Using Embeddings for Both Entity Recognition and Linking on Tweets
Download as PPTX, PDF1 like486 views
This paper presents an approach for named entity recognition (NER) and linking in tweets using deep learning techniques. The approach uses bidirectional LSTMs with character-level embeddings to perform NER, and averages word embeddings around entity mentions to link them to DBpedia entries. Training data was expanded from the provided 1629 tweets to nearly 14,000 tweets by adding additional annotated data. Experimental results showed the effectiveness of embeddings for entity disambiguation. The best system achieved an overall score of 0.541 on the test data.
Using Embeddings for Both Entity Recognition and Linking on Tweets
- 1. Using Embeddings for Both Entity Recognition and Linking in Tweets Giuseppe Attardi, Daniele Sartiano, Maria Simi, Irene Sucameli Dipartimento di Informatica Università di Pisa Università di Pisa
- 2. Task Description Annotate named entity mentions in tweets and link them to corresponding DBpedia entry Training set provided by the organizers consists of just 1629 tweets
- 3. Approach Two stages: 1. NER 2. Entity linker NER requires more training data Added to training set 6439 tweets from the PoSTWITA task Trained NER applied to set of 7100 tweets, then manually corrected Final NER training set 13,945 tweets
- 4. Approach 1. Train word embeddings on a large corpus of Italian tweets 2. Train a bidirectional LSTM character-level Named Entity tagger, using the pre-trained word embeddings 3. Build a dictionary mapping Italian titles to English Dbpedia title , e.g. Milano (http://dbpedia.org/resource/Milan, Location) 4. Map anchor texts from Wikipedia to the above titles: Person Cristoforo_Colombo Colombo
- 5. Approach 2 5. Create word embeddings from the Italian Wikipedia 6. For each page abstract compute the average of the word embeddings of its tokens and map it to its URL 7. Perform Named Entity tagging on the test set 8. For each extracted entity, compute the average of the word embeddings for a context of words around the entity. 9. Annotate the mention with the DBpedia entity which is closest to those abstracts 10. For the Twitter mentions, use the Twitter API to obtain the real name and set the category to Person if present in a gazetteer of names.
- 6. Note The last step is somewhat in contrast with the task guidelines
- 7. Bi-LSTM Character-level NER Character-level features are learned No hand-engineered features like prefix and suffix rMari o rMari rMar rMa lM lMa lMar lMari M a r i rMario eMario lMario rM lMari o
- 8. NER Tagger B- PER I- PER O B- LOC c1 c2 c3 c4 r1 r2 r3 r4 l1 l2 l3 l4 Mario Monti a Roma CRF Layer Bi LSTM Encoder Word Embeddings
- 9. NER Accuracy (dev set) Category Precision Recall F1 Character 50.00 16.67 25.00 Event 92.48 87.45 89.89 Location 77.51 75.00 76.24 Organization 88.30 78.13 82.91 Person 73.71 88.26 88.33 Product 65.48 60.77 63.04 Thing 50.00 36.84 42.42
- 10. Official results Run Mention ceaf Strong typed mention match Strong link match Final score UniPI.3 0.561 0.474 0.456 0.5034 UniPI.1 0.561 0.466 0.443 0.4971 Team2.base 0.530 0.472 0.477 0.4967 UniPI.2 0.561 0.463 0.443 0.4962 UniPI.3 without mention check 0.616 0.531 0.451 0.541
- 11. Discussion Effectiveness of embeddings in disambiguation See improvement in the strong link match score from run UniPI.2 to UniPI.3 Liverpool_F.C. vs Liverpool Italy_national_football_team vs Italy S.S._Lazio vs Lazio Diego_Della_Valle vs Pietro_Della_Valle Nobel_Prize vs Alfred_Nobel Bad cases: Maria_II_of_Portugal for Maria Luke_the_Evangelist for Luca Tesla GPU granted by NVIDIA
- 12. Conclusions Deep Learning approach Char embedding helpful for dealing with noise in tweets in NER Word emeddings used for semantic relatedness in entity linking Side product: a new gold resource of 13,609 tweets (242,453 tokens) annotated with NE categories, leveraging the resource from the Evalita PoSTWITA task