Aspect-based Sentiment Analysis
Aspect-based Sentiment Analysis
Aspect-based Sentiment Analysis
通常的情感分析只关注情感极性，而Aspect-based Sentiment Analysis是情感分析的精细化扩展，关注针对事物不同具体维度的情感分析。
With the proliferation of user-generated content, interest in mining sentiment and opinions in text has grown rapidly, both in academia and business. The majority of current approaches, however, attempt to detect the overall polarity of a sentence, paragraph, or text span, irrespective of the entities mentioned (e.g., laptops, battery, screen) and their attributes (e.g. price, design, quality).
Aspect-based Sentiment Analysis System
Aspect-based Sentiment Analysis的一种典型方法参见论文。
http://www.aclweb.org/anthology/P15-4010
系统的输入是中文字符序列，输出是事物的维度和对应的情感极性。
The input to the pipeline is a string of Chinese characters; the output is a set of relationships between evaluations and their targets. The main goal is to demonstrate how knowledge about sentence structure can increase the precision, insight value and granularity of the output.
分析过程主要包括两个步骤：单元识别（unit identification）和关系抽取（relation extraction）。
Formulate the task of sentiment analysis in two steps, namely unit identification and relation extraction.
The presented model is complemented by a probabilistic model which performs topic and polarity classification on the sentence and the document levels.
算法的核心思想是中文的表达通常具有一定的模式（例如主谓宾或主系表结构），因此可以按照词/词性的序列模式中进行抽取。
The basic assumption on which the model builds is that language follows rules. Chinese phrase structure is largely head-final (Huang 1982, Li 1990, i. a.): nominal modifiers precede their head nouns, whereas degree and negation adverbs normally precede the adjectives or verbs they modify. A small set of correspond- ing phrase-level rules achieves a high coverage.
The goal of aspect-based sentiment analysis is to derive the opinions of a speaker about an entity and its features.
系统的输入和输出如下例所示。
Given an opinion statement on a specific product, translate the statement into a set of (f eature, < evaluation|emotion > ) pairs in two processing steps:
通常情感极性对应的判断词包含了维度信息（例如"好看"隐含了外观维度），因此需要对隐含维度进行补足。
It has long been observed that many evaluation words come with implicit features; for example, the evaluation beautiful implicitly contains the feature VisualAppearance. In order to preserve this meaning, we adopt a scalar representation of evaluations (cf. Kennedy and McNally (2005) for a linguistic analysis of scalar expressions): evaluations are represented as pairs of a feature and a numerical value which "maps" the evaluation to some point on the feature scale [-3, 3].
The final mapping goes from sequences of features to numerical evaluations.
系统的整体算法过程如下。
Lexical basis
Semantic categories are relevant for the interpretation of opinions. The top-level semantic categories are:
Processing steps
Use a phrase rule grammar which is based on regular expressions involving the POS tags of the words
In the following, present some of the most common phrase structures for features and evalu- ations/emotions.
Feature phrases
Evaluation and emotion chunks
Relation extraction
The causer relation. The causer relation is a fairly well-delimited relation which describes the causer of some state of event.
The theme relation. The theme relation is expressed differently for evaluations and emotions.
SemEval-2015 Task 12
SemEval每年会提出10-20个和NLP相关的任务，以比赛的形式展开。
SemEval包含范围有一定广度，其中包括了Aspect-based Sentiment Analysis问题、
SemEval-2015 Task 12: Aspect Based Sentiment Analysis
SemEval中针对Aspect-based Sentiment Analysis问题的描述如下。
In aspect-based sentiment analysis (ABSA) the aim is to identify the aspects of entities and the sentiment expressed for each aspect. The ultimate goal is to be able to generate summaries listing all the aspects and their overall polarity such as the example shown in Fig.
In addition, SE-ABSA15 will include an out-of-domain ABSA subtask, involving test data from a domain unknown to the participants, other than the domains that will be considered during training. In particular, SE-ABSA15 consists of the following two subtasks.
SemEval中Aspect-based Sentiment Analysis包含两个子问题。
Subtask 1: In-domain ABSA
Given a review text about a laptop or a restaurant, identify the following types of information:
- Slot 1: Aspect Category (Entity and Attribute). Identify every entity E and attribute A pair E#A towards which an opinion is expressed in the given text.
- Slot 2 (Only for the restaurants domain): Opinion Target Expression (OTE). An opinion target expression (OTE) is an expression used in the given text to refer to the reviewed entity E of a pair E#A.
- Slot 3: Sentiment Polarity. Each identified E#A pair of the given text has to be assigned a polarity, from a set P = {positive, negative, neutral}.
Subtask 2: Out-of-domain ABSA
The input for the participating systems will be reviews from a particular domain (e.g., laptop or restaurant reviews). Each system will be required to construct all the {E#A, P} tuples for the laptops domain and the {E#A, OTE, P} tuples for the restaurants domain.
Datasets
SemEval中Aspect-based Sentiment Analysis问题所采用的数据集如下。
Two datasets of ~550 reviews of laptops and restaurants annotated with opinion tuples (as shown in Fig. 2) will be provided for training. Additional datasets will be provided to evaluate the participating systems in Subtask 1 (in-domain ABSA). Information about the domain adaptation dataset of Subtask 2 (out-of-domain ABSA) will be provided in advance.
相关连接
- http://www.aclweb.org/anthology/P15-4010
- http://alt.qcri.org/semeval2015/task12/
- http://alt.qcri.org/semeval2016/task5/
- https://www.cs.uic.edu/~liub/FBS/SentimentAnalysis-and-OpinionMining.pdf
