Research And Implementation On Natural Language Understanding Oriented New Slot Value Recognition

Natural language understanding(NLU)is an important component of task-oriented dialogue system,including domain recognition,intent recognition and semantic labeling.As the main task in NLU,semantic labeling aims to recognize the slots and corresponding values,so it is always named as slot filling.In lots of real-world applications,NLU not only needs to recognize the slot value by semantic labeling algorithm,but it also needs to map the slot value into one of the labels in a predefined(or pre-built based on training dataset)slot value list.However,a predefined list hardly includes all possible values so that in real conversations there may be new slot values inot in the predefined list or training dataset.The ability to recognize new slot values efficiently is important for the robustness of NLU and the expandability of dialogue system.But traditional methods based on sequence labeling or classification can not recognize new slot values efficiently.To address the problem,this thesis has fully investigated related researches and does some works taking the real demands into consideration.More specifically,contents in this thesis include:Propose an attention based joint semantic labeling model and a negative-sampling based training method.The model combines a sequence tagger and a classifier.The tagger locates the slot value and the classifier obtains the label of standard slot value or the new slot value.The training method based on negative sampling enables the model with supervision by constructing negative samples.Experimental results on two datasets show that training with negative samples boosts the performance of new slot value recognition greatly and the attention mechanism discovers important information automatically and thus improve the performance.Study on the extension of joint model and try different approaches to model internal relations in multi-word values,including concatenating features,improving attention mechanism and constructing structured representation of utterances.Also,this thesis explores methods in multi-slots condition.Experimental results prove that the model can be adapted to multi-slots applications and recognizes new slot values in each slot.Realize a complete NLU module and apply it into a cell-phone service dialogue system of an enterprise based on the technique described above.Functional testing shows that the module successes in extracting standard slot values new slot values and reveals good performance.