Sentiment analysis of natural language texts needs to deal with linguistic phenomena like negation, intensi cation or adversative clauses. In this talk, I present an approach to tackle such phenomena by means of syntactic information. Our approach combines machine learning and symbolic processing: the former is used to obtain dependency trees for input sentences, and the latter to obtain the sentiment polarity for each sentence using handwritten rules that traverse the tree. Thanks to universal guidelines for syntactic annotation, our approach is applicable to multiple languages without rewriting the rules. Additionally, very accurate parsing is not needed for our approach to be helpful: fast and simple parsers will do, even if they lag behind state-of-the-art accuracy.
Polarity classi cation is a basic task in sentiment analysis of natural language texts, consisting on determining
whether the expressed opinion in a piece of text is positive, negative or neutral. Since the seminal work by Pang
et al. [
This phone is expensive, and not really very good. / This phone is good, and not really very expensive.
What sets them apart is word order, which in turn determines their syntactic structure, i.e., the way in which words interact with each other to form meaningful sentences. Namely, their polarities di er mainly due to the di erent roles played by the negation \not". In the rst example, it modi es the positive word \good", making the phone \not good" and thus expressing a negative view, while in the second example, it modi es the negative word \expensive", making it \not expensive" which is positive. Since syntactic relations like these can happen at arbitrarily long distances, n-grams cannot always capture them { but a syntactic dependency tree can.
Syntax can be incorporated into a sentiment analysis model in various ways: for example, by decomposing
trees into features for a supervised classi er [
Our approach is described in detail in [
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). The sentence is tokenized, part-of-speech tagged and parsed to a syntactic dependency tree. Any o -the-shelf dependency parser can be used for this purpose.
Standard third-party sentiment lexicons are used to assign individual polarities to subjective words (for example, in a scale of -5 to +5, the word \excellent" could be assigned a polarity of +5, and \good" could be assigned +3).
Handwritten rules are used to propagate the polarity in a top-down fashion, from individual words to the root of the tree (which yields the global polarity of the sentence). Rules are written to deal with syntactic phenomena that in uence polarity. For example, given a negation, we use rules on the dependency tree to identify the scope of negation, and we modify the polarity of the negated element by subtracting -4 if it is positive, or +4 if it is negative (e.g. \not good" will be assigned 3 4 = 1). Rules are also written to deal with adversative clauses, intensi ers and conditional statements.
Two conditioning factors need to be taken into account when writing the rules to ensure generalizability of the system across languages: rst, rules are dependent on syntactic annotation criteria (but the appearance of universal annotation criteria that can be applied cross-linguistically, like Universal Dependencies, means that this is not an obstacle for multilinguality). Second, lists of negation words, intensi ers, adversative conjunctions and words introducing conditionals are needed to make the rules generalizable across languages, but these are available with standard sentiment lexicons.
Our experiments show that our approach outperforms existing unsupervised alternatives, as well as supervised
models when evaluated outside of their corpus of origin. Using a high-accuracy parser is not vital [