Simple method of extracting events

Re-build event graph, event extraction methods can not help but to use, here we quoted a summary of someone else's used as a reference

https://www.cnblogs.com/cyandn/p/10915394.html

Relation extraction

 

Defined : automatic identification with certain semantic relationships between entities. The number of participating entities can be divided into binary relation extraction (two entities) and polyols relation extraction (three or more entities).

 

Semantic relationship between two entities of interest can be obtained (arg1, Relation, arg2) triples, wherein arg1 and arg2 two entities, Relation indicates the semantic relationship between entities.

 

Depending on the data source, the extraction can be divided into the following three relationships:

• Relationship-oriented structured text extraction: a document including tables, XML documents, databases and other data
• Relations for unstructured text extraction: plain text
• Relationship oriented semi-structured text extraction: between structured and unstructured

Depending on the range of text extraction, relation extraction can be divided into the following two:

• Sentence Relation Extraction: determining a sentence from what entities between two semantic relationship
• Corpus (chapters) relation extraction stage: Context is not limited to two target entities appearing

Based on the extracted field division, relation extraction can be divided into the following two:

• Relationship extraction qualified domain: decimating semantic relationships among the entities in the area of ​​one or more defined category defining relationship, it can be regarded as a text classification task
• Open field relation extraction: the relationship between categories is not limited

 

Qualified domain relation extraction method:

• Or by manually editing the learned template text entity relationship decimating and discrimination, it is limited by the quality and coverage of the template, the expandable not strong: template-based relation extraction method
• The extraction method based on machine learning relationship: the relationship is drawn as a classification problem

In which the relationship extraction method based on machine learning can be divided into supervised and weak supervision.

There supervised relation extraction method:

• Engineering based on feature: the need to convert a display example of the relationship between the classifier to the feature vector can accept
• Kernel function based on: a tree structure is directly processed, when calculating the relationship between the distance between the feature vectors no longer in use but with the inner product kernel
• Neural Network: direct and effective automatic learning features from the input text, said end

Relation extraction method weak supervision: does not require manual tagging large amounts of data.

Distance Supervision: Automatic training samples labeled with an open mapping knowledge, does not require manual tagging one by one, is a weak supervised relation extraction.

 

Open domain relation extraction method:

No need to define the relationship between classes, some of the words used entity context to describe the relationship between entities.

 

to sum up:

Qualified domain relation extraction is the main direction of research.

The traditional template-based extraction method relationship scalability poor relation extraction method based on machine learning is the hot topic at present.

Extraction requires manual annotation based on the relationship of supervised learning a lot of training data, time-consuming, weakly supervised learning based on the relationship extraction gaining more and more attention.

 

Event extraction

 

Definitions : description of the event information extracted from the text the user is interested in the event and presented in a structured form.

Steps: First, identify the type of event and, secondly to identify the elements involved in the incident (generally entity), the final element needed to determine the role each played in the incident.

 

Extract the relevant event concept:

• Event allegations: natural language description of the specific events that occur conduct an objective, usually a sentence or group of sentences
• Event Trigger words: word refers to an event that best represents the alleged incident, the decision is an important feature event category, usually a verb or a noun
• Event elements: event participants, mainly by the entity and attribute values ​​make up time
• The role of elements: What is the role in the event element corresponding event
• Event Category: event elements and trigger word determines the category of the event (defined in a number of categories and sub-categories)

 

Event field defining extraction: Before performing the extraction, the type of event and a good target specific configuration of each type (which contain the specific event element) defined in advance, will generally give a certain amount of data are denoted.

Event field defining extraction method:

• The method of pattern matching based: on some type of event identification and extraction is carried out under the direction of some of the modes (steps of: acquiring mode, pattern matching)
  • Supervised event pattern matching: Get corpus based entirely on manual annotation pattern
  • Weak oversight event pattern matching: corpus does not need to fully dimensioned, only artificial corpus of certain pre-sorting mode or the development of a small amount of seed
• Method based on machine learning
  • Supervised event extraction method: the extraction modeled as a multi-event classification
    • Project-based method features: the need to display the event instance to convert into a classifier acceptable feature vectors, research is focused on how to extract distinguishing features have sexual
    • Based on Neural Network: automatically acquired from the text feature extraction and then complete the event, avoiding the use of a conventional natural language processing problem of accumulation of errors caused by the tool
  • Weak oversight event extraction method: no manual tagging a large number of samples, but needs to be given with a labeled training data specification semantic tags (event category, role name, etc.)
    • Based on Bootstrapping event extraction: automatic generation of large-scale tagging data using the small number of manually labeled data (high confidence results will be drawn as training samples, and then training, continuous iteration)
    • Based Distant Supervison event extraction: fully automatic generation of events stated in the catalog, using a structured knowledge base of the event back to standard training samples directly in the unstructured text

 

Event open field extraction: before the event recognition is performed, the possible event type and structure of the event is unknown, so the task is usually not marked data, based mainly unsupervised and distribution hypothesis.

Distribution hypothesis: if the candidate or the candidate event trigger word event elements have a similar context, then these candidate events trigger words tend to trigger the same type of event, candidate events corresponding elements tend to play the same event elements.

Open field event extraction method:

• Event-based content feature extraction method
• Event extraction method based anomaly detection

 

Event relation extraction to BSU event, event logic implemented deep extraction and detection, comprising:

• Event co-reference relation extraction
• Event causal relation extraction
• Sub-event relation extraction
• Event timing relation extraction

 

to sum up:

Organize and build contains both entities, entity relationships, event mapping knowledge event, the event relation has been more and more attention.