Rosette API で関係の検出がさらに容易に


Rosette API 1.5の新機能や改善点について本ブログに連載していきます。初回は特定タイプの関係抽出の紹介です。
関係抽出はエンティティと文中の関連情報との関係を識別する機能です。新バージョンは、主にビジネスや家族関係(間柄)に特化した17の関係タイプを抽出するようにトレーニングされているので、2つのエンティティ間のこれらの関係を正確に抽出できます。もちろん、この17タイプ以外の関係を抽出することもできます。
——-

Relationship discovery gets even easier with Rosette API

January 18, 2017

Rosette API 1.5 is our most ambitious update since the inception of our cloud API offering last spring, with new features, capabilities, and improvements. To help our users get the most out of the release, we’re bringing you a series of posts highlighting some of the bigger changes, starting with the introduction of targeted relationship extraction.

What is relationship extraction?

If you’re a text analytics veteran, you’ve likely utilized entity extraction tools in the past. Building on the results of entity extraction and linking, Rosette relationship extraction identifies how different entities are related to each other.

Using linguistic tools and Wikidata to bootstrap its knowledge, Rosette identifies the exact action connecting the entities and other related information contained within a sentence. Machine learning methods applied over parse trees and entity mentions allows Rosette to analyze the connection, and then return the components that make up the relationships.

What’s new?

The /relationships endpoint now returns targeted relationships, specific relationships that connect two entities (think “Bill Gates was educated at Harvard). We trained Rosette to identify 17 common relationship types, focused primarily on business and family connections.

But what if you’re looking to discover new relationship types or the type you’re interested in isn’t included in our pre-built list? Don’t worry, Rosette can still return “open” relationships as well, just set the option “discoveryMode” to “true” to find any and all explicit relationships between entities, e.g., person, location, organization, time, etc., in unstructured text.

The open relationship extractor focuses on two-party relationships expressed through verbs, possessives, or appositions. For example, given:

“New Ebola patient, Martin Salia, may be flown to a Nebraska hospital,”

the open relationship extractor will return:

“fly to” between the named entity “Martin Salia” and the noun phrase “Nebraska hospital”

Rosette’s relationship extraction capabilities now have three underlying components:

  1. *NEW* A collection of pre-built targeted relationship extractors
  2. An open relationship extractor
  3. A human-in-the-loop workflow for building targeted relationship extractors *on-premise only*

The new pre-built collection of targeted relationship types is useful for both financial services and open source intelligence use cases, and also serves as a great starting point if you’re new to relationship extraction in general.

Try it yourself

Sign up for a free API account (no credit card required) and start extracting entity relationships from your text today. To make it even easier, we even plugin to RapidMiner. As always, let us know if you use Rosette to power any interesting projects and we’ll feature you on our blog!

Go further

For those extra tricky problems, our on-premise solutions also include custom relationship extraction training. The targeted extraction workflow combines automation and human guidance to construct a targeted relationship extractor that fits your goals and data. Talk to our sales team to learn more.