Search many languages with high accuracy

Every language, including English, presents unique and difficult challenges for search applications to deliver relevant and precise results. Rosette® Base Linguistics (RBL) enables enterprise applications to effectively search or process text in many languages by providing a complete set of linguistic services. RBL enriches the original text in its native language for best-of-class natural language processing, improving speed and accuracy.

What is base linguistics?

Base linguistics refers to the core morphological building blocks that prepare your text for further analysis. In Chinese and Japanese base linguistics answers “what are the words?” for these languages written without spaces between words. In many European languages, it is associating related words based on their dictionary form, such as beau/beaux/belle/belles, which all mean “beautiful” in French, but vary by gender and number.

The leaders in multilingual search

Intelligent, successful search is about semantics. People want to type in a question and get an answer. To find more relevant results, the engine must associate words by meaning, that is, trace words back to their dictionary form based on the word’s context. For example, without looking at context, a reference to “spoke,” as part of a bicycle wheel, can be easily confused with the past tense of the verb “to speak.” While open source platforms now provide the basic framework for inverted full-text search engines, the challenges of accurate search are compounded as you add more languages to the queries and results. Rosette fills the linguistic need in Elasticsearch, Apache Solr, and applications that need to search across 30+ languages.

Product highlights

• Supports multiple languages
• Sentence tagging
• Tokenization
• Lemmatization
• Part-of-speech tagging
• Decompounding
• Chinese/Japanese readings
• Fast and scalable
• Cloud, on-premise, search plugin deployments

Language-specific features

• Japanese Base Linguistics
• Chinese Base Linguistics
• Arabic Base Linguistics

Part of speech tagging

As part of the lemmatization process, statistical modeling is used to determine the correct part of speech, even with ambiguous words. Each token is then tagged for enhanced comprehension and search relevancy. Because different languages have different grammars, part-of-speech tags differ.

Our base linguistics support the Universal POS Tag standard from which the developer can map to Penn Treebank or other POS tag systems.

Decompounding

Decompounding breaks compound words into sub-components and delivers each individual element to be indexed. This is especially useful for increasing search recall in languages such as German and Korean.

Example: German

“Samstagmorgen” is a compound word formed with “Samstag” (Saturday) and “morgen” (morning). Decompounding allows for an appropriate match when searching for “Samstag.”

Chinese and Japanese readings

Our base linguistics functionality understands the difference between Chinese and Japanese text when they are written in Han script, and accurately returns the pronunciation information. For Chinese text in Hanzi, Rosette returns the pronunciation information in pinyin transcriptions. For Japanese content, Rosette returns furigana transcriptions in Katakana. For example, if you call Rosette with “医療番組,” it will return this reading: “イリョウ”, “バングミ”.

Lemmatization

To associate related words, most search engines utilize a crude method of chopping off characters at the end of a word in the hopes of finding a common root form. This method, called stemming, often results in more recall, but poorer precision, associating unrelated words such as arsenic/arsenal, which share a stem (arsen). Our base linguistics tools find the true dictionary form of each word, known as a lemma, by using vocabulary, context, and advanced morphological analysis. Indexing the root form increases search precision and recall, and slims the search index by not indexing all inflected forms. Alternative lemmas are also made available to supplement indexing. Linguistic analysis is useful for every language, improving search recall and precision.

Tokenization

Many search tools use n-grams to break text into overlapping consecutive characters to create a search index in languages written without spaces between words. N-grams result in a larger index size and a reduction in precision. Our tools, in contrast, accurately identify and separate each word through advanced statistical modeling. The resulting token output (also known as segmentation) minimizes index size, enhances search accuracy, and increases relevancy.

Sample output:

{
"tokens": [
"The",
"fact",
"is",
"that",
"the",
"geese",
"just",
"went",
"back",
"to",
"get",
"a",
"rest",
"and",
"I",
"'m",
"not",
"banking",
"on",
"their",
"return",
"soon"
],
"lemmas": [
"the",
"fact",
"be",
"that",
"the",
"goose",
"just",
"go",
"back",
"to",
"get",
"a",
"rest",
"and",
"I",
"be",
"not",
"bank",
"on",
"they",
"return",
"soon"
]
}