A large part of the world’s information is unstructured. That means it is available in natural language text that is accessible to humans who are willing to invest the time to read it, but inaccessible to any but the most rudimentary automated analysis. As such, unstructured information is still the step child of the big data revolution.

Looking at Wikipedia and its structured siblings, DBPedia and Wikidata, it becomes obvious that more information is generated in unstructured form than is codified in formats accessible to computation. The fate of the semantic web painfully illustrates this.

Humans continuously curate knowledge in the form they know best: natural language.

ayfie makes that trove accessible to analysis and prediction algorithms by unearthing the sematic structure behind the text.

Documents can be collected from almost any common storage system or application, from cloud storage providers to ECM tools to information stored in CRM systems and historical databases.

Documents can also be pushed through an API, in order to facilitate integration into third-party applications. After collection, documents are converted into a common text representation, keeping all metadata and structural information intact.

ayfie benefits from over a decade of experience in the enterprise search market where connecting to many different systems – from content management to email servers – is key to success. Over 70 connectors are available and can be extended using a robust SDK. Text and metadata can be extracted from all popular document formats. Text contained in images is extracted using OCR technology.

By applying large electronic dictionaries that codify information about semantic and syntactic properties of words, ayfie enriches texts. It does so with base forms, synonyms, spelling variants, and dependency properties for all words that occur in the text.

ayfie finds and annotates all external structure that exists in the documents, such as headings, paragraphs, or sentences, and makes them accessible to further analysis.

Many of our customers use this functionality alone in order to greatly improve the search experience on their customer-facing portals. Dealing with these linguistic properties of words is key to consistent and relevant search results, especially in domain-specific applications.

ayfie builds on more than 30 years of research in compiling large scale electronic dictionaries and other linguistic resources. Inflectional forms, synonyms, and other phenomena (such as decomposition in Norwegian and German) are handled for all major European languages.

ayfie analyzes the semantic structure inside the document insofar as it is relevant to the use case. It then brings it into a form where it is more efficiently consumable for humans and formal enough for machines.

What distinguishes ayfie from all previous and current approaches to text analysis is our view of the basic elements of meaning in language. Our algorithms do not simply manipulate isolated words - which are always either strongly ambiguous or extremely vague - but complex semantic constructions which express meanings at a higher level. We apply combinations of very large semantic dictionaries that encode a lot of information about the entities mentioned in the texts together with millions of semantically typed parsing rules that know how relations are expressed syntactically. We are therefore able to extract the names of entities, the facts and the opinions expressed in the text.

Franz Guenthner, professor of Computational Linguistics at the Center for Information and Language Processing at the Ludwig-Maximilian-University (LMU) and Technology Advisor at ayfie Inc.

For instance, there are thousands of ways to describe the acquisition of one company by another. ayfie recognizes them all and makes them accessible to search applications and further statistical analysis.

Word inflections continuously lead to inconsistent search results on Google.

This effect is of course multiplied for multi-term queries.

In the same manner, doctors use many different expressions, ranging from colloquial to very formal, in describing obesity in patient notes. ayfie understands all of these variations and can map them to the correct ICD 10 code.

For dealing with structure extraction and semantics, ayfie builds on well-researched linguistic frameworks based on the works of Zelig Harris, Maurice Gross and Franz Guenthner. This sound theoretical foundation is combined with our blazingly fast proprietary extraction engine and exhaustive resources in many languages and domains.

All extracted information is stored in the context of the original documents. It is then represented in different forms in order to facilitate research, further processing, and analysis. ayfie’s storage architecture is scalable, from indexing thousands of documents for an eDiscovery case, to analyzing an arbitrary number of nodes, such as scientific articles from the biggest publishers in the world.

We support both graph-based and search-based access patterns to efficiently execute different types of algorithms on the extracted information and raw data. A vast set of operations is included out of the box. Additional ones can be added in the form of Spark jobs or ElasticSearch plugins to access further text and results of all prior structurization.

ayfie uses a solid architecture that scales horizontally to any content size, combining open source innovations and proprietary procedures.

By extracting the most important information from any kind of document, ayfie can power or improve a wide variety of applications in a lot of different domains.

While it is relatively simple to construct a decent suggest / type-ahead search functionality from structured data such as an ecommerce product catalogue, the same is not true for unstructured data such as news text.

ayfie can, for instance, extract the most important concepts, person names, locations and organizations from financial news so they can be used on ”Page Zero“, as Microsoft calls it:

ayfie can automatically extract persons, locations, organizations, key phrases and many more entities out of the box that are perfect ”hand rails“ into the content to be searched.

ayfie’s linguistic preprocessing and extraction capabilities make search applications much more consistent, without sacrificing relevancy. It does so by examining phenomena like synonyms, inflectional forms, and other variations.

ayfie’s advanced extraction engine can even be used to power natural language search by reducing both the query and the document content to their semantic core.

Thus, the query

matches the document content

when processed by ayfie, because both sentences are reduced to a meaning representation of


and are thus considered a match of each other.

By aggregating extracted predicates of the same type, ayfie can build tables of events or facts in any domain required. For instance, a large industrial manufacturer uses ayfie to analyze incident reports about turbine malfunctions. While these reports are written in plain text by service personnel, ayfie is able to turn them into tabular structures:

Ring Position Finding Recommendation
Xla2 16 Crack Replaced
Xla3 7 Wear Smoothing

Based on the tables created by ayfie, the manufacturing company can now precisely answer questions like

"Which part failed most often because of outside heat?"

"Which malfunction is the most likely for a certain class of parts?"

"Which environmental conditions are the most detrimental to the overall reliability of the component?"

Thus, all free-text service reports and even historical texts can be subjected to rigid analyses that would normally only be possible on pre-structured data.

Pure statistical and machine learning approaches promise to yield good results with any kind of text once the initial work of implementing the algorithm is done. However, training these algorithms inevitably requires training data, and lots of it if the results are to be useful.

The more structure we can give to a text by breaking it into tokens, finding the boundaries of sentences, dealing with synonyms and inflectional forms, extracting salient phrases and entities, the closer we get to a structured representation of the text.

Thanks to this structure, the data can be processed more easily by learning algorithms that look for patterns and trends in the data. For instance, clustering and categorization yield higher precision when ayfie’s structured representation of the text is used instead of simple token vectors. This produces a higher gain in precision than using a better learning algorithm on the unstructured text.

ayfie comes with a wide range of visualization and analysis options and can even drive 3rdparty data analysis tools.

At its core, ayfie is a technological platform capable of powering many different applications in a variety of markets. Currently, we are focussing on eDiscovery and Compliance in text-heavy and regulated industries including Legal, Insurance, Finance, and Healthcare.

For more information on our go to market strategy, please contact us below.

We solve business problems with data analysis tailored to your needs by leveraging machine learning, linguistic analysis and years of experience solving complicated problems.

Contact us

— an ayfie whitepaper