論文鏈接:http://pdfs.semanticscholar.org/893a/9ea38da739af53d4cb8ec5d0e722b0e6c6e4.pdf
本文的任務(wù):extract aspect and opinion terms/phrases for each review
In aspect-based sentiment analysis, the core component is to extract aspects or features of a product/service from a review, along with the opinions being expressed. aspect-based sentiment analysis任務(wù)是做什么
過去的方法主要分為兩類:
第一種:從一個(gè)seed集合共苛,使用句法規(guī)則和aspect及opinion之間的關(guān)聯(lián)來積累aspect terms和opinion terms匀钧。但是這種方法很依賴與手動(dòng)定義的規(guī)則,并且嚴(yán)格遵循特定的詞性規(guī)則姿骏,例如opinion詞是形容詞。
第二種:sequence labeling classifier,例如CRFs和HMMs,使用feature engineering芯丧,詞典和有標(biāo)注的數(shù)據(jù)集。This approach requires extensive efforts for designing hand-crafted features, and only combines features linearly when a CRF/HMM is applied
使用深度學(xué)習(xí)進(jìn)行情感分析的方法分為兩類:一類是句子級(jí)別的情感預(yù)測世曾,一類是phrase/word-level情感預(yù)測。
本文的方法:
包括兩部分:基于每句話的依存樹構(gòu)造的recursive neural network,為了學(xué)習(xí)到句子中每個(gè)詞的上下文的high-level representation轮听,輸出會(huì)送到Conditional random field(CRF)學(xué)習(xí)從high-level特征到標(biāo)簽的映射骗露。Because CRFs have proven to be promising for this kind of sequence tagging problems.
與本文方法類似的是:【1】,使用標(biāo)準(zhǔn)的recurrent neural network血巍,很依賴于word embeddings的質(zhì)量萧锉,除此之外,沒有考慮句子結(jié)構(gòu)中的依存關(guān)系述寡。
The tree structure used for RNNs generally adopts two forms: constituency tree and dependency tree柿隙。在constituency tree中,所有的詞都在葉節(jié)點(diǎn)中鲫凶,每個(gè)內(nèi)部節(jié)點(diǎn)表示一個(gè)短語或句子的一部分禀崖,跟節(jié)點(diǎn)代表整個(gè)句子;dependency樹螟炫,每個(gè)節(jié)點(diǎn)代表一個(gè)詞波附,用依存關(guān)系與其他節(jié)點(diǎn)關(guān)聯(lián)。
每個(gè)節(jié)點(diǎn)n昼钻,是與一個(gè)詞w相關(guān)聯(lián)的掸屡,每個(gè)依存關(guān)系r與一個(gè)矩陣關(guān)聯(lián)
先計(jì)算葉子結(jié)點(diǎn)的隱狀態(tài):
CRF:
In a linear-chain CRF, which is empolyed in this paper, there are two different cliques: unary clique(U) representing input-output connection, pairwise clique(P) representing adjacent output connection.
【1】Fine-grained Opinion Mining with Recurrent Neural Networks and Word Embeddings仅财,論文鏈接:paper