Automatic acquisition of lexical semantic information using medium to small corpora

Since many speech and text processing techniques are portable with a limited amount of work from one language to another, the most daunting task for NLP and SP practitioners becomes to build the resources needing for those tools to operate, In particular, the constitution of “high-level” resources, such as advanced corpus annotations or linguistically motivated lexicons, can be extremely work-intensive. We present in this paper a system to assist the creation of semantic lexicons using small to medium-sized corpora, thanks to the combination of semantic class constitution and topic detection, and the development of specific statistical data analysis techniques for relatively small datasets. By reducing the amount of data needed for semi-automatic semantic lexicon acquisition, traditionally applied to 100 million-word corpus or more, we make this help for lexical resource acquisition applicable to the case of underresourced languages.