Research on Optimization Methods of groove filling mode

Research on the Mode Optimization Method in slot filling (Master's thesis) Shen Xiaowei's Implementation Method for slot filling tasks

There are three main types:

The first is to convert the slot filling task into a link extraction task. The traditional information extraction method is used to implement the slot filling task, information extraction methods can be divided into pattern matching and classifier-based methods.

The second is based on the Q & A System (question answering, QA). The slot filling task is completed by parsing each slot defined in the slot filling task into an equivalent question set.

The last is the rule-based method, that is, the corresponding Pattern Library or rule repository is built for each slot directly using the manual method, then, based on these modes or rules, the test corpus is retrieved to meet the standard instance to complete the slot filling task.

Among them, the first method: Information Extraction Method occupies the majority.

Common tool kits

Stanford natural language processing tool, WordNet.

Http://wordnet.princeton.edu/wordnet/

Link Extraction

The implementation of link extraction tasks can be divided into rule-based and machine learning-based methods.

The former sets some lexical, syntactic, and semantic pattern rules manually based on the characteristics of the relationships to be extracted, and then searches for matching relational instances from free texts;

The latter adopts the machine learning method, especially the statistical machine learning method. First, a relational extraction model is obtained through corpus tagging training, and then the model is used to identify the relational instances in free text.

Usage and performance:

Stanford (classifier-based Information Extraction Method (distant supervision) uses Entity-slot value instance pairs in KB) P = 10.54%, r = 21.41%, F1 = 14.12%

IBM (based on classifier, maximum entropy model distant supervision noise) P = 31.0%, r = 25.9%, F1 = 28.2%

NYU (based on pattern matching, bootstrapping) P = 54.3%, r = 11.6%, F1 = 19.1%

IIRG (the Q & A system converts the slot value query to the question set) P = 66.5%, r = 18.6%, F1 = 29.1%

Usfd (manual rule) P = 3.1%, F = 24.1%, F1 = 5.5%

Dependency syntax

Dependency defined by Stanford

1. The representation of the baseline (basic.
2. Collapsed representation.
3. The merge representation mode of the Extended Parallel Relationship (propagation of conjunct dependencies.

WordNet

WordNet establishes synonym, async ym, holonym, Meronym, hypernym, and hyponym between synonym sets) and Other semantic relationships to connect them into semantic networks.

The focus is on nouns and verbs. Nouns are the main component of ontology.

Training Phase

(1) ing from the knowledge base to the slot. Because the object attributes defined in the knowledge base are not consistent with the slot defined in the slot filling task, you must first map the information of the entity in the knowledge base to the slot type defined in the slot filling task. In system implementation, Stanford's lexical analysis, Named Entity recognition, and other Resolvers are used as tools, ing and conversion of entities in the Knowledge Base Based on the ing relationship between object attribute names provided by kbp to the slot.

(2) refer to replacement and search example sentences. After the first phase of the ing operation, you can obtain a large number of Object-slot value pairs from the knowledge base, based on these object-slot value pairs, you can retrieve example sentences from the free text corresponding to the object. In order to obtain as many example sentences as possible, you need to replace the personal and all lattice Pronouns that appear in the text, for example (co-finger relationship)

(3) extract the dependency path and generate the mode library. After the example sentences are retrieved for each slot, You need to obtain the mode from the example sentence. In this document, the model in the baseline system is constructed based on the dependency path between the object and the slot, dependency paths use 53 dependency relationships from Stanford,

Test phase

(1) first, retrieval of relevant documents;

(2) Extraction of target object slot values to obtain candidate items;

(3) The third step is to select the final answer from these candidates. (WordNet, object type, Domain value)

 

(1) modify the dependency syntax of the Stanford toolkit and (2) generate the pattern filtering method. This is not my concern, so I did not continue reading it.

 

Subsequent papers:

Classification Methods: Stanford, UBC, And NYU (bootstrapping)

QA Methods IIRG and usfd

Integrate multiple methods CUNY

Research on Optimization Methods of groove filling mode