Incorporating Linguistic Expertise Using ILP for Named Entity Recognition in Data Hungry Indian Languages

Developing linguistically sound and data-compliant rules for named entity annotation is usually an intensive and time consuming process for any developer or linguist. In this work, we present the use of two Inductive Logic Programming (ILP) techniques to construct rules for extracting instances of various named entity classes thereby reducing the efforts of a linguist/developer. Using ILP for rule development not only reduces the amount of effort required but also provides an interactive framework wherein a linguist can incorporate his intuition about named entities such as in form of mode declarations for refinements (suitably exposed for ease of use by the linguist) and the background knowledge (in the form of linguistic resources). We have a small amount of tagged data - approximately 3884 sentences for Marathi and 22748 sentences in Hindi. The paucity of tagged data for Indian languages makes manual development of rules more challenging, However, the ability to fold in background knowledge and domain expertise in ILP techniques comes to our rescue and we have been able to develop rules that are mostly linguistically sound that yield results comparable to rules handcrafted by linguists. The ILP approach has two advantages over the approach of hand-crafting all rules: (i) the development time reduces by a factor of 240 when ILP is used instead of involving a linguist for the entire rule development and (ii) the ILP technique has the computational edge that it has a complete and consistent view of all significant patterns in the data at the level of abstraction specified through the mode declarations. The point (ii) enables the discovery of rules that could be missed by the linguist and also makes it possible to scale the rule development to a larger training dataset. The rules thus developed could be optionally edited by linguistic experts and consolidated either (a) through default ordering (as in TILDE[1]) or (b) with an ordering induced using [2] or (c) by using the rules as features in a statistical graphical model such a conditional random field (CRF) [3]. We report results using WARMR [4] and TILDE to learn rules for named entities of Indian languages namely Hindi and Marathi.

[1]  Luc De Raedt,et al.  Mining Association Rules in Multiple Relations , 1997, ILP.

[2]  Douglas E. Appelt,et al.  FASTUS: A Finite-state Processor for Information Extraction from Real-world Text , 1993, IJCAI.

[3]  Ralph Grishman,et al.  Message Understanding Conference- 6: A Brief History , 1996, COLING.

[4]  Erik F. Tjong Kim Sang,et al.  Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition , 2003, CoNLL.

[5]  Ganesh Ramakrishnan,et al.  Learning Decision Lists with Known Rules for Text Mining , 2008, IJCNLP.

[6]  Raymond J. Mooney,et al.  Relational Learning of Pattern-Match Rules for Information Extraction , 1999, CoNLL.

[7]  Hendrik Blockeel,et al.  Top-Down Induction of First Order Logical Decision Trees , 1998, AI Commun..

[8]  Andrew McCallum,et al.  Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data , 2001, ICML.

[9]  Luc De Raedt,et al.  Top-down induction of logical decision trees , 1997 .

[10]  Stephen Soderland,et al.  Learning Information Extraction Rules for Semi-Structured and Free Text , 1999, Machine Learning.

[11]  Kalina Bontcheva,et al.  GATE: an Architecture for Development of Robust HLT applications , 2002, ACL.

[12]  Ellen Riloff,et al.  Automatically Constructing a Dictionary for Information Extraction Tasks , 1993, AAAI.

[13]  Nancy Chinchor,et al.  Overview of MUC-7 , 1998, MUC.

[14]  Erik F. Tjong Kim Sang,et al.  Introduction to the CoNLL-2003 shared task , 2003 .