Abstract | ||
---|---|---|
Text sentiment transfer models modify sentence sentiments while retaining their semantic content. The main challenge is to separate sentiment-independent content information from the semantic information of the sentence. The previous works usually expect to utilize the model encoder to infer the sentence-level representation that removed sentiment information. However, the strength of the models’ abilities to reconstruct text is difficult to control which resulting encoder infer sentence-level sentiment-independent content embedding failed. In this paper, we address this challenge by using word-level representation. We first use the POS-Tagging technique to tag the part of speech of word sequence, then extracting content keywords by three schemes and use them as input, rather than the entire sentence, to obtain a purer word-level sentiment-independent content representation. In this way, the model does not require to infer the sentiment-independent representation, which avoids the instability of the adversarial training process. Experiments show that our method achieves the state-of-the-art performance and is also effective in long text sentiment transfer tasks. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1007/978-3-030-55789-8_4 | IEA/AIE |
DocType | Citations | PageRank |
Conference | 0 | 0.34 |
References | Authors | |
0 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Shengwei Hu | 1 | 0 | 0.34 |
Bicheng Li | 2 | 0 | 0.34 |
Kongjie Lin | 3 | 0 | 0.34 |
Rui Wang | 4 | 8 | 5.36 |
Kai Liu | 5 | 0 | 0.34 |