Relationship extraction -- casrel

Relationship extraction –CASREL

Relation extraction is a basic task in naturallanguageprocessing . Relation extraction usually uses triples (subject, relation, object) Express . There are two ways to solve relationship extraction ：
（1） Two entities are known subject and object, Use the classification model to get the relationship between entities
（2） Extract entities , Predict possible relationships between entities . If you first extract entities and then predict relationships , This is called pipline The formula extraction ; If we extract entities and relationships between entities at the same time , This method is called joint extraction .
The data set of relation extraction is complex , Entities and relationships in the dataset overlap , An ideal data set for relational extraction subject and object Corresponding to a relationship , Real data sets subject and object Corresponding to a variety of relationships , Entities overlap . Here's the picture ：

among EPO Indicates that the entity is repeated ,SPO Represents a single entity repetition .
When relational triples (subject, relation, object) When overlapping , Relational classification models are difficult to handle overlapping data . If there are not enough training examples , It is difficult for classifiers to tell which relationship entities participate in , Extracted triples are usually incomplete and inaccurate . However CASREL The model can effectively deal with overlapping relational triples .

CASREL Model

novel cascade binary tagging framework(CASREL) The model is A Novel Cascade Binary Tagging Framework for Relational Triple Extraction Proposed ,CASREL The model has been refreshed SOTA Result .CASREL The model is divided into two steps ：
（1） Through pre training BERT The model gets all the possible subject
（2） For each subject, We apply a relationship specific marker to identify all possible relationships and corresponding object.
The purpose of extracting relational triples is to identify all possible （subject,relation,object）, Some of these relationships may be the same as sharing subject or object Entity . therefore CASREL The objective function of is expressed as ：

CASREL The model structure is shown below

Subject Tagger

Subject Tagger The model in is decoded directly N layer BERT The encoding vector generated by the encoder hN To identify all possible in the input sentence subject, In fact, two identical two classifiers are used (0/1) To mark subject Start and end of , The formula is as follows ：

In a given sentence subject The maximum likelihood function of is ：

Relation-specific Object Taggers

Relation-specific object taggers Considering subject Characteristics of , Instead of directly decoding pre training bert Model HN,relation-specific object taggers The formula is as follows ：

Relation-specific object taggers The maximum likelihood function of is

Chinese experimental results

On Chinese data and English data sets tokenizer The treatment is basically the same , Every Chinese character is followed by [unused1], use chinese_L-12_H-768_A-12 Pre training model . The format of Chinese dataset processing is as follows ：

{
   "text": " How to play your part well , Please read 《 Self-cultivation of actors 》《 The king of comedy 》 The unique secret collection of Stephen Chow rising from poverty ",
   "triple_list": [
     [
       " The king of comedy ",
       " starring ",
       " Stephen Chow "
     ]
   ]
 }

The model parameters are as follows ：
max_length=128, batch_size=16, lr=1e-5, epoch=16
The evaluation results of the model are as follows ：

f1: 0.7827, precision: 0.7736, recall: 0.7921, best f1: 0.7944

The prediction results of the model are as follows ：

{
    "text": "《 The magic show of love 》 It's a song sung by Anxia , Written by wuyiwei ,MartinHansen/StefanDouglasHayOsson Composing music , Included on album 《 Single best 》 in ",
    "triple_list_gold": [
        {
            "subject": " The magic show of love ",
            "relation": " The album ",
            "object": " Single best "
        },
        {
            "subject": " The magic show of love ",
            "relation": " singer ",
            "object": " Anxia "
        }
    ],
    "triple_list_pred": [
        {
            "subject": " The magic show of love ",
            "relation": " The album ",
            "object": " Single best "
        },
        {
            "subject": " The magic show of love ",
            "relation": " singer ",
            "object": " Anxia "
        },
        {
            "subject": " The magic show of love ",
            "relation": " Lyrics ",
            "object": " Wuyiwei "
        }
    ],
    "new": [
        {
            "subject": " The magic show of love ",
            "relation": " Lyrics ",
            "object": " Wuyiwei "
        }
    ]
}

If there is an error , Welcome to correct .