDDINGS
To allow knowledge-enriched tokens’ representations to propagate to each other in the text, we
use a fully-connected word graph Gt, with knowledge-enriched representations from outputs of
the previous step, and employ the self-attention mechanism similar to KT-NET to update token’s
embedding. The ﬁnal representation for i-th token in the text is ~hf
i2R6(dt+d1++djQj).
3.3 O UTPUT MODULE
3.3.1 E XTRACTIVE -STYLE MRC TASK
A simple linear transformation layer and softmax operation are used to predict start and end positions
of answers. For i-th token, the probabilities to be the start and end position of answer span are:
ps
i=exp(wT
s~hf
i)
nP
j=1exp(wTs~hf
j);pe
i=exp(wT
e~hf
i)
nP
j=1exp(wTe~hf
j), wherews;we2R6(dt+d1++djQj)are trainable vectors
andnis the number of tokens. The training loss is calculated by the log-likelihood of the true start
and end positions: L= 1
NNP
i=1(logps
ys
i+logpe
ye
i), whereNis the total number of examples in the
dataset,ys
iandye
iare the true start and end positions of i-th query’s answer, respectively. During
inference, we pick the span (a;b)with maximum ps
ape
bwhereabas predicted anwser.
3.3.2 M ULTIPLE -RESPONSE -ITEMS -STYLE MRC TASK
Since answers to a given question are independent of each other, to predict the correct probability
of each answer, a fully connected layer followed by a sigmoid function is applied on the ﬁnal
representation of [CLS]token in BERT.
4 E XPERIMENTS
4.1 D ATASETS
In this paper, we empirically evaluate KELM on both two types of MRC benchmarks in Super-
GLUE (Wang et al., 2020a): ReCoRD (Zhang et al., 2018) (extractive-style) and MultiRC (Khashabi
6Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
et al., 2018) (multiple-response-items-style). Detailed descriptions of the two datasets can be found
in Appendix B. On both datasets, the test set is not public, one has to submit the predicted results to
the organization to get the ﬁnal test score. Since frequent submissions to probe the unseen test set are
not encouraged, we only submit our best model once for each of the datasets, thus the statistics of the
results (e.g., mean, variance, etc.) are not applicable. We use Exact Match (EM) and (macro-averaged)
F1 as the evaluation metrics.
External Knowledge We adopt knowledge sources the same as used in KT-NET: WordNet and
NELL (Carlson et al., 2010). Representations of injected knowledge are initialized by resources
provided by (Yang & Mitchell, 2017). The size of these embeddings is 100. We retrieve related
knowledge from the two KGs in a given sentence and construct UKET graph (as shown in Section
3.2.1). More details about entity embedding and concepts retrieval are available in Appendix B.
4.2 E XPERIMENTAL SETUPS
Baselines and Comparison Setting Because we use BERT largeas the base model in our method,
we use it as our primary baseline for all tasks. For fair comparison, we mainly compare our results
with two ﬁne-tune-based knowledge-enhanced models: KT-NET and SKG, which also evaluate
their results on ReCoRD with BERT largeas the encoder part. As mentioned in the original paper
of KT-NET, KT-NET mainly focuses on the extractive-style MRC task. We also evaluate KT-NET
on the multiple-response-items-style MRC task and compare the results with KELM. We evaluate
our approach in three different KB settings: KELM WordNet ,KELM NELL, and KELM Both, to inject KG
from WordNet, NELL, and both of the two, respectively (The same as KT-NET). Implementation
details of our model are presented in Appendix C.
Dev Test
Model EM F1 EM F1
BERT large 70.2 72.2 71.3 72.0
SKG+BERT large 70.9 71.6 72.2 72.8
KT-NET WordNet 70.6 72.8 - -
KT-NET NELL 70.5 72.5 - -
KT-NET BOTH 71.6 73.6 73.0 74.8
KELM WordNet 75.4 75.9 75.9 76.5
KELM NELL 74.8 75.3 75.9 76.3
KELM Both 75.1 75.6 76.2 76.7
Table 1: Result on ReCoRD.Dev Test
Model EM F1 EM F1
BERT large - - 24.1 70.0
KT-NET
BOTH 26.7 71.7 25.4 71.1
KELM WordNet 29.2 70.6 25.9 69.2
KELM NELL 27.3 70.4 26.5 70.6
KELM Both 30.3 71.0 27.2 70.8
Table 2: Result on MultiRC. [*] are from our
implementation.
4.3 R ESULTS
The results for the extractive-style MRC task and multiple-response-items-style MRC task are given in
Table 1 and Table 2, respectively. The scores of other models are taken directly from the leaderboard
of SuperGLUE6and literature (Qiu et al., 2019; Yang et al., 2019). In this paper, our implementation
is based on a single model, and hence comparing with ensemble based models is not considered. Best
results are labeled in bold and the second best are underlined.
Results on the dev set of ReCoRD show that: (1) KELM outperforms BERT large, irrespective of
which external KG is used. Our best KELM offers a 5.2/3.7 improvement in EM/F1 over BERT large.
(2) KELM outperforms previous SOTA knowledge-enhanced PLM (KT-NET) by +3.8 EM/+2.3 F1 .
In addition, KELM outperforms KT-NET signiﬁcantly in all three KB settings. On the dev set of
MultiRC, the best KELM offers a 3.6improvement in EM over KT-NET. Although the performance
on F1 drop a little compared with KT-NET, we still get a gain of +2.9 (EM+F1) over the former
SOTA model7.
Results on the test set further demonstrate the effectiveness of KELM and its superiority over the
previous works. On ReCoRD, it signiﬁcantly outperforms the former SOTA knowledge-enhanced
PLM (ﬁnetuning based model) by +3.2 EM/+1.9 F1 . And on MultiRC, KELM offers a 3.1/0.8
improvement in EM/F1 over BERT large, and achieves a gain of +1.5 (EM+F1) over KT-NET.
6https://super.gluebenchmark.com/leaderboard (Nov.14th, 2021)
7The best model is chosen according to the EM+F1 score (same as KT-NET).
7Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
5 C ASE STUDY
This section uses an example in ReCoRD to show how KELM avoids knowledge ambiguity issue
and selects the most relevant mentioned entities adaptively w.r.t the textual context. Recall that given
a tokenti, the importance of a mentioned entity mq
jinq-th KG is scored by the attention weight
q
ijin Eq.2. To illustrate how KELM can select the most relevant mentioned entities, we analyze
the example that was also used in the case study part of KT-NET. The question of this example is
“Sudan remains a XXX-designated state sponsor of terror and is one of six countries subject to the
Trump administration’s ban”, where the “XXX” is the answer that needs to be predicted. The case
study in KT-NET shows the top 3 most relevant concepts from WordNet for the word “ban” are
“forbidding.n.01”, “proscription.n.01”, and “ban.v.02”, with the weights of 0.861, 0.135, and 0.002,
respectively. KT-NET treats all synsets of a word as candidate KG concepts, both “forbidding.n.01”
and “ban.v.02” will be the related concepts of the word “ban” in the text. Although KT-NET can
select relevant concepts and suppress the knowledge noise through its specially designed attention
mechanism, we still observe two problems from the previous case study: (1) KT-NET cannot select
the most relevant mentioned entities in KG that share the same string in the input text. (2) Lack
of ability to judge the part of speech (POS) of the word (e.g. “ban.v.02” gets larger weights than
“ban.n.04”).
Word in text
(prototype)The most relevant
mentioned entity in
WordNet (predicted)Golden mentioned entity
ford ford.n.05 (0.56) ford.n.05
pardon pardon.v.02 (0.86) pardon.v.02
nixon nixon.n.01 (0.74) nixon.n.01
lead lead.v.03 (0.73) lead.v.03
outrage outrage.n.02 (0.62) outrage.n.02
Table 3: Case study. Comparisons between the golden label
with the most relevant mentioned entity in WordNet. The
importance of selected mentioned entities is provided in the
parenthesis.For KELM, by contrast, we focus on
selecting the most relevant mentioned
entities to solve the knowledge ambi-
guity issue (based on the “entity link
to token graph” part of UKET). For in-
jecting WordNet, by allowing message
passing on the extracted sub-graphs
(“KG only” part of UKET), knowl-
edge context can be dynamically em-
bedded according to the textual con-
text. Thus the neighbors’ information
of mentioned entities in WordNet can
be used to help the word in a text to
correspond to a particular POS based on its context. The top 3 most relevant mentioned entities in
WordNet for the word “ban” in the above example are “ban.n.04”, “ban.v.02”, and “ban.v.01”, with
the weights of 0.715, 0.205, and 0.060, respectively.
To vividly show the effectiveness of KELM, we analyze ambiguous words in the motivating example
show in Fig. 1 (The example comes from ReCoRD):
“President Ford then pardoned Richard Nixon, leading to a further ﬁrestorm of outrage. ”
Table. 3 presents 5 words in the above passage. For each word, the most relevant mentioned entity in
WordNet with the highest score is given. The golden mentioned entity for each word is labeled by
us. Deﬁnitions of mentioned entities in WordNet that correspond to the word examples are listing in
Table 5 of Appendix.
6 C ONCLUSION
In this paper, we have proposed KELM for MRC, which enhances PLM representations with
structured knowledge from KGs based on the ﬁne-tuning process. Via a uniﬁed knowledge-enhanced
text graph, KELM can embed the injected knowledge dynamically, and select relevant mentioned
entities in the input KGs. In the empirical analysis, KELM shows the effectiveness of fusing external
knowledge into representations of PLM and demonstrates the ability to avoid knowledge ambiguity
issue. Injecting emerging factual knowledge into PLM during ﬁnetuning without re-pretraining
the whole model is quite important in the application of PLMs and is still barely investigated.
Improvements achieved by KELM over vanilla baselines indicate a potential direction for future
research.
8Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
ACKNOWLEDGEMENTS
The authors thank Ms. X. Lin for insightful comments on the manuscript. We also thank Dr. Y . Guo
for helpful suggestions in parallel training settings. We also thank all the colleagues in AI Application
Research Center (AARC) of Huawei Technologies for their supports.
REFERENCES
Antoine Bordes, Nicolas Usunier, Alberto García-Durán, J. Weston, and Oksana Yakhnenko. Trans-
lating embeddings for modeling multi-relational data. In NIPS , 2013.
A. Carlson, J. Betteridge, B. Kisiel, B. Settles, E.R. Hruschka Jr., and T.M. Mitchell. Toward an
architecture for never-ending language learning. In Proceedings of the Conference on Artiﬁcial
Intelligence (AAAI) , pp. 1306–1313. AAAI Press, 2010.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep
bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of
the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, Volume 1 (Long and Short Papers) , pp. 4171–4186, Minneapolis, Minnesota, June
2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https:
//www.aclweb.org/anthology/N19-1423 .
P. Ferragina and Ugo Scaiella. Tagme: on-the-ﬂy annotation of short text fragments (by wikipedia
entities). Proceedings of the 19th ACM international conference on Information and knowledge
management , 2010.
Lu Haonan, Seth H. Huang, Tian Ye, and Guo Xiuyan. Graph star net for generalized multi-task
learning, 2019.
Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. Looking
beyond the surface:a challenge set for reading comprehension over multiple sentences. In Pro-
ceedings of North American Chapter of the Association for Computational Linguistics (NAACL) ,
2018.
Thomas N. Kipf and Max Welling. Semi-supervised classiﬁcation with graph convolutional networks,
2017.
Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. K-bert:
Enabling language representation with knowledge graph, 2019a.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike
Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining
approach, 2019b.
Edward Loper and Steven Bird. Nltk: The natural language toolkit. In Proceedings of the ACL-02
Workshop on Effective Tools and Methodologies for Teaching Natural Language Processing and
Computational Linguistics - Volume 1 , ETMTNLP ’02, pp. 63–70, USA, 2002. Association for
Computational Linguistics. doi: 10.3115/1118108.1118117. URL https://doi.org/10.
3115/1118108.1118117 .
Haonan Lu and Hailin Hu. Dense: An enhanced non-abelian group representation for knowledge
graph embedding, 2020.
George A. Miller. Wordnet: A lexical database for english. COMMUNICATIONS OF THE ACM , 38:
39–41, 1995.
Matthew E. Peters, Mark Neumann, Robert L. Logan IV au2, Roy Schwartz, Vidur Joshi, Sameer
Singh, and Noah A. Smith. Knowledge enhanced contextual word representations, 2019.
Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual
knowledge vs. name-based reasoning in unsupervised qa. ArXiv , abs/1911.03681, 2019.
9Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
Delai Qiu, Yuanzhe Zhang, Xinwei Feng, Xiangwen Liao, Wenbin Jiang, Yajuan Lyu, Kang Liu, and
Jun Zhao. Machine reading comprehension using structural knowledge graph-aware network. In
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and
the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pp.
5896–5901, Hong Kong, China, November 2019. Association for Computational Linguistics. doi:
10.18653/v1/D19-1602. URL https://www.aclweb.org/anthology/D19-1602 .
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi
Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a uniﬁed text-to-text
transformer, 2020.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. Squad: 100,000+ questions for
machine comprehension of text, 2016.
Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible alternatives:
An evaluation of commonsense causal reasoning. In Logical Formalizations of Commonsense
Reasoning, Papers from the 2011 AAAI Spring Symposium, Technical Report SS-11-06, Stanford,
California, USA, March 21-23, 2011 . AAAI, 2011. URL http://www.aaai.org/ocs/
index.php/SSS/SSS11/paper/view/2418 .
Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max
Welling. Modeling relational data with graph convolutional networks, 2017.
Yusheng Su, Xu Han, Zhengyan Zhang, Peng Li, Zhiyuan Liu, Yankai Lin, Jie Zhou, and Maosong
Sun. Cokebert: Contextual knowledge selection and embedding towards enhanced pre-trained
language models, 2020.
Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang.
Colake: Contextualized language and knowledge embedding, 2020.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
Kaiser, and Illia Polosukhin. Attention is all you need, 2017.
Petar Veli ˇckovi ´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua
Bengio. Graph attention networks, 2018.
Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer
Levy, and Samuel R. Bowman. Superglue: A stickier benchmark for general-purpose language
understanding systems, 2020a.
Minjie Wang, Da Zheng, Zihao Ye, Quan Gan, Mufei Li, Xiang Song, Jinjing Zhou, Chao Ma,
Lingfan Yu, Yu Gai, Tianjun Xiao, Tong He, George Karypis, Jinyang Li, and Zheng Zhang. Deep
graph library: A graph-centric, highly-performant package for graph neural networks, 2020b.
Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Jianshu ji, Guihong Cao, Daxin
Jiang, and Ming Zhou. K-adapter: Infusing knowledge into pre-trained models with adapters,
2020c.
Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian
Tang. Kepler: A uniﬁed model for knowledge embedding and pre-trained language representation,
2020d.
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,
Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von
Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama
Drame, Quentin Lhoest, and Alexander M. Rush. Huggingface’s transformers: State-of-the-art
natural language processing, 2020.
Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. Pretrained encyclopedia:
Weakly supervised knowledge-pretrained language model, 2019.
Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto. Luke: Deep
contextualized entity representations with entity-aware self-attention, 2020.
10Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
An Yang, Quan Wang, Jing Liu, Kai Liu, Yajuan Lyu, Hua Wu, Qiaoqiao She, and Sujian Li.
Enhancing pre-trained language representations with rich knowledge for machine reading com-
prehension. In Proceedings of the 57th Annual Meeting of the Association for Computational
Linguistics , pp. 2346–2357, Florence, Italy, July 2019. Association for Computational Linguistics.
doi: 10.18653/v1/P19-1226. URL https://www.aclweb.org/anthology/P19-1226 .
Bishan Yang and Tom Mitchell. Leveraging knowledge bases in LSTMs for improving machine
reading. In Proceedings of the 55th Annual Meeting of the Association for Computational Lin-
guistics (Volume 1: Long Papers) , pp. 1436–1446, Vancouver, Canada, July 2017. Association for
Computational Linguistics. doi: 10.18653/v1/P17-1132. URL https://www.aclweb.org/
anthology/P17-1132 .
Bishan Yang, Wen tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and
relations for learning and inference in knowledge bases, 2015.
Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V . Le.
Xlnet: Generalized autoregressive pretraining for language understanding, 2020.
Donghan Yu, Chenguang Zhu, Yiming Yang, and Michael Zeng. Jaket: Joint pre-training of
knowledge graph and language understanding, 2020.
Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme.
Record: Bridging the gap between human and machine commonsense reading comprehension,
2018.
Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. Ernie: Enhanced
language representation with informative entities, 2019.
ASUMMARY AND COMPARISON OF RECENT KNOWLEDGE -ENHANCED PLM S
Table 4 shows a brief summary and comparison of recent knowledge-enhanced PLMs. Most of recent
work concentrated on injecting external knowledge graphs during pre-training phase, which makes
them inefﬁcient in injecting external knowledge (e.g. LUKE takes about 1000 V100 GPU days to
re-pretraining the RoBERTa based PLM model). Also, nearly all of them uses an additional entity
linking tool to align the mentioned entities in the Wikidata to the entity mentions in the pre-trained
corpus (English Wikipedia) uniquely. These methods never consider to resolve the knowledge
ambiguity problem.
B D ATATSET AND KNOWLEDGE GRAPH DETAILS
ReCoRD (an acronym for the Reading Comprehension with Commonsense Reasoning Dataset) is a
large-scale dataset for extractive-style MRC requiring commonsense reasoning. There are 100,730,
10,000, and 10,000 examples in the training, development (dev), and test set, respectively. An example
of the ReCoRD consists of three parts: passage, question, and answer. The passage is formed by
the ﬁrst few paragraphs of an article from CNN or Daily Mail, with named entities recognized and
marked. The question is a sentence from the rest of the article, with a missing entity speciﬁed as the
golden answer. The model needs to ﬁnd the golden answer among the entities marked in the passage.
Questions that can be easily answered by pattern matching are ﬁltered out. By the design of the
process of data collection, one can see that to answer the questions, external background knowledge
and ability of reasoning are required.
MultiRC (Multi-Sentence Reading Comprehension) is a multiple-response-items-style MRC dataset
of short paragraphs and multi-sentence questions that can be answered from the content of the
paragraph. Each example of MultiRC includes a question that associates with several choices for
answer-options, and the number of correct answer-options is not pre-speciﬁed. The correct answer is
not required to be a span in the text. The dataset consists of 10K questions ( 6k multiple-sentence
questions), about 60% of this data make training/dev data. Paragraphs in the dataset have diverse
provenance by being extracted from 7 different domains such as news, ﬁction, historical text etc., and
hence are expected to be more complicated in their contents as compared to single-domain datasets.
11Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
Model Downstream Task Used KGsNeed
Pre-trainDynamically
Embedding KG
ContextInject external
KG’s RepresentationsSupport
Multi-relationalSupport
Multi-hopHandle Knowledge
Ambiguity IssueBase Model
ERNIE (Zhang et al., 2019)Glue, Ent Typing
Rel CLSWikidataYes
(MLM, NSP,
Ent Mask task)NoInject pretrained
entity embeddings
(TransE) explicitlyNo
(only entity embedding)NoNo
(anchored entity mention to
the unique id of Wikidata)BERT base
K-BERT (Liu et al., 2019a)Q&A, NER
Sent CLSCN-DBpedia
HowNet, MedicalKGOptional
(MLM, NSP)No NoYes
(treat relations as words)NoNo
(designed ATT mechanism
can solve KN issue)BERT base
KnowBERT (Peters et al., 2019)Rel Extraction
Ent TypingCrossWikis,
WordNetYes
(MLM, NSP,
Ent Linking task)NoInject both pretrained
entity embeddings (TuckER)
and entity deﬁnition explicitlyNo
(only entity embedding)NoYes
(weighed entity embeddings
shared the same text)BERT base
WKLM (Xiong et al., 2019) Q&A, Ent Typing WikidataYes
(MLM, Ent
replacement task)No No No NoNo
(anchored entity mention to
the unique id of Wikidata)BERT base
K-Adapter (Wang et al., 2020c)Q&A,
Ent TypingWikidata
Dependency ParsingYes
(MLM,
Rel predition task)No NoYes
(Via Rel prediction task
during pretraining)NoNo
(anchored entity mention to
the unique id of Wikidata)RoBERTa large
KEPLER (Wang et al., 2020d)Ent Typing
Glue, Rel CLS
Link PredictionWikidataYes
(MLM,
Link predition task)YesInject embeddings of
entity and relation’s
description explicitlyYes
(Via link prediction task
during pretraining)NoNo
(anchored entity mention to
the unique id of Wikidata)RoBERTa base
JAKET (Yu et al., 2020)Rel CLS, KGQA
Ent CLSWikidataYes
(MLM, Ent Mask task,
Ent category prediction,
Rel type prediction)YesInject embeddings of
entity descriptionsYes
(Via Rel type prediction
during pretraining)YesNo
(anchored entity mention to
the unique id of Wikidata)RoBERTa base
CoLAKE (Sun et al., 2020)Glue, Ent Typing
Rel ExtractionWikidataYes
(MLM, Ent Mask task,
Rel type prediction)Yes NoYes
(treat relations as words)NoNo
(anchored entity mention to
the unique id of Wikidata)RoBERTa base
LUKE (Yamada et al., 2020)Ent Typing, Rel CLS
NER, Q&AEnt from
WikipediaYes
(MLM, Ent Mask task)No No No NoNo
(treat hyperlinks in Wikipedia
as entity annotations)RoBERTa large
CokeBERT (Su et al., 2020)Rel CLS
Ent TypingWikidataYes
(MLM, NSP,
Ent Mask task)YesInject pretrained
entity embeddings
(TransE) explicitlyYes
(Via S-GNN to encode KG
context dynamically)YesNo
(anchored entity mention to
the unique id of Wikidata)RoBERTa large
SKG (Qiu et al., 2019) MRC WordNet, ConceptNet No YesInject pretrained
entity embeddings
(BILINER) explicitlyYes
(Via multi-relational
GNN to encode KG
context dynamically)Yes No BERT large
KT-NET (Yang et al., 2019) MRC WordNet, NELL No NoInject pretrained
entity embeddings
(BILINER) explicitlyNo
(only entity embedding)NoYes
(dynamically selecting
KG context)BERT large
KELM MRC WordNet, NELL No YesInject pretrained
entity embeddings
(BILINER) explicitlyYes
(Via multi-relational
GNN to encode KG
context dynamically)YesYes
(dynamically selecting
related mentioned entity)BERT large
Table 4: A brief summary and comparison of recent knowledge-enhanced PLMs. The full names
of some abbreviations are as follows. MLM : masked language model, NSP: next sentence pre-
diction, Ent: entity, Rel: relation, CLS : classiﬁcation, Sent: sentence, ATT : attention. Com-
ments/descriptions of features are written in parentheses. Desired properties are written in bold .
WordNet contains 151,442 triplets with 40,943 synsets459and 18 relations. We look up mentioned
entities in the WordNet by string matching operation, and link all tokens in the same word to the
retrieved mentioned entities (tokens are tokenized by Tokenizer of BERT). Then, we extract all 1-hop
neighbors for each mentioned entity and construct sub-graphs. In this paper, our experiment results
are based on the 1-hop case. However, our framework can be generalized to multi-hop easily, and we
leave this for future work.
NELL contains 180,107 entities and 258 concepts. We link entity mentions to the whole KG, and
return associated concepts.
C I MPLEMENTATION DETAILS
Our implementation is based on HuggingFace (Wolf et al., 2020) and DGL (Wang et al., 2020b).
For all three settings of KELM, parameters of the encoding layer of BERT largeare initialized with
pre-trained model released by Google. Other trainable parameters in HMP are randomly initialized.
The total number of trainable parameters of KELM is 340.4M (Roughly the same as BERT large, which
has 340M parameters). Since including all neighbors around mentioned entities of WordNet is not
efﬁcient, for simplicity, we use top 3 most common relations in WordNet in our experiment (i.e.
hyponym, hypernym, derivationally_related_form). For both datasets, we use a “two stage” ﬁne-tune
strategy to achieve our best performance, the FullTokenizer built in BERT is used to segment input
words into wordpieces.
For ReCoRD, the maximum length of answer during inference is set to 30, and the maximum length
of question is set to 64. Questions longer than that are truncated. The maximum length of input
sequenceT8is set to 384. Input sequences longer than that are segmented into chunks with a stride
of 128. Fine-tuning our model on ReCoRD costs about 18 hours on 4 V100 GPU with a batch size of
48. We freeze parameters of BERT and use Adam optimizer with a learning rate of 1e-3 to train our
knowledge module in the ﬁrst stage. The maximum number of training epochs of the ﬁrst stage is
10. The purpose of this is to provide a good weight initialization for our HMP. For the second stage,
the pre-trained BERT parameters and our HMP part will be ﬁne-tuned together. The max number of
training epochs is chosen from f4;6;8g. The learning rate is set to be 2e-5 with a warmup over the
8Refer to the PLM Encoding Module of Methodology Section.
12Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
ﬁrst6%of max steps, and linear decay until up to max epochs. For both two stages, early stopping is
applied according to the best EM+F1 score on the dev set every 500 steps.
For MultiRC, the maximum length of input sequence Tis set to 256. The summation of length of
question (Q) and length of candidate answer ( A) is not limited. Paragraph ( P) is truncated to ﬁt the
maximum length of input sequence. Fine-tuning KELM on MultiRC needs about 12 hours on 4 V100
GPU with a batch size of 48. For the ﬁrst stage ﬁnetuning, learning rate is 1e-4 and the maximum
number of training epochs is 10. For the second stage, the max number of training steps is chosen
fromf10000;15000;20000g. The learning rate is set to be 2e-5 with a warmup over the ﬁrst 10% of
max steps.
D S UPPLEMENTATION OF THE CASE STUDY SECTION
We provide deﬁnitions of the top 3 most relevant mentioned entities in WordNet that correspond to the
word examples mentioned in Case Study Section . Descriptions are obtained by using NLTK (Loper
& Bird, 2002). By comprehending the motivating example in the case study section, we can see that
KELM can correctly select the most relevant mentioned entities in the KG.
Word in text
(prototype)Mentioned entity
in WordNetDeﬁnition
banban.n.04 (0.72) an ofﬁcial prohibition or edict against something
ban.v.02 (0.21) prohibit especially by legal means or social pressure
ban.v.01 (0.06) forbid the public distribution of ( a movie or a newspaper)
fordford.n.05 (0.56)38th President of the United States;
appointed vice president and succeeded
Nixon when Nixon resigned (1913-)
ford.n.07 (0.24) a shallow area in a stream that can be forded
ford.v.01 (0.08) cross a river where it’s shallow
pardonpardon.v.02 (0.86) a warrant granting release from punishment for an offense
sentinel (0.10) -
pardon.n.02 (0.04) grant a pardon to
nixonnixon.n.01 (0.74)vice president under Eisenhower and 37th President
of the United States; resigned after the Watergate
scandal in 1974 (1913-1994)
sentinel (0.26) -
leadlead.v.03 (0.73) tend to or result in
lead.n.03 (0.12) evidence pointing to a possible solution
lead.v.04 (0.05) travel in front of; go in advance of others
outrageoutrage.n.02 (0.62) a wantonly cruel act
sentinel (0.38) -
Table 5: Deﬁnitions of mentioned entities in WordNet corresponding to the word examples in the case
study. The importance of mentioned entities is provided in the parenthesis. “sentinel” is meaningless,
which is used to avoid knowledge noise.
E E XPERIMENT ON COMMONSENSE CAUSAL REASONING TASK
To further explore the generalization ability of KELM, we also evaluate our method on COPA (Roem-
mele et al., 2011) (Choice of Plausible Alternatives), which is also a benchmark dataset in SuperGLUE
and can be used for evaluating progress in open-domain commonsense causal reasoning. COPA
consists of 1000 questions, split equally into development and test sets of 500 questions each. Each
question is composed of a premise and two alternatives, where the task is to select the alternative
that more plausibly has a causal relation with the premise. Similar to the previous two MRC tasks,
the development set is publicly available, but the test set is hidden. One has to submit the predicted
results for the test set to SuperGLUE to retrieve the ﬁnal test score. Since the implementation of
KELM is based on BERT large, we use it as our baseline for the comparison. The result of BERT large
is directly taken from the leaderboard of SuperGLUE. Table 6 shows the experiment results. The
injected KG is WordNet here, and we use accuracy as the evaluation metric.
The huge improvement over the baseline in this task demonstrates that knowledge in WordNet is
indeed helpful for BERT to improve the generalization ability to the out-of-domain downstream task.
13Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
Model dev test
BERT large - 70.6
KELMBERTlarge
WordNet76.1 78.0
Table 6: Performance comparison on COPA. The effectiveness of injecting knowledge (WordNet) are
shown.
F KELM: A FRAMEWORK OF FINETUNE -BASED MODEL -AGNOSTIC
KNOWLEDGE -ENHANCED PLM
We implement KELM based on the RoBERTa large, which has a similar number of trainable parameters
asBERT largebut uses nearly 10 times of training corpus than BERT large. Since the performances of
RoBEATa on the leaderboard of SuperGLUE are based on ensembling, we also ﬁnetune RoBERTa large
on ReCoRD to produce the results of a single model. Comparisons of the results can be found
in Table 7, where you can also see an improvement there. However, that improvement is not
as signiﬁcant as we observed in BERT large. Reasons are two-fold: (1) Passages in ReCoRD are
collected from articles in CNN/Daily Mail, while BERT is pre-trained on BookCorpus and English
Wikipedia. RoBERTa not only uses the corpus that used in BERT (16 GB), but also an additional
corpus collected from the CommonCrawl News dataset (76 GB). ReCoRD dataset is in-domain
for RoBERTa but is out-of-domain for BERT. It seems that the improvements of KELM with
injecting general KGs (e.g. WordNet) on the in-domain downstream tasks are not as large
as the out-of-domain downstream tasks . A similar phenomenon can be also observed in the
experiment of SQuAD 1.1 (Refer to Appendix E). (2) The same external knowledge (WordNet,
NELL) can not help RoBERTa largetoo much, since RoBERTa is pre-trained on a much larger corpus
than BERT, knowledge in WordNet/NELL has been learned in RoBERTa.
Dev Test
Model EM F1 EM F1
PLM w/o
external knowledgeBERT large 70.2 72.2 71.3 72.0
RoBERTa
large 87.9 88.4 88.4 88.9
knowledge enhanced PLM
(ﬁnetune-based)KELMBERTlarge
Both75.1 75.6 76.2 76.7
KELMRoBERTalarge
Both88.2 88.7 89.1 89.6
knowledge enhanced PLM
(pretrain-based)LUKE 90.8 91.4 90.6 91.2
Table 7: Comparison of the effectiveness of injecting external knowledge between BERT and
RoBERTa. [*] Results are from our implementation.
We also list the results of LUKE (Yamada et al., 2020) in Table 7. LUKE is a pretrain-based
knowledge enhanced PLM and uses Wiki-related golden entities (one-to-one mapping) as the injected
knowledge source (about 500k entities9). It has more 128 M parameters than the total number
of parameters of the vanilla RoBERTa. As we summarized in Table 4 in the main text, the pre-
training task is also different compared with RoBERTa. Although LUKE gets better results compared
with vanilla RoBERTa and KELM, it needs 16 NVIDIA Tesla V100 GPUs and the training takes
approximately 30 days. Relying on hyperlinks in Wikipedia as golden entity annotations, lacking
the ﬂexibility to adapt the external knowledge of other domains, and needing re-pretraining when
incorporating knowledge, these limitations hinder the abilities of applications.
G E XPERIMENT ON SQUAD 1.1
SQuAD1.1 (Rajpurkar et al., 2016) is a well known extractive-style MRC dataset that consists of
questions created by crowdworkers for Wikipedia articles . It contains 100,000+ question-answer
pairs on 536 articles. We implement KELM based on the BERT large, and compare our results on the
development set of SQuAD 1.1 with KT-NET (Best result of KT-NET is based on injecting WordNet
only). Results are shown in Table 8
9For KELM, we only use 40943 entities in WordNet and 258 concepts in NELL.
14Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
Dev
Model EM F1
PLM w/o external knowledge BERT large 84.4 91.2
knowledge enhanced PLM
(ﬁnetune-based)KT-NETBERTlarge
WordNet85.1 91.7
KELMBERTlarge
WordNet84.7 91.5
Table 8: Performance comparison on the development set of SQuAD 1.1.
Results on KELM show an improvement over vanilla BERT. Both BERT and RoBERTa use English
Wikipedia as the corpus for pretraining. Since SQuAD is also created from Wikipedia, it is an
in-domain downstream task for both BERT and RoBERTa (while ReCoRD dataset is in-domain for
RoBERTa but is out-of-domain for BERT). This explains why RoBERTa achieves a much larger
improvement over BERT on the result of ReCoRD ( 71:3!88:4in EM on test set) than the one
on SQuAD 1.1 ( 84:1!88:9). The rest of the improvement is because RoBERTa uses 10 times of
training corpus than BERT and different pre-training strategies they used.
Interestingly, we ﬁnd the performance of KELM on SQuAD 1.1 is sub-optimal compared with KT-
NET. As we mentioned in the last paragraph of the Related Work Section , KT-NET treats all synsets
of entity mentions within the WN18 as candidate KB concepts. Via a specially designed attention
mechanism, KT-NET can directly use all 1-hop neighbors of the mentioned entities. Although this
limits the ability of KT-NET to select the most relevant mentioned entities (as we discussed in Case
Study Section ), information of these neighbors can be directly considered. Using neighbors of the
mentioned entities indirectly via the HMP mechanism makes it possible for KELM to dynamically
embed injected knowledge and to select semantics-related mentioned entities. However, SQuAD
is an in-domain downstream task for BERT, the problem of ambiguous meanings of words can be
alleviated by pretraining model on the in-domain corpus. Compared with KT-NET, a longer message
passing path in KELM may lead to sub-optimal improvement on the in-domain task.
H F URTHER DISCUSSIONS ABOUT THE NOVELTY W .R.TSKG/KT-NET
UKET deﬁned in KELM consists of three subgraphs in a hierarchical structure, each subgraph
corresponds to one sub-process of our proposed HMP mechanism and solves one problem presented
in the Hierarchical Knowledge Enhancement Module part of Methodology Section . SKG only
uses GNN to dynamically encode the extracted KG which corresponds to the ﬁrst part of UKET, it can
not solve the knowledge ambiguity issue and forbids interactions among knowledge-enriched tokens.
KT-NET deﬁnes a similar graph as the third part of UKET. However, the ﬁrst and second subgraphs
of UKET are absent. The second subgraph of UKET is independent of ideas of KT-NET and SKG,
thus KELM is not a simple combination of these two methods. We are the ﬁrst to unify text and
KG into a graph and to propose this hierarchical message passing framework to incorporate two
heterogeneous information. SKG/KT-NET can be interpreted as parts of the ablation study of
components of KELM . The result of SKG is ablation with the component only related to the ﬁrst
subgraph of UKET. While KT-NET only contains the third subgraph with a modiﬁed knowledge
integration module. KELM uses a dedicatedly designed HMP mechanism to let the information of
farther neighbors to be considered. However, longer information passing path than KT-NET makes
it less efﬁcient. In our experiments, KELM takes more 30% training time than KT-NET on both
ReCoRD and MultiRC.
I L IMITATIONS AND FURTHER IMPROVEMENTS OF KELM
Limitations for KELM are two-fold: (1) Meanings of mentioned entities in different KGs that share
the same entity mentions in the text may conﬂict with each other. Although HMP can help to select
the most relevant mentioned entities in a single KG, there’s no mechanism to guarantee the selections
across different KGs; (2) Note the knowledge-enriched representation in Eq.3 is obtained by simple
concatenation of the embeddings from different KGs. Too much knowledge incorporation may
divert the sentence from its correct meaning (Knowledge noise issue). We expect these two potential
improvements to be a promising avenue for future research.
15Accepted at the ICLR 2022 Workshop on Deep Learning on Graphs for Natural Language Processing
J F URTHER ANALYSIS AND DISCUSSION
KELM incorporates knowledge in KGs into the representations in the last hidden layer of PLM
(Refer to Methodology Section ). It is essentially a model-agnostic, KG-agnostic, and task-agnostic
framework for enhancing language model representations with factual knowledge from KGs. It
can be used to enhance any PLM, with any injected KGs, on any downstream task. Besides the
two Q&A-related MRC tasks we mentioned in the main text, we also evaluate KELM on COPA
and SQuAD 1.1 based on BERT large, results are presented in Appendix E and Appendix G, respec-
tively. To demonstrate KELM is a model-agnostic framework, we also implement KELM based on
RoBERTa largeand evaluate it on ReCoRD. The experiment is presented in Appendix F. Improvements
achieved by KELM over all vanilla base PLM models indicate the effectiveness of injecting external
knowledge.
However, the improvements of KELM over RoBERTa on ReCoRD and BERT on SQuAD 1.1 are
marginal compared with the ones on ReCoRD/MultiRC/COPA ( BERT largebased). The reason behind
this is that pretraining model on in-domain unlabeled data could boost performance on downstream
tasks. Passages in ReCoRD are collected from articles in CNN/Daily Mail, while BERT is pre-trained
on BookCorpus and English Wikipedia. RoBERTa not only uses the corpus that used in BERT
(16 GB), but also an additional corpus collected from the CommonCrawl News dataset (76 GB).
ReCoRD is in-domain for RoBERTa but is out-of-domain for BERT. Similarly, SQuAD 1.1 is
created from Wikipedia, it is an in-domain downstream task for both BERT and RoBERTa. This
partially explains why RoBERTa achieves a much larger improvement over BERT on the result of
ReCoRD ( 71:3!88:4in EM on test set) than the one on SQuAD 1.1 ( 84:1!88:9). A similar
analysis can be also found in T5 (Raffel et al., 2020). From our empirical results, we can summarize
that general KG (e.g. WordNet) can not help too much for the PLMs pretrained on in-domain data.
But it can still improve the performance of the model when the downstream tasks are out-of-domain.
Further detailed analysis can be found in our appendix.
Finding a popular NLP task/dataset that is not related to the training corpus of modern PLMs is
difﬁcult. Pre-training on large-scale corpus is always good if we have unlimited computational
resources and plenty of in-domain corpus. It has been evident that the simple ﬁnetuning of PLM is
not sufﬁcient for domain-speciﬁc applications. KELM can provide people another choice when they
do not have such a large-scale in-domain corpus and want to incorporate incremental domain-related
structural knowledge into the domain-speciﬁc applications.
16