OCR and Automated Translation for the Navigation of non-English Handsets: A Feasibility Study with Arabic

OCR and Automated Translation for the Navigation of non-English Handsets: A Feasibility Study with Arabic Jennifer Biggs and Michael Broughton Defence Science and Technology Organisation Edinburgh, South Australia {firstname.lastname}@dsto.defence.gov.au Abstract In forensics, mobile phones or handsets store potentially valuable information such as Contact lists, SMS Messages, or possibly emails and Calendar appoint- ments. However, navigating to this con- tent on non-English configured handsets, when the operator is untrained in the lan- guage, becomes a difficult task. We dis- cuss a feasibility study that explored the performance of optical character recogni- tion (OCR) systems against Arabic men- us on handset LCD screens. Further, a method of automated spell correction and translation is explored considering fully automated or user-interactive workflow options. A capability technology demon- strator for non-English handset naviga- tion was implemented based on outcomes of these studies, providing a platform for investigating workflow and usability. 1 Introduction Some cellular exploitation tools support imaging of the handset display after the operator has nav- igated the handset menus to the content of inter- est. Such tools may support any handset type. However, navigating to this content on handsets configured for languages other than English (LOTE) is challenging for operators not trained in the language. We undertook several feasibility studies to in- vestigate the navigation of LOTE handsets for CELLEX purposes. The studies investigated the merits of: 1) applying Commercial-Off-The- Shelf (COTS) Optical Character Recognition (OCR) tools to photographed displays of hand- sets; and 2) combining LOTE OCR outputs with a method of automated translation. 1.1 OCR accuracy COTS OCR systems are typically optimised for recognition of text at resolutions in excess of 100 dots per inch (dpi), such as scans of printed doc- uments, newspapers or magazines, advertising accuracy rates of up to 97%. Batawi and Abulnaja (2012) report accuracy rates of between 94.8% and 97.8% for a selection of printed newspaper and magazine Arabic texts. Recogni- tion of non-degraded printed pages may still re- quire identification of optimal image pre- processing options (Yale University, 2008a; 2008b). Recognition accuracy for degraded doc- uments may be expected to be significantly de- creased (Herceg et al. 2005). To utilise a COTS OCR application within a larger system architecture or workflow, where images do not meet application parameters, addi- tional image pre-processing can be applied. Chang et al. (2007) and Chang et al. (2009) used Sakhr Automatic Reader v8.0 on photographed images of text. 1.2 Automated translation of OCR outputs When applying Machine Translation (MT) pro- cessing to OCR output text, OCR errors are compounded. Chang et al. (2009) combined Sakhr Automatic Reader v8.0 with a statistical MT system known as PanDoRA (Zhang and Vo- gel, 2007), noting that word errors introduced by OCR were amplified during MT processing. For example, in translation of generated images of text from the Basic Travel Expression Corpus (BTEC) (Eck & Hori, 2005), the BLEU score of image text translation without errors was 43.12, while a 10.7% word recognition error rate se- verely drops the BLEU score to 28.56 …

Content truncated to 500 words