In this thesis, we focus on Answer Sentence Selection (A2S) that is the core task of retrieval based question answering. A2S consists of selecting the sentences that answer user queries from a collection of documents retrieved by a search engine. Over more than two decades, several solutions based on machine learning have been proposed to solve this task, starting from simple approaches based on manual feature engineering to more complex Structural Tree Kernels models, and recently Neural Network architectures.
In particular, the latter requires little human effort as they can automatically extract relevant features from plain text. The development of neural architectures brought improvements in many areas of A2S, reaching unprecedented results. They substantially increase accuracy on almost all benchmark datasets for A2S. However, this has come with the cost of a huge increase in the number of parameters and computational costs of the models. A large number of parameters has led to two drawbacks. The model requires a massive amount of data to train effectively, and huge computational power to maintain an acceptable transaction per second in a production environment. Current state-of-the-art techniques for A2S use huge Transformer architectures, having up to 340 million parameters, pre-trained on a massive amount of data, e.g., BERT. The latter and related models in the same family, such as RoBERTa, are general architectures, i.e., they can be applied to many tasks of NLP without any architectural change.
In contrast to the trend above, we focus on specialized architectures for A2S that can effectively encode the local structure of the question and answer candidate and global information, i.e., the structure of the task and the context in which the answer candidate appears.
In particular, we propose solutions to effectively encode both the local and the global structure of A2S in efficient neural network models. (i) We encode syntactic information in a fast CNN architecture exploiting the capabilities of Structural Tree Kernel to encode the syntactic structure. (ii) We propose an efficient model that can use semantic relational information between question and answer candidates by pretraining word representations on a relational knowledge base. (iii) This efficient approach is further extended to encode each answer candidate's contextual information, encoding all answer candidates in the original context. Lastly, (iv) we propose a solution to encode task-specific structure that is available, for example, available on the community Question Answering task.
The final model, which encodes different aspects of the task, achieves state-of-the-art performance on A2S compared with other efficient architectures. The proposed model is more efficient than attention based architectures and outperforms BERT by two orders of magnitude in terms of transaction per second during training and testing, i.e., it processes 700 questions per second compared to 6 questions per second for BERT when training on a single GPU.
| Identifer | oai:union.ndltd.org:unitn.it/oai:iris.unitn.it:11572/275116 |
| Date | 25 September 2020 |
| Creators | Bonadiman, Daniele |
| Contributors | Bonadiman, Daniele, Moschitti, Alessandro |
| Publisher | Università degli studi di Trento, place:Trento |
| Source Sets | Università di Trento |
| Language | English |
| Detected Language | English |
| Type | info:eu-repo/semantics/doctoralThesis |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | firstpage:9, lastpage:112, numberofpages:104 |
Page generated in 0.0024 seconds