The task of Named Entity Extraction and Linking (NEEL) challange 2015 [5] is considered as two successive tasks : Named Entity Extraction (NEE) from the tweets and Named Entity Linking (NEL) with DBpedia. For NEE task we use CRF++ [1] to create a language model on the given training data. For entity linking, we use DBpedia Spotlight.
The entity linking problem is well explored on normal text. However, the existing techniques of entity linking do not work well on short messages as the microblogs do not have sucient context to classify (or disambiguate) the mentions. In this work we have identified the mention by creating an entity recognition model on the given training data and link them to the DBpedia using DBpedia Spotlight.
The rest of the paper is organized as follows: Section 2 describes our proposed approach which includes data preparation and feature selection for named entity recognition model Copyright c 2015 held by author(s)/owner(s); copying permitted only for private and academic purposes.
Published as part of the #Microposts2015 Workshop proceedings, available online as CEUR Vol-1395 (http://ceur-ws.org/Vol-1395) creation and entity linking method. Section 3 describes the setup for web access. The result of our work is discussed in Section 4. Section 5 illustrates the future scope of our work followed by the references.
In our approach we have divided the Named Entity
Extraction and Linking (NEEL) [
The NER task is viewed here as a sequence labeling problem. Given an input tweet, this step aims to identify the word sequences that constitute a Named Entity and classify each such entity into one of the predefined classes. For entity recognition and classification task, we have developed a model on the given training data using Conditional Random Fields (CRFs) which is an undirected graphical model used mainly for sequence labeling.
As we have discussed in the previous section that the
context of the tweets is short, sometimes noisy and informal and
thus, their syntactic structures are not always comparable to
the normal texts. [
In the data preparation step, we have identified the word
sequences refering to a Named Entity(NE) in the training
data using the gold standard. The training data is
tokenized, part-of-speech(POS) tagged using Twitter NER[
We have experimented with various feature types, various window lengths and their combinations and come up with the following feature set which gave us a good result. We experimented with some context window lengths and 5 gave us good results.
• Contextual (Word) Features: a context window of size five: Wi 2 Wi 1 Wi Wi+1 Wi+2
Pi 2 Pi 1 Pi Pi+1 Pi+2 • Part-of-Speech (POS) Features: a context of size five: • Word having Capitalization: binary feature • Word having Punctuation: binary feature • Is a Digit: binary feature • Word having a Dot: binary feature • Word having hashtag: binary feature
For linking, we use the annotations returned by DBpedia Spotlight REST API as the candidates and look for the longest matching surface forms.
We take the output of the NEE task and collect the named entities that are extracted and their categories. To identify correct start position we check for # and @. For each tweet, using the B/I tags we find the longest consecutive entities that make up a single entity. For example, in the tweet above, ”London riots” would be treated as a single entity. For each tweet, DBpedia Spotlight REST API is accessed with confidence and support set to 0 with accepted return text in XML. We use the DBpedia Spotlight’s annotate endpoint to obtain all the links at once. For each entity returned from DBpedia Spotlight, if the surface form is found to be a substring of any of the entities and if a substring match is found the corresponding URI is returned. For named entities for which no match is found, if it is an existing nil entity then the nil id is returned, else the nil counter is incremented and returned.
We used perl for transforming the data. We used the CMU
Twitter NLP[
We use JSP to create our REST API, which uses perl which
in turn uses curl to connect to DBpedia Spotlight[
The precision for strong link match with the training set itself is 30.49%, recall is 30.29% and f1 is 30.39%. For the tagging of correct entity type the precision with the training set itself is 82.89%, recall 82.35% and f1 82.62%. The precision for strong link match with the development set is 14.82%, recall is 7.97% and f1 is 10.37%. For the tagging of correct entity type the precision with the training set itself is 41.65%, recall 22.41% and f1 29.14%.
As we can see using the CMU POS tagger[