1、Motivationu Problem Definitionu Civil Law System:we need to decide charges based on law articles被告人被告人吴吴某某以外出打工为名，将贵州省盘县特区马场乡村民吴某诱骗出门诱骗出门，后与王合伦（在逃）、曹某（已判刑）将吴某带到江苏省丹阳市麦溪镇黄庄村以以29002900的价格卖给陈某为妻的价格卖给陈某为妻，三人共同分赃。案发后被告人吴春丽向吴某家人退还人民币600元。罪名：拐卖罪名：拐卖人口罪人口罪相关法条：刑法相关法条：刑法第十二条、第十二条、第二十五条第一款、第六十第二十五条第一款、第六十七条第三
2、款七条第三款Motivationu How do we deal with a criminal case?Collect FactsFind Relevant Law ArticlesDetermine ChargesDetermine Punishments.Obtaining Basic InformationArgumentationMotivationu Challenges with charge predictionu Civil law system:decisions rely on law articlesu Infer from both facts and law ar
3、ticlesu Differences between charges can be subtleu Intentional Injury V.S.Intentional Homicide u One case may also relate to multiple chargesMotivationu Challenges with article findingu we need to infer from both facts and article descriptionu one case may relate to multiple law articlesu we have ma
4、ny law articlesu 452 articles in Criminal Law of ChinaDatau Good News!u We have publicly available data,by government!u Millions of judgement documents on http:/ cl eExtractorf act sar t i cl e1ar t i cl e2ar t i cl ek Docum ent EncoderfArti cl eExtractorf act sar t i cl e1ar t i cl e2ar t i cl ek d
5、fufwufsMethodu Find Relevant Articles u Facts embedding needed for both charge and article prediction Methodu Document Modelingu Both fact description and law articles are documentsu Documents:sequence of sentencesu Sentence:sequence of wordsDocument Encoderusuwsent1s1Sequence Encoderdsent2sentnSequ
6、ence EncodersSequence EncodersSequence Encoderss2sndMethodu Document Modelingu Both fact description and law articles are documentsu Bi-GRUu(Guided)AttentionAttentive Sequence EncoderDocum ent EncoderfArti cl eExtractorf act sar t i cl e1ar t i cl e2ar t i cl ekDocum ent EncoderaDocum ent EncoderaDo
7、cum ent Encodera a1a2a3uawuasdfufwufsMethodu Find Relevant Articles u Article embedding deeply understandingDocum ent EncoderfArti cl eExtractorf act sar t i cl e1ar t i cl e2ar t i cl ekdaDocum ent EncoderaDocum ent EncoderaDocum ent Encodera a1a2a3uawuasArti cl e AggregatoruaddfufwufsMethodu Find
8、Relevant Articles u Article attention Article selection output relevant articlesDocum ent EncoderfArti cl eExtractorf act sar t i cl e1ar t i cl e2ar t i cl ekdaDocum ent EncoderaDocum ent EncoderaDocum ent Encodera a1a2a3uawuasArti cl e AggregatoruaddfSoftm axufwufsMethodu Charge Predictionu Combin
9、e both facts and articles,thresholding for multi-chargesExperimentsu Charge Predictionu SVM is a strong baseline,but can utilize retrieved relevant articles u Using articles can improve charge predictionExperimentsu Trade-off Between Charge and Articleu More weights on article can improve article pe
10、rformanceu High weights on article will harm charge predictionExperimentsu Experiments on News Datau Can model trained on judgement documents work for non-legal professionals?YES!Conclusionu One model charge prediction+article findingu Using law articles can help charge predictionu Model trained on judgement documents can generalize to fact descriptions written by non-legal professionals
