哑炮 - 麻省理工学院人工智能研究专业

哑炮

The Language Resource Switchboard

Claus Zinn
University of T ¨ubingen
Department of Computational
语言学
claus.zinn@uni-tuebingen.de

The CLARIN research infrastructure gives users access to an increasingly rich and diverse set
of language-related resources and tools. Whereas there is ample support for searching resources
using metadata-based search, or full-text search, or for aggregating resources into virtual col-
lections, there is little support for users to help them process resources in one way or another.
In spite of the large number of tools that process texts in many different languages, there is no
single point of access where users can ﬁnd tools to ﬁt their needs and the resources they have. 在
this squib, we present the Language Resource Switchboard (LRS), which helps users to discover
tools that can process their resources. 为了这, the LRS identiﬁes all applicable tools for a given
resource, lists the tasks the tools can achieve, and invokes the selected tool in such a way so that
processing can start immediately with little or no prior tool parameterization.

1. 介绍

The pan-European CLARIN project is building an eScience infrastructure for language-
related resources and tools (Hinrichs and Krauwer 2014). The project is driven by about
20 national consortia, each of which is sharing and making available their digital lan-
guage data and services. This produces an aggregate of resources that is both abundant
and diverse, but that also needs to be managed. Among the pillars of the infrastructure
is the Virtual Language Observatory (VLO), which gives users a metadata-based access
to language resources [1],1 the Federated Content Search (FCS) that gives users a full-
text search across resources [2], and the Virtual Collection Registry (VCR), 用户在哪里
can collect resources in a virtual set [3]. CLARIN uses the Component MetaData infra-
结构 (CMDI) to describe resources in a common, ﬂexible language (Broeder et al.
2010), and persistent identiﬁers based on the Handle system to ensure a persistent URL
addressing of resources.

到目前为止, information about tools was scattered throughout the CLARIN national
communities and consortia. Knowledgeable users would install their tool of choice

1 Web references are marked by bracketed number order and can be found at the end of this squib.

提交材料已收到: 28 六月 2017; 收到修订版: 14 行进 2018; 接受出版:
22 可能 2018.

土井:10.1162/coli a 00329

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

locally on their desktop machines, or they would direct their browser to the Web-
based tools they knew. 然而, less knowledgeable users and newcomers to the ﬁeld
struggled to keep up with the dynamics of the pan-European tool space.

The Language Resource Switchboard (LRS) aims at bridging the gap between re-
来源 (as identiﬁed in the VLO, FCS, and VCR) and tools that can process these
resources in one way or another. The LRS can be seen as an intelligent virtual tool
registry or a tool discovery device. For a given resource in question, it identiﬁes all tools
that can process the resource. It then sorts the tools in terms of the tasks they perform,
and presents a task-oriented list to the user. Users can then select and invoke the tool
of their choosing. By invoking the tool, all relevant information about the resource in
question is passed on to the tool, and the tool uses this information to set (一些) 它是
参数. This makes it easy for users to identify the right tools for their resource,
but also to use the chosen tool in the most effective way possible.

2. The Language Resource Switchboard

数字 1 displays the main interface of the switchboard. When users access it directly,
they are given three possible input methods: they can upload a ﬁle (左边), paste a URL
(中间), or type plain text (正确的). The switchboard can also be invoked with a ﬁle from
the VLO; the following scenario covers the VLO–switchboard link.

2.1 Usage Scenario

The VLO gives users access to about 1.6 million metadata records on language re-
来源; for about 40 mostly European languages, the VLO hosts at least a thousand
records each. 现在, consider an educational scenario in which a student of linguistics
would like to learn about tools that perform syntactic analyses on Polish texts. 为了这,
she ﬁrst searches the VLO to ﬁnd an interesting text that she would like to investigate
更远. On the VLO search results page, the student then clicks on the · · · area to invoke
the LRS with this ﬁle (见图 2(A)). In a new browser tab, the LRS opens and shows
a resource pane that depicts all relevant information about the ﬁle (见图 2(乙)).
The user is free to correct this metadata, before clicking on “Show Tools” to get to the
task-oriented view, shown in Figure 2(C). If the student is interested in the dependency
parsing task, 例如, then she may wish to get more information about the three
tools offered, in which case more detailed information about the chosen tool is given (看
数字 2(d)). When the student then clicks on “Click to start tool,” the chosen tool—here

数字 1
The switchboard’s main interface.

632

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Zinn

The Language Resource Switchboard

(A) The VLO – LRS interface.

(乙) The LRS Resource pane.

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

(d) The LRS Tool Detail view.

数字 2
The LRS in action.

the MaltParser [19]—opens in a new browser tab. The MaltParser obtains from the LRS
a reference to the resource, loads the resource, and sets all relevant parameters so that
the user is left to click on MaltParser’s “Analyze” command to start the process. 不
further user action is required to parameterize the MaltParser for this. If the student
would like to compare MaltParser’s result with the result of another dependency parser,
then the student returns to the browser tab holding the switchboard. 这里, only a simple
click is needed to start another applicable tool (例如, UDPIPE [18]) to process the same
resource.

2.2 Architecture

The back-end of the LRS consists of three main components: a resource proﬁler, 一个
application registry, and a matcher.

633

计算语言学

体积 44, 数字 4

数字 3
The metadata entry for UDPIPE.

The resource proﬁler is responsible for identifying those resource characteristics
that support identifying applicable tools that can process the resource. The current
version of the LRS takes into account the resource’s mimetype and language, 谁的
values are transferred from the VLO to the LRS. In the standalone version of the switch-
board [4], the proﬁler makes use of Apache Tika to identify these characteristics [5].

The application registry manages a set of metadata descriptions. 数字 3 depicts
the metadata entry for UDPIPE, a versatile tool from the Czech LINDAT/CLARIN
项目 [18]. For the LRS user, relevant metadata include the title of the application,
an English description about its capabilities, and contact information about the tool
提供者. For the LRS, the relevant parts are the tool’s task description (using a con-
trolled vocabulary, 例如, “tokenization,” “part-of-speech-tagging,” “dependency pars-
ing”), an ISO 639-3–based identiﬁcation of the languages the tool can handle, 和
media types it can process. For tool invocation, the metadata holds the tool’s Web
address and a list of parameters that the tool understands. With this information,
a URL can be constructed where all relevant information is URL-encoded.

The matcher uses the resource’s metadata from the proﬁler and the tools’ metadata
from the application registry to ﬁnd matches. For the given resource proﬁle, it computes
a list of all applicable tools and the analyses they offer. For the time being, only the
resource’s mimetype and language are taken into account.

The LRS is implemented using React, a Javascript library for building user interfaces
[6] and a number of nodejs-based libraries. The source code of the LRS is maintained in
a GitHub repository [7]. For more details, please consult Zinn (2016).

2.3 Tool Integration Effort

To attract a good number of tool makers to connect their tools to the LRS, 我们有
attempted to minimize their integration efforts for this. Web-based tools usually offer a
browser-based user interface where tool users are expected to specify the tool’s input as
well as a number of tool parameters before asking the tool to process the input with the
given parameterization. When the Web-based tool is called via the switchboard, a part
of the tool’s parameterization is already known—namely, 输入, its language, 和它的

634

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Zinn

The Language Resource Switchboard

media type. When the switchboard invokes the tool via a URL request, all three pieces
of information can be encoded in the URL. But it is the tool developer who decides
the names for the URL parameters, and whether all three parameters, or other default
参数, should be passed on. As a bare minimum, only the input parameter needs
encoding. A tool connected to the LRS should be able to extract this parameter from the
tool’s invocation URL and fetch the resource from the given location.

The browser-based tools must hence be extended to parse and interpret parameters
added to the tools’ URL invocation path. In case the tool already provides a parameter
passing via URL encoding, the switchboard “honors” the existing parameter names.
为了这, standard parameter names used by the LRS such as “input” are mapped to
a tool’s parameter name, 说, “URL.” The metadata entry for each tool captures such
信息 (see the “mapping” structure in Figure 3). 当然, the tool maker should
also update the browser-based interface, making it clear to its users that the tool is aware
of the parameterization it received from its LRS-based invocation. On the LRS side, 这
main integration task is the creation of the metadata entry for the tool (见图 3).

2.4 Status

At the time of writing (六月 2018), a total of 60 browser-based applications have been
connected to the switchboard; 见表 1 for a selection thereof. Note that the WebLicht
work engine has multiple entries with predeﬁned workﬂows to match the tasks and
languages at hand. These workﬂows engage well-known NLP tools from Stanford’s
CoreNLP toolset [14], the OpenNLP library [15], and the Berkeley NLP Group [16].

The LRS also supports a “batch mode” by accepting a zip archive of ﬁles sharing a
common media type and language. 笔记, 然而, that the switchboard itself does not
perform the batch processing itself. 相当, it delegates the task to those tools that can
handle such zip archives.

3. 相关工作

There are a number of directory services for language processing software. A faceted
search on the Virtual Language Observatory, 例如, yields around 400 不同的
metadata entries for language-related processing tools and services. Each metadata
record comes with a short description and sometimes has a link to the tool’s homepage
where more information is available, such as the tool’s download location. The VLO has
no systematic classiﬁcation of the tools so it is hard for users to identify tools of interest.
LT World is one of the older Web sites on language technology and maintains
a classiﬁed list of tools, especially for processing written language: Dimensions are
“Tokenization,” “Naming Entities Detection,” “Lemmatizer,” “Language Guesser,“ 和
so forth (J ¨org, Uszkoreit, and Burt 2010); also see [8]. Whereas the entries from the
VLO are automatically harvested from various metadata providers, the LT World list
seems to be maintained manually, and links to a tool’s homepage are sometimes broken.
还, most tools of LT World are not Web-based and must be installed on a user’s
desktop machine. The LINDAT/CLARIN Web site is a Web site that goes beyond a
simple yellow-paging of tools [9]. Although its focus is predominantly on tools for
the processing of Czech text ﬁles, it allows users to invoke each of the Web services
via a user interface with a common look and feel across the services. 这里, users can
deﬁne their input, and inspect the output of all REST-based services. The Institute of

635

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

桌子 1
Tools connected to the Language Resource Switchboard (分段).

Tool(s)

语言(s)

任务

Tokenization

Morphology Analysis

Lemmatization

PoS Tagging

Shallow Parsing
Dependency Parsing

Named Entity
Recognition

机器翻译
Text Analytics

Spelling correction

NLP suite for Dutch
N-Gramming

Text Summarization

Ucto
CST Tokenizer
WebLicht-Tokenization-TUR
WebLicht-Morphology-X
Morfeusz 2
CST Lemmatizer
WebLicht-Lemmas-X
WebLicht-POSTags-Lemmas-X
Concraft
Concraft→Spejd
WebLicht-Dep-Parsing-X
MaltParser
Alpino
UDPIPE
CST Name Recognizer
WebLicht-NamedEntities-X
GATE NER
Liner 2
NER NLTK
Oersetter (NLD-FRY, FRY-NLD)
Voyant Tools
T-scan
WebSty
Fowlt
Valkuil
Frog
FoLiA-stat
Colibri Core
Concraft→Bartek→NicolasSummarizer
Summarize

nld, eng, deu, fra, ita, fry
many languages
tur
deu, eng
pol
many languages
eng, deu
deu, fra, ita, eng
pol
pol
nld, eng, deu, slv, hrv, srp
pol
nld
for languages, 比照. 如图. 3
和
deu, eng, slv
eng, deu
pol
eng
nld, fry
many languages
nld
pol
eng
nld
nld
nld, generic
many languages
pol
pol
pol
pol
pol

Coreference Resolution Concraft→Bartek
Word Sense Disamb.
Sentiment Analysis

WoSeDon
Concraft→Sentipejd

计算机科学, Polish Academy of Sciences, has a well organized Web page on
language tools and resources for Polish [10]. 这里, each tool comes with its own Web
页, often with background information with references to publications, download
locations, installation instructions, and sometimes with a demo page where the tool can
be tried online.

WebLicht is a workﬂow engine that gives users access to a good range of natural
language processing tools (Hinrichs, Hinrichs, and Zastrow 2010). WebLicht offers
predeﬁned workﬂows (“easy-chains”), but also an advanced mode, where users can
construct their own processing chains. For a tool to be integrated into WebLicht, 它必须
be adapted to read and write TCF-compliant data. Each tool in the workﬂow reads
its input from the TCF source, and extends the TCF ﬁle with its processing results.
WebLicht’s tool landscape is dynamic. At regular intervals, it harvests tool metadata
from CLARIN repositories; the metadata lists the speciﬁc input–output behavior of the
tool, informing the WebLicht orchestrator about possible workﬂow constructions.

636

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Zinn

The Language Resource Switchboard

The Language Application Grid (LAPPS Grid) [11] is an open, Web-based infras-
tructure that offers a very good range of language-related tools. 而且, the tools
can be composed into tool chains using a graphical editor. Similar to WebLicht, 为了
tools to take part of the Grid, they need to be adapted so that they can read and
write LAPPS Grid formats. Tool developers should be aware of the LAPPS Interchange
Format (LIF) and the Web Services Exchange Vocabulary (WSEV). The LAPPS Grid also
offers additional features such as visualization and the sharing of various types of data
(such as LAPPS interaction histories, workﬂows, and visualizations).

4. 讨论

The task of the LRS is rather simple: given some information about a resource, 帮助
users to ﬁnd and start a tool that can process the resource in one way or another. 一次
the user has been directed to the tool, LRS’s engagement ends; it is a deliberate design
decision that the LRS is unaware of the processing, and that it will not know whether
the processing succeeded or failed. 当然, users may manually save the output of
the tool to a ﬁle, and then upload it to the standalone version of the LRS to ﬁnd post-
processing tools. But such tools are rarely available, as most tools have their proprietary
output format that other tools cannot read.

The LRS was never designed to be a workﬂow engine such as WebLicht or the
LAPPS Grid. 尽管如此, a good share of tools connected to the switchboard execute
predeﬁned processing workﬂows, which in turn make use of well-known NLP software.
Also note that both LAPPS Grid and WebLicht come with their build-in user interface
能力, and that they run all tools or services in the background. 相比之下, 这
LRS is mainly advertising tools that come with their own Web-based GUI, each of which
is optimized for the tool at hand. Naturally, the LRS and its tools are ill-suited for big
data processing, as browser-based tools face http request timeouts. For big data, 用户
should directly use services amenable to such processing such as WebLicht as a service;
看 [12].

The ease of tool integration gives the switchboard a head start for growth. Tools
need to be amended only to process a number of URL-encoded input parameters and to
update their user interface to reﬂect that parameters were successfully processed. No in-
ternal data formats need to be changed. In this respect, note that the integration of tools
into WebLicht or LAPPS Grid is very labor-intensive because tools must be adapted
to become capable of reading and writing WebLicht- or LAPPS Grid–compliant data.
还, we do not see the switchboard as a competitor to WebLicht or the LAPPS Grid.
相反, each new predeﬁned workﬂow in WebLicht can be advertised by and
directly invoked from the switchboard, hence directing more user trafﬁc to WebLicht.
In a similar way, we aim at integrating LAPPS Grid tools with the switchboard, 和
hence making available many more popular NLP tools at users’ ﬁngertips.

The switchboard is also being integrated with B2DROP, a cloud storage open to
all European researchers [17]. 这里, a plug-in enables users to invoke the switch-
board with their cloud resources (Zinn 2018), hence guiding further trafﬁc towards the
switchboard.

Two big issues affect the usability of the switchboard: the quality of the metadata
in the VLO, and access restrictions to resources and tools. The VLO lists a considerable
number of resources with incorrect or incomplete mimetypes. Searching, 例如,
for resources of type “text/plain” might yield textual resources that are not plain, 但
enriched with other annotations. When users load such a resource in the LRS, 和

637

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

计算语言学

体积 44, 数字 4

subsequently process the resource with a tool of their choice, they may ﬁnd that the
resource’s content is not what they expected. 相似地, accessibility and authentication
issues affect the usability of the switchboard. When the user invokes an applicable
tool in the switchboard, the tool might not be able to download the resource from
the resource provider. 还, some tools require authentication, and some users may
not have the proper access rights to make use of the tools. 为了这, a number of user
delegation issues need to be tackled, with many technical intricacies involved (Blumtritt
等人. 2014). 目前, users can often side-step the problem by ﬁrst downloading the
resource from the provider, and then uploading it to the standalone version of the LRS.
With a tool gaining greater visibility through the “business leads” of the switch-
board, tool developers must address operational challenges. In an ideal world, all the
tools listed by the LRS are indeed online, but sometimes, the server hosting the tool is
ofﬂine, the tool’s behavior has changed so that an LRS invocation fails, or another error
发生. For the time being, the LRS does not “ping” the tools’ server to check whether
the tools are available for invocation. 将来, the LRS might want to replicate
the system status report facility of CLARIN [13] to inform developers about the ofﬂine
status of their tools.

5. 结论

Given a language-related resource, the LRS helps users to identify the tools capable of
processing the resource and then to invoke the tool with a default conﬁguration that
minimizes user involvement. The switchboard solves this task well, and it has received
positive feedback from the linguistics community. The LRS is indeed perceived as the
missing low-technology link between language-related resources and the tools that can
process them. The LRS fulﬁlls an important role within and across the CLARIN com-
社区. The switchboard’s discovery service adds enormous visibility to all the tools
connected to the switchboard, 因此, increases their use. It also helps users to better
get to know the many tools that each consortium partner contributes to the alliance. 这
switchboard, 然而, is not restricted to the tools that originate in the CLARIN world.
It already features quite a few classic, well-known NLP tools, and it expands its tool
registry steadily. Tool makers are invited to contact the author to get their tool integrated
with the switchboard so that the LRS gives a more accurate reﬂection of the existing
tool landscape. 同样地, students and researchers are encouraged to use the switch-
board for their work, and to report their feedback to steadily improve the switchboard
服务.

致谢
The project has been funded from the
European Union’s Horizon 2020 研究
and innovation programme under grant
agreement no. 676529 (CLARIN-PLUS).
Thanks to the many tool developers who
helped to integrate their tool with the
switchboard. Thanks to Twan Goosen for
linking the VLO with the LRS. Thanks to
Dieter Van Uytvanck for valuable feedback
on usability aspects, and for promoting the
LRS in the CLARIN community. 谢谢
to the anonymous referees for their
comments.

参考
Blumtritt, J。, 瓦. Elbers, 时间. Goosen,

中号. Hinrichs, 瓦. Qiu, 中号. Sall´e, 和
中号. Windhouwer. 2014. User delegation in
the CLARIN Infrastructure. Link¨oping
Electronic Press, (116):14–24.

Broeder, D ., 中号. Kemps-Snijders, D. 货车

Uytvanck, 中号. Windhouwer, 磷. Withers,
磷. Wittenburg, 和C. Zinn. 2010. A data
category registry- and component-based
metadata framework. 在诉讼程序中
7th International Conference on Language
Resources and Evaluation (LREC’10),
pages 43–47, Valetta.

638

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

Zinn

The Language Resource Switchboard

Hinrichs, E., 中号. Hinrichs, 和T. Zastrow.

2010. WebLicht: Web-based LRT services
for German. In Proceedings of the ACL 2010
系统演示, http://dl.acm
.org/citation.cfm?id=1858933
.1858938.

Hinrichs, 乙. 瓦. 和S. Krauwer. 2014. 这

CLARIN research infrastructure:
Resources and tools for ehumanities
学者. In Proceedings of the 9th
International Conference on Language
Resources and Evaluation (LREC’14),
pages 25–29, Uppsala.

J ¨org, B., H. Uszkoreit, 和一个. Burt. 2010. LT

世界: Ontology and reference
information portal. In Calzolari, Nicoletta,
Khalid Choukri, Bente Maegaard, 约瑟夫
Mariani, Jan Odijk, Stelios Piperidis, Mike
Rosner, and Daniel Tapias, 编辑,
Proceedings of the 7th International Conference
on Language Resources and Evaluation
(LREC’10), pages 1002–1006, Valetta.
Zinn, C. 2016. The CLARIN Language

Resource Switchboard. In Proceedings
of the CLARIN Annual Conference,
Aix-en-Provence.

Zinn, C. 2018. A Bridge from EUDAT’s
B2DROP cloud service to CLARIN’s
Language Resource Switchboard.
Link¨oping Electronic Press, (147):36–45.

Web References (Accessed: 六月 15, 2018)

[1] vlo.clarin.eu
[2] clarin.eu/contentsearch
[3] clarin.eu/vcr
[4] switchboard.clarin.eu
[5] tika.apache.org
[6] reactjs.org
[7] github.com/clarin-eric/LRSwitchboard
[8] www.lt-world.org
[9] lindat.mff.cuni.cz/en/services
[10] clip.ipipan.waw.pl/LRT
[11] www.lappsgrid.org
[12] weblicht.sfs.uni-tuebingen.de/WaaS
[13] status.clarin.eu
[14] stanfordnlp.github.io/CoreNLP
[15] opennlp.apache.org
[16] nlp.cs.berkeley.edu/software.shtml
[17] b2drop.eudat.eu
[18] ufal.mff.cuni.cz/udpipe
[19] maltparser.org

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

639

我

D
哦
w
n
哦
A
d
e
d

F
r
哦
米
H

t
t

:
/
/

d
我
r
e
C
t
.

米

我
t
.

e
d
你
/
C
哦

我
我
/

我

A
r
t
我
C
e
–
p
d

F
/

4
4
4
6
3
1
1
8
0
9
9
1
1
/
C
哦

我
我

_
A
_
0
0
3
2
9
p
d

乙
y
G
你
e
s
t

哦
n
0
8
S
e
p
e
米
乙
e
r
2
0
2
3

640 Squibs image

下载pdf