We present a novel approach for detecting, classifying, and linking entities from Twitter posts (tweets). The task is challenging because of the noisy, short, and informal nature of tweets. Consequently, the proposed approach introduces several methods that robustly facilitate successful realization of the task with enhanced performance in several measures.
Microblogging services, such as Twitter, are rapidly
becoming virtually ubiquitous. This is attributable to the fact
that they are extremely valuable mechanisms that enable
us to obtain live and raw information in real time. In this
paper, we describe our approach to the #Microposts 2015
NEEL challenge [
The main diculty inherent in this task stems from the noisy, short, and informal nature of tweets. The performance of previous approaches su↵ered because they tended to focus on well-written, long texts such as news articles. Our system explicitly focuses on tweets and addresses the problem using a variety of methods working together.
Our proposed system addresses the task in an end-to-end
manner. Unlike most of the previous approaches, the system
does not use an external named entity recognition system
(NER) to generate candidates of the entity mentions
because the current NER typically performs badly for tweets
[
Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
Consequently, we constructed three supervised machinelearning models to detect NIL entity mentions and predict the types (e.g., PERSON and LOCATION ) of the detected mentions. 2.
Our proposed system addresses the task using a procedure comprising the following five steps: 1) preprocessing, 2) mention candidate generation, 3) mention detection and disambiguation, 4) NIL mention detection, and 5) type prediction. 2.1
We tokenize a tweet and assign part-of-speech tags to the
resulting tokens using ARK Twitter Part-of-Speech Tagger
[
In this step, the candidates of the entity mentions are generated from the tweet using the methods described below.
The system generates candidates using the mention-entity dictionary; it first takes all the n-grams (n < 10) from the tweet and performs queries to the dictionary using the text surface of each of these n-grams. The following four methods are used to retrieve candidates: • Exact search retrieves mention candidates that have text surfaces exactly equal to the query text. • Fuzzy match searches the mention candidates that have text surfaces within a certain distance of the query text measured by edit distance. • Approximate token search obtains mention candidates whose text surfaces have a significant ratio of words in common with the query text. • Acronym search retrieves mention candidates with possible acronyms2 that include the query text. 2We generate acronyms by tokenizing the mention surface and simply taking the first characters of the resulting tokens.
The system first generates possible mention candidates using the above methods, sorts these candidates according to the number of occurrences in which the mention appear as a link to the referent entity, and selects the top k candidates (k = 100 for exact search and k = 30 for other methods). Additionally, we experimentally set the maximum allowed edit distance of fuzzy match to two and the minimum ratio of approximate token search to 66% because these settings achieve the best scores in our experiments. 2.3
In this step, we first assign a score to mention candidates using a supervised machine-learning model. In this case, we used random forest as the machine-learning algorithm.
We started out using features similar to those proposed in
previous works [
Finally, the overlapped entity mentions are resolved. We start with the beginning of the tweet and iterate over the candidate entity mentions. Then, we detect the mention if the corresponding span of the mention has not already been detected and the score assigned to the mention is above the threshold. If multiple mentions are found, the mention with the highest score is selected. 2.4
We formulate the task of detecting NIL mentions from a tweet as a supervised classification task to assign a binary label to each of all possible n-grams (n < 10). Random forest is again used as our machine-learning algorithm.
We extract several features from the output of the Stanford NER3 using two types of models: 1) a standard threeclass model, and 2) a model that does not use capitalization as a feature. We also use the ratio of capitalized words as an indicator of the reliability of the capitalization in the tweet. Additionally, various other features are used, such as partof-speech tags of the surrounding words and the length of the n-grams. 2.5
We cast the task of detecting types of mentions as a multiclass supervised classification task. In the previous steps, we extracted two types of mentions: entity mentions and NIL mentions. Thus, we are able to build two separate classifiers to predict the entity types for each type of mention. We developed two machine-learning models using logistic regression and random forest and created the final model by building an ensemble model on top of these models in order to boost the performance.
The primary features used to detect types of entity mentions are the corresponding entity classes retrieved from DBpedia and Freebase (e.g., FictionalCharacter, SportsTeam). Name strong link match strong typed mention match mention ceaf We also use our 300 dimensional entity-embeddings constructed from Wikipedia and the predicted entity types of the Stanford NER.
In order to detect the types of NIL mentions, we use
features extracted from word embeddings. Here, the GloVe
Twitter 2B model [
In our experiments, we used the #Microposts 2015 dataset
[
Table 1 shows a summary of our experimental results. We evaluated our system using the following three measures: strong link match to evaluate the performance of linking entities, strong typed mention match to measure the performance of mention detection and entity typing, and mention ceaf for calculating the performance of clustering detected mentions into entity mentions or NIL mentions.4 We successfully achieved accurate performance in all of the measures.
In this paper, we described our approach for detecting, classifying, and linking entity mentions in tweets. We introduced a novel machine-learning approach specifically targeted at tweets and successfully achieved enhanced performance on the #Microposts2015 dataset.