SpeedRead: A Fast Named Entity Recognition Pipeline

Proceedings of COLING 2012: Technical Papers, pages 51–66, COLING 2012, Mumbai, December 2012. SpeedRead: A Fast Named Entity Recognition Pipeline Rami Al − Rf ou′ Steven Skiena Department of Computer Science Stony Brook University NY 11794, USA {ralrfou, skiena}@cs.stonybrook.edu ABSTRACT Online content analysis employs algorithmic methods to identify entities in unstructured text. Both machine learning and knowledge-base approaches lie at the foundation of contemporary named entities extraction systems. However, the progress in deploying these approaches on web-scale has been been hampered by the computational cost of NLP over massive text corpora. We present SpeedRead (SR), a named entity recognition pipeline that runs at least 10 times faster than Stanford NLP pipeline. This pipeline consists of a high performance Penn Treebank- compliant tokenizer, close to state-of-art part-of-speech (POS) tagger and knowledge-based named entity recognizer. KEYWORDS: Tokenization, Part Of Speech, Named Entity Recognition, NLP pipelines. 51 1 Introduction Information retrieval (IR) systems rely on text as a main source of data, which is processed using natural language processing (NLP) techniques to extract information and relations. Named entity recognition is essential in information and event-extraction tasks. Since NLP algorithms require computationally expensive operations, the NLP stages of an IR system become the bottleneck with regards to scalability (Pauls and Klein, 2011). Most of the relevant work, conducted by researchers, was limited to small corpora of news and blogs because of the limitation of the available algorithms in terms of speed. Most of the NLP pipelines use previously computed features that are generated by other NLP tasks, which adds computational cost to the overall NLP pipeline. For example, named entity recognition and parsing need POS tags; co-reference resolution requires named entities. In effect, we anticipate lower speed for future tasks. A conservative estimate of a sample of the web news and articles can add up to terabytes of text. On such scale, speed makes a huge difference. For example, considering the task of annotating 10 TiBs of text with POS tags and named entities using a 20 CPU cores computer cluster would take at least 4 months using the fastest NLP pipeline available for researchers, our calculations show. Using our proposed NLP pipeline the time is reduced to a week. Several projects have tried to improve the speed by using code optimization. Figure 1a shows that Stanford POS tagger has improved throughout the years, increasing its speed by more than 10 times between 2006 and 2012. However, the current speed is twice slower than the SENNA POS tagger. (a) POS taggers performance. (b) NER taggers performance. Figure 1: Performance of NLP pipelines through the years over POS and NER tagging. Stanford POS tagger uses L3W model, its speed in 2006 is slow to be apparent in the graph. Stanford tagger uses CONLL 4 classes model. SENNA pipeline was first released in 2008 In this paper, we present a new NLP pipeline, SpeedRead, where we integrate global knowledge extracted from large corpora with machine learning algorithms to achieve high performance. Figures 1a and 1b show …

Content truncated to 500 words