Significant amount of literature is available on compound splitting of long words albeit for non-English languages-especially European. Not surprisingly, there has been not much work for English as it is not a compounding language like some of its European counterparts. However, Internet domain names in general are compound English words, e.g. "bankofamerica.com". Compound splitting can be effectively employed to extract information from domain names. In this paper, an data-driven learning technique for splitting English compound words is described which among others uses features like normalized frequency, length of parts and n-gram. The splitting F-measure is higher than the published approaches. We applied this technique on a real life web search application where the queries are mistyped domain names routed through sources like ISPs and browsers. Relevant and meaningful keywords were extracted out and shown to the user as a value added search option. Results show a very high click-through rate and increased commercial value.

Abstract-Developments in semantic search technology have motivated the need for efficient and scalable entity annotation techniques. We demonstrate RAD: a tool for Rapid Annotator Development on a document collection. RAD builds on a recent approach [1] that translates entity annotation rules into equivalent operations on the inverted index of the collection, to directly generate an annotation index (which can be used in search applications). To make the framework scalable, we use an industrial strength indexer, Lucene [2] and introduce some modifications to its API.The index also serves as a suitable representation for making quick comparisons with an indexed ground truth of annotations on the same collection to evaluate precision and recall of the annotations. RAD achieves at least an order of magnitude speedup over the standard approach of annotating a document-at-a-time as adopted by GATE [3]. The speedup factor increases with increase in the size of the collection, making RAD scalable. We cache intermediate results from the index operations, enabling quick update of the annotation index as well as speedy evaluation when rules are modified. This makes RAD suitable for rapid and interactive development of annotators.

Entity annotation is emerging as a key enabling requirement for search based on deeper semantics: for example, a search on 'John's address', that returns matches to all entities annotated as an address that co-occur with 'John'. A dominant paradigm adopted by rulebased named entity annotators is to annotate a document at a time. The complexity of this approach varies linearly with the number of documents and the cost for annotating each document, which could be prohibiting for large document corpora. A recently proposed alternative paradigm for rule-based entity annotation [16], operates on the inverted index of a document collection and achieves an order of magnitude speed-up over the document-based counterpart. In addition the index based approach permits collection level optimization of the order of index operations required for the annotation process. It is this aspect that is explored in this paper. We develop a polynomial time algorithm that, based on estimated cost, can optimally select between different logically equivalent evaluation plans for a given rule. Additionally, we prove that this problem becomes NP-hard when the optimization has to be performed over multiple rules and provide effective heuristics for handling this case. Our empirical evaluations show a speed-up factor upto 2 over the baseline system without optimizations.

Protein sequences vary in their length and are not readily amenable to conventional data mining techniques that need mapping in a fixed dimensional space. Thus, majority of the current methods for protein sequence classification are based on alignment of the query sequence either with a sequence or a profile of the sequence family. We present a method for mapping of protein sequences in a fixed dimensional descriptor space. The descriptors such as amino acid content and amino acid pair association rules were used along with routinely available classification methods. An experiment on one hundred Pfam families showed classification accuracy of 98% with support vector machines classifier. Information gain based feature selection helped simplify the model and improve accuracy. Interestingly, a large number of the selected features were based on the association rules of Glycine or Aspartic acid residues suggesting their role in the conserved loops among evolutionarily related proteins. Further, in another experiment, the approach was tested for classification of proteins from 39 Pfam families of protein kinases. Support vector machines classifier provided an accuracy of approximately 96%. The method provides an alternative to conventional profile based methods for protein sequence classification.