📄 BootCaT: Bootstrapping Corpora and Terms from the Web

Summary

Translators and terminologists often need small, current, domain-specific corpora for a single job. Collecting these by hand from the web takes too long. BootCaT is a modular set of perl scripts that automates the work and needs only a handful of seed terms. Random tuples of seeds are sent to Google, and the top pages are downloaded and converted to text. New single-word terms are then found by comparing word frequencies against a reference corpus with the log odds ratio. These terms become seeds for the next round. Once the final corpus is built, a recursive procedure extracts multi-word terms, using connectors learned from the data together with frequency constraints. The authors tested the toolkit by building English and Italian psychiatry resources to support translating one article. About two thirds of the sampled pages were judged acceptable. English term extraction performed reasonably well, while Italian was clearly weaker. Even so, both corpora contained most of the article's single-word terms. The work shows that throwaway specialized corpora can be produced quickly and cheaply.

Key points

• English run. The run used 6 seeds (abstract words absent from the Brown corpus) and two iterations. It produced:
  • a resulting corpus totaling roughly 1.5 million words;
  • 1800 unigram terms;
  • 1507 multi-word terms of up to 5 words.
• Italian run. It started from only two acronyms (eeg, ptsd). It queried seed pairs instead of triplets to increase recall and used a reference corpus of roughly 17 million words of Italian web text.
• Page quality. In random samples of 30 pages, 20 English and 21 Italian pages were judged acceptable. Most rejected pages were uninformative, such as reference lists or staff lists.
• Multi-word term precision. In samples of 100 terms, 73 English terms were judged good, compared with 32 Italian ones. Thirty-one of the Italian candidates were English chunks, and these were excluded from the good count.
• Recall against the translation task.
  • English: the term lists covered 38 of 43 unigram terms (88%) and 11 of 29 multi-word terms (38%) from the source text.
  • Italian: the lists contained almost none of the text's terms. However, the corpus itself included 85% of the unigram terms and 32% of the multi-word terms.

Interesting ideas

• I1. Choosing seeds by contrast with a general corpus. Words from a domain text that never appear in a general reference corpus make cheap, automatic seeds. This could bootstrap domain crawls for any language that has a balanced reference corpus.
• I2. Acronyms as seeds across languages. Technical acronyms tend to stay the same across languages, so they can seed a target-language corpus without translating anything. This is handy when building comparable corpora, for example Ukrainian–English.
• I3. Querying random seed combinations. Sending random tuples of seeds as queries pulls in diverse but on-topic pages. Using shorter tuples raises recall for languages with less web content, which is a simple tuning knob for low-resource settings.
• I4. Multi-word term extraction from data-driven connectors and subsumption. Connectors are learned from words that occur between known terms, and a constant k filters out nested candidates whose frequencies are nearly equal. Together they yield multi-word terms without POS tagging, which suits languages with limited tools.
• I5. Task-based recall evaluation. True recall cannot be measured on the open web. Instead, the authors checked how many terms from the actual document to be translated appear in the extracted lists and in the corpus. This practical evaluation design transfers well to other unsupervised harvesting pipelines.