Named Entity rEcognition and Linking (NEEL) from text is an essential task in many Natural Language Processing (NLP) applications because it enables a better understanding of the content. However in the context of Social Media, NEEL is challenging due to the higher level of writing mistakes, fast language dynamics and often lack of context. To this end, we adapted Kanopy - an unsupervised graph-based topic disambiguation system - to be used for the task of NEEL in the domain of Twitter, a fast-paced micro-blogging platform. We describe the design of our solution and report the results obtained by our system using the official corpus of Tweets for the NEEL 2016 Challenge [10].
We designed our system, Kanopy4Tweets, inspired on
Kanopy [
Our system must deal with Tweets, very noisy and short pieces of text. Kanopy was not intended for this type of content, hence we need to adapt it. Kanopy has three stages: Name Entity Recogniser (NER), Entity Linking and Entity Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citatCioonpoynrtihgehfirtst pcag2e0.1C6ophyerilgdhtsbyforacuotmhpoorn(esn)t/soowf ntheirs(ws)o;rkcowpnyeindgbypoetrhmersitttheadn AoCnMly mfourstpbreihvoantoereadn. dAbasctraadcteimngiwcipthucrrpedoisteis.permitted. To copy otherwise, or repPuublbislhis, htoedpoastsopnasretrvoerfs tohreto#reMdisitcrriboupteostotsl2is0ts1,6reqWuiorersksphriorpsppercoificepeedrminisgssio,n aanvda/oilraabfleee.oRneliqnuesatspeCrmEisUsiRonsVforol-m16p9er1m(ishstiotnps:@//acmeu.orr-gw.s.org/Vol-1691) Disambiguation. We need to adapt each one of them in order to process tweets. 2.1
We integrated a NER specially designed for Twitter into
Kanopy based on the well-known GATE NLP Framework [
In recent versions, GATE now includes a ready to use
Twitter-specific pipeline for IE called TwitIE [
To integrate GATE into Kanopy4Tweets we use an embedded engine in an optimised multi-threading REST API server. This server operates in two recognition modes: (1) the TWITTER mode uses the above described TwitIE pipeline for text processing, and (2) the NORMAL mode uses the standard ANNIE pipeline included in GATE for regular text processing. Those two modes are used during the on-line Tweet processing and the off-line DBpedia indexing stages respectively. More details are in Section 3. 2.2
We built a DBpedia index using a selection of datasets in order to find suitable resource candidates for each extracted named entity. The datasets used in our system are: 1 DBpedia ontology, mapping based types, mapping based properties, titles, short abstracts, article categories, categories(labels), categories(skos) and redirects.
We stored the DBpedia datasets in a single binary file
using the HDT RDF format [
After the index is built, we use the named entities found by our NER component to query DBpedia. For each entity we obtain a list of resource candidates using the following topdown strategy: first, we search a resource with exactly the same title property as the name entity. If a result is found, we use it immediately for the next stage as a candidate, otherwise we query again the index looking for a resources with a title similar to the named entity. Again, if we find some candidates we use them for the next stage. Each time we query the index we use a query that is more general than the previous one. The last query finds matches of the named entity in the abstract property. This process stop when at least one candidate is selected. If we found no candidates, the named entity is assgined as a NIL resource. 2
One challenge we found using this approach is that the number of found resource candidates tends to be large, increasing the processing time for the disambiguation stage. In such situations, we reduce the number of candidates by ranking them according to the document score assigned by the indexing engine (details in Section 3) and selecting the top-k elements. 2.3
The selected resource candidates for each named entity
found in a Tweet are processed by Kanopy in an
unsupervised graph-based approach for joint disambiguation that
combines the same datasets mentioned in Section 2.2,
resulting in a graph of linked resources [
The disambiguation process takes as input a list of named
entities, each containing a list of candidate DBpedia
resources after the linking stage. The output is the selection of
the best candidate resource for each input named entity. For
this, relatedness of the candidates with respect to the
candidate resources of all the other entities is calculated based
on the number of paths between the resources weighted by
the exclusivity of the edges of these paths [
The input named entities are then jointly disambiguated
and linked to the candidate resources with the highest
combined relatedness. This process is fully detailed in [
When no resource in the knowledge base can be found to be linked to a named entity, a NIL resource is assigned. However, we might have sets of unlinked named entities that can be similar under some criteria. Hence we are interested on further clustering all the found NIL-linked named entities.
For this clustering, we use the following hierarchical and
incremental simple approach: we iterate over each NIL-linked
entity and aggregate them into clusters one by one. The first
element is assigned into an initial cluster, then the next item
is compared to the previous ones using the Monge-Elkan
similarity measure [
Kanopy4Tweets has two different pipelines as shown in Figure 1. The first pipeline is used during a one-time initial off-line processing, while the other is used during the normal on-line processing of input Tweets.
As seen in Figure 1, the off-line pipeline starts with a list of input datasets to create a single HDT file. The number of datasets depends on the context of the task. In our case, we selected a small number of datasets that are mostly composed by text but we discarded those containing RDF literals. Furthermore, each output RDF triple is filtered in order to decide whether it should be included for indexing or not. For example, titles, redirections and disambiguation links are among the key properties considered.
One fundamental resource property used is the short abstract. For this attribute, we use our NER service in NORMAL mode to extract named entities to be used for indexing. On the other hand, the HDT file is also the input for the graph building. As mentioned in Section 2.2, this graph contains all the DBpedia resources that are indexed.
Both processes, the index and graph creation, are slow and usually take a few hours to complete. However the offline pipeline is only required to run if the knowledge base is modified. For the indexing engine we use Elastic 3 and for handling the graph we use Neo4j 4. Both solutions offer high performance and robustness.
The on-line pipeline is the main process that reads an input TSV file, extracts a list of Tweets and generates an ouput TSV file with the found linked named entities, all as shown in Figure 1. Each input Tweet is sent to our NER service in TWITTER mode and the returned list of named entities are labelled according to their function of in the text. Next, each named entity is used to search the index and a list of candidates is generated. Later, the Tweet is sent to the disambiguation service that determines which candidate DBpedia resources are best suited for the Tweet and named entity. If the link to an entity can not be determined, the named entity is goes to the NIL clustering component where it assigned a numbered NIL cluster.
Regardless of the NER operation mode (which is defined per instance), the REST API exposes a single POST method for entity extraction from text. This method accepts an array of texts to process and returns an array of the same texts including a per-text array of found named entities. These entities returned contain the position indices in the text where they are located, their class and any meta-data that the GATE pipeline identified, i.e. first and last names for a Person class or additional data from a gazetteer for Locations, among others. This meta-data can be later used to
refine and/or further disambiguate the initially found named entities. The returned entities classes belong to the set of annotations available in the TwitIE and ANNIE pipelines: Hashtag, UserID, URL, Address, Date, Identifier, Location, Money, Organization, Percent, Person and Phone. When operating in NORMAL mode (using ANNIE), the Hashtag, UserID and URL classes are never returned.
We evaluated Kanopy4Tweets using the NEEL 2016 dev and training datasets. The results for each dataset are shown in the table 1. The three measures used to evaluate our system are: i) Strong-typed Mention Match (SMM) is a micro-averaged evaluation of entity mentions including the entity offset and type, ii) Strong Link Match (SLM) is a micro-averaged evaluation of links, iii) Mention Ceaf (MC) is based on a one-to-one alignment between system and gold entity clusters.
We presented Kanopy4Tweets, a system for NEEL of Twitter data. Our approach uses Kanopy, an unsupervised system that exploits the graph structure of DBpedia in order to disambiguate a list of candidate resources for named entities. Furthermore, our system implements a hierarchical clustering approach for grouping unlinked named entities according to their textual similarity.
Our results reported a relatively poor performance, however we believe this is mostly because of the small number of entities found in the Tweets and the hardness of lack of context in Twitter data. In addition, hashtags (words used by users to tag their Tweets) and user mentions require further special considerations in the NER phase. Finally, we want to investigate better approaches for traversing the disambiguation graph. 5.
This publication has emanated from research conducted with the financial support of Science Foundation Ireland (SFI) under Grant Numbers SFI/12/RC/2289 and 14/TIDA /2419, and National University of Ireland Galway under the Hardiman Scholarship.