| Game,as an intellectual sport,has been deeply loved by people.There are two types of games: complete information games and incomplete information games..Compared with complete information game,incomplete information game hides part of scene information,making the whole game process full of interesting.In daily life,we commonly incomplete information game including mahjong,bridge and so on,attracting a large number of players.At present,most people acquire incomplete information game knowledge by reading books or communicating on forums,but these methods cannot accurately and quickly return the required knowledge to the player.In addition,the dialogue corpus in the information field of incomplete game has the problem of missing key information and serious colloquial language,which greatly increases the difficulty of human-computer interaction in the information field of incomplete game.Based on Rasa framework,The research in this paper focuses on multi-round dialogues within the incomplete game domain of information field.The completed work can be categorized into the following parts:1.A map of knowledge has been created for the domain of multi-region incomplete information games.Because the corpus in the imperfect game field has the characteristics of short sentences and limited effective information,it causes the problem of missing key information in the process of model training.In order to solve this problem,this paper firstly analyzes the characteristics of the original corpus,then makes corresponding preprocessing according to different characteristics,and then carries out iterative screening manually to remove invalid corpus.Finally,BIO was used for annotation,and data was stored through Neo4 j graph database.2.Ro BERTa-based model for named entity recognition in incomplete information games.Since the Rasa framework is not very satisfactory in dealing with the task of natural language understanding in the incomplete game domain,the original components of Rasa are optimized according to the problems in the research process.Among them,Enhancing the CRF model is a crucial component in optimizing the named entity recognition task,as stated in this paper,introduced the pre-training model and bidirectional long and short memory network,and added the dynamic mask strategy and L2 loss function,and finally designed the Ro BERTa-BiLSTM-CRF joint model.Then,the entity recognition model is compared from different pretraining models and different downstream structural models.The results of the experiment indicate that the combined model proposed achieved an F1 score of 89.07%,which is5.54% and 1.55% higher than that of BiLSTM-CRF and BERT-BiLSTM-CRF,respectively.3.Proposed is a model based on domain map for recognizing intention and filling slots jointly.Due to the characteristics of incomplete information game domain corpus key information missing.In this paper,the semantic information of key entities in the domain map is embedded into the joint model,and the intention recognition and slot filling are compared from three perspectives: whether the pre-training model is included,whether external information is introduced,and whether it is a single task model.The experimental results show that the F1 value of this model is further increased by 1.55% and 0.88% compared with that of BERT and K-BERT.4.A GRU-based conversational management model with an attention mechanism was designed to accommodate the peculiarities of incomplete information game domain dialogue,which often features short sentences and missing semantic information,this paper chooses a variant model GRU with lower complexity than LSTM model used by Keras Policy component in Rasa Core,and adds an attention layer to its benchmark.The model’s attention to key semantic information is strengthened.The experimental results show that compared with LSTM,F1 value of this model is increased by 6.11%,MLP by 1.81% and GRU by 1.31%. |
PDF Full Text Request