Scientific information extraction is a crucial step for understanding scientific publications. In this paper, we focus on scientific keyphrase extraction, which aims to identify keyphrases from scientific articles and classify them into predefined categories. We present a neural network based approach for this task, which employs the bidirectional long short-memory (LSTM) to represent the sentences in the article. On top of the bidirectional LSTM layer in our neural model, conditional random field (CRF) is used to predict the label sequence for the whole sentence. Considering the expensive annotated data for supervised learning methods, we introduce self-training method into our neural model to leverage the unlabeled articles. Experimental results on the ScienceIE corpus and ACL keyphrase corpus show that our neural model achieves promising performance without any hand-designed features and external knowledge resources. Furthermore, it efficiently incorporates the unlabeled data and achieve competitive performance compared with previous state-of-the-art systems.

Keyphrases provide core information for users to understand the document. Most previous works utilize machine learning based methods for keyphrases extraction and achieve promising performance. However, these methods focus on identify keyphrases from the input text, and can not extract keyphrases that do not appear in the text. In this paper, we present an encoder-decoder framework, which incorporating copying mechanism, to generate keyphrases for the given text. This framework (CopyNet) integrates the generation part and copying part. The generation part generates the keyphrase from the predefined vocabulary, and the copy part gets the keyphrases from the source text. Furthermore, we improve the CopyNet by using different probability of the two parts. To incorporate more related information for keyphrase generation, the automatically built keyphrase semantic web is merged into the dataset to participate in the training process of the neural network. Semantic similarity based and word co-occurrence based methods are used for keyphrase semantic web construction. We build a large-scale biomedical keyphrase dataset to evaluate the system performance. Experiments show that our improved CopyNet can achieve better performance with different portions of the generation and copying part, and the incorporation of the semantic web also effectively improves the keyphrase generation.

Named entity recognition (NER) is one fundamental task in natural language processing, which is usually viewed as a sequence labeling problem and typically addressed by neural conditional random field (CRF) models, such as BiLSTM-CRF. Intuitively, the entity types contain rich semantic information and the entity type sequence in a sentence can globally reflect the sentence-level semantics. However, most previous works recognize named entities based on the feature representation of each token in the input sentence, and the token-level features cannot capture the global-entity-type-related semantic information in the sentence. In this paper, we propose a joint model to exploit the global-type-related semantic information for NER. Concretely, we introduce a new auxiliary task, namely sentence-level entity type sequence prediction (TSP), to supervise and constrain the global feature representation learning process. Furthermore, a multitask learning method is used to integrate the global-type-related semantic information into the NER model. Experiments on the four datasets in different languages and domains show that our final model is highly effective, consistently outperforming the BiLSTM-CRF baseline and leading to competitive results on all datasets.