About text mining

Today, an increasing amount of information is being held in unstructured and semistructured formats, such as customer e-mails, call center notes, open-ended survey responses, news feeds, web forms, etc. This abundance of information poses a problem to many organizations that ask themselves, "How can we collect, explore, and leverage this information?"

Text mining is the process of analyzing collections of textual materials in order to capture key concepts and themes and uncover hidden relationships and trends without requiring that you know the precise words or terms that authors have used to express those concepts. Although they are quite different, text mining is sometimes confused with information retrieval. While the accurate retrieval and storage of information is an enormous challenge, the extraction and management of quality content, terminology, and relationships contained within the information are crucial and critical processes.

Text mining and data mining

For each article of text, linguistic-based text mining returns an index of concepts, as well as information about those concepts. This distilled, structured information can be combined with other data sources to address questions such as:

  • Which concepts occur together?
  • What else are they linked to?
  • What higher level categories can be made from extracted information?
  • What do the concepts or categories predict?
  • How do the concepts or categories predict behavior?

Combining text mining with data mining offers greater insight than is available from either structured or unstructured data alone. This process typically includes the following steps:

  1. Identify the text to be mined. Prepare the text for mining. If the text exists in multiple files, save the files to a single location. For databases, determine the field containing the text.
  2. Mine the text and extract structured data. Apply the text mining algorithms to the source text.
  3. Build concept and category models. Identify the key concepts and/or create categories. The number of concepts returned from the unstructured data is typically very large. Identify the best concepts and categories for scoring.
  4. Analyze the structured data. Employ traditional data mining techniques, such as clustering, classification, and predictive modeling, to discover relationships between the concepts. Merge the extracted concepts with other structured data to predict future behavior based on the concepts.

Text analysis and categorization

Text analysis, a form of qualitative analysis, is the extraction of useful information from text so that the key ideas or concepts contained within this text can be grouped into an appropriate number of categories. Text analysis can be performed on all types and lengths of text, although the approach to the analysis will vary somewhat.

Shorter records or documents are most easily categorized, since they are not as complex and usually contain fewer ambiguous words and responses. For example, with short, open-ended survey questions, if we ask people to name their three favorite vacation activities, we might expect to see many short answers, such as going to the beach, visiting national parks, or doing nothing. Longer, open-ended responses, on the other hand, can be quite complex and very lengthy, especially if respondents are educated, motivated, and have enough time to complete a questionnaire. If we ask people to tell us about their political beliefs in a survey or have a blog feed about politics, we might expect some lengthy comments about all sorts of issues and positions.

The ability to extract key concepts and create insightful categories from these longer text sources in a very short period of time is a key advantage of using Text Analytics. This advantage is obtained through the combination of automated linguistic and statistical techniques to yield the most reliable results for each stage of the text analysis process.

Linguistic processing and NLP

The primary problem with the management of all of this unstructured text data is that there are no standard rules for writing text so that a computer can understand it. The language, and consequently the meaning, varies for every document and every piece of text. The only way to accurately retrieve and organize such unstructured data is to analyze the language and thus uncover its meaning. There are several different automated approaches to the extraction of concepts from unstructured information. These approaches can be broken down into two kinds, linguistic and nonlinguistic.

Some organizations have tried to employ automated nonlinguistic solutions based on statistics and neural networks. Using computer technology, these solutions can scan and categorize key concepts more quickly than human readers can. Unfortunately, the accuracy of such solutions is fairly low. Most statistics-based systems simply count the number of times words occur and calculate their statistical proximity to related concepts. They produce many irrelevant results, or noise, and miss results they should have found, referred to as silence.

To compensate for their limited accuracy, some solutions incorporate complex nonlinguistic rules that help to distinguish between relevant and irrelevant results. This is referred to as rule-based text mining.

Linguistics-based text mining, on the other hand, applies the principles of natural language processing (NLP)—the computer-assisted analysis of human languages—to the analysis of words, phrases, and syntax, or structure, of text. A system that incorporates NLP can intelligently extract concepts, including compound phrases. Moreover, knowledge of the underlying language allows classification of concepts into related groups, such as products, organizations, or people, using meaning and context.

Linguistics-based text mining finds meaning in text much as people do—by recognizing a variety of word forms as having similar meanings and by analyzing sentence structure to provide a framework for understanding the text. This approach offers the speed and cost-effectiveness of statistics-based systems, but it offers a far higher degree of accuracy while requiring far less human intervention.

To illustrate the difference between statistics-based and linguistics-based approaches during the extraction process, consider how each would respond to a query about reproduction of documents. Both statistics-based and linguistics-based solutions would have to expand the word reproduction to include synonyms, such as copy and duplication. Otherwise, relevant information will be overlooked. But if a statistics-based solution attempts to do this type of synonymy—searching for other terms with the same meaning—it is likely to include the term birth as well, generating a number of irrelevant results. The understanding of language cuts through the ambiguity of text, making linguistics-based text mining, by definition, the more reliable approach.

Understanding how the extraction process works can help you make key decisions when fine-tuning your linguistic resources (libraries, types, synonyms, and more). Steps in the extraction process include:

  • Converting source data to a standard format
  • Identifying candidate terms
  • Identifying equivalence classes and integration of synonyms
  • Assigning a type
  • Indexing and, when requested, pattern matching with a secondary analyzer

Step 1. Converting source data to a standard format

In this first step, the data you import is converted to a uniform format that can be used for further analysis. This conversion is performed internally and does not change your original data.

Step 2. Identifying candidate terms

It is important to understand the role of linguistic resources in the identification of candidate terms during linguistic extraction. Linguistic resources are used every time an extraction is run. They exist in the form of templates, libraries, and compiled resources. Libraries include lists of words, relationships, and other information used to specify or tune the extraction. The compiled resources cannot be viewed or edited. However, the remaining resources can be edited in the Template Editor or, if you're in an interactive workbench session, in the Resource Editor.

Compiled resources are core, internal components of the extraction engine within Text Analytics. These resources include a general dictionary containing a list of base forms with a part-of-speech code (noun, verb, adjective, and so on).

In addition to those compiled resources, several libraries are delivered with the product and can be used to complement the types and concept definitions in the compiled resources, as well as to offer synonyms. These libraries—and any custom ones you create—are made up of several dictionaries. These include type dictionaries, synonym dictionaries, and exclude dictionaries.

Once the data have been imported and converted, the extraction engine will begin identifying candidate terms for extraction. Candidate terms are words or groups of words that are used to identify concepts in the text. During the processing of the text, single words (uniterms) and compound words (multiterms) are identified using part-of-speech pattern extractors. Then, candidate sentiment keywords are identified using sentiment text link analysis.

Note: The terms in the aforementioned compiled general dictionary represent a list of all of the words that are likely to be uninteresting or linguistically ambiguous as uniterms. These words are excluded from extraction when you are identifying the uniterms. However, they are reevaluated when you are determining parts of speech or looking at longer candidate compound words (multiterms).

Step 3. Identifying equivalence classes and integration of synonyms

After candidate uniterms and multiterms are identified, the software uses a normalization dictionary to identify equivalence classes. An equivalence class is a base form of a phrase or a single form of two variants of the same phrase.The purpose of assigning phrases to equivalence classes is to ensure that, for example, side effect and 副作用 are not treated as separate concepts. To determine which concept to use for the equivalence class—that is, whether side effect or 副作用 is used as the lead term— the extraction engine applies the following rules in the order listed:

  • The user-specified form in a library.
  • The most frequent form, as defined by precompiled resources.

Step 4. Assigning type

Next, types are assigned to extracted concepts. A type is a semantic grouping of concepts. Both compiled resources and the libraries are used in this step. Types include such things as higher-level concepts, positive and negative words, first names, places, organizations, and more.

Linguistic systems are knowledge sensitive—the more information contained in their dictionaries, the higher the quality of the results. Modification of the dictionary content, such as synonym definitions, can simplify the resulting information. This is often an iterative process and is necessary for accurate concept retrieval. NLP is a core element of Text Analytics.