PolyakovPYuKalininaMVPleshkoVV (Аннотации)

Автоматическое определениетональности объектовс использованием семантическихшаблонов и словарейтональной лексикиПоляков П. Ю. (pavel@rco.ru),Калинина М. В. (kalinina_m@rco.ru),Плешко В. В. (vp@rco.ru)ООО «ЭР СИ О», Москва, РоссияКлючевые слова: определение тональности, анализ мнений, тональность объектов, тональность атрибутов, синтактико-семантическийанализ, семантические шаблоныPolyakov P. Yu., Kalinina M. V., Pleshko V. V. 1. IntroductionThe task of automatic sentiment analysis of natural language texts has becomeextremely in demand.

Many commercial companies producing goods and servicesare interested in monitoring social networking websites and blogs for users’ opinionsabout their products and services. However, until recently there were no tagged textcorpora in Russian on which developers could test and compare quality of their methods.

This gap was filled by ROMIP and later SentiRuEval sentiment analysis evaluationconferences with their sentiment analysis tracks. However, the task of the previousconferences was to detect general sentiment of a text (for example, see Chetviorkin I.,Braslavski P. I., Loukachevitch N.

[2]), while at the present conference the task wasbrand new—object-oriented sentiment analysis, which is more difficult and requiresmore sophisticated algorithms; for, in case of general sentiment detection, selectionof positive and negative terms and defining of their weights are important, while,in case of object-oriented sentiment detection, syntactic relations between a targetobject and a word expressing sentiment are also of great importance.Such object-oriented method is not new for us; we have already used similar approach in our previous research. For instance, we evaluated sentiment-oriented opinions in regard to car makes on the material of the LiveJournal blog AUTO_RU (see description of the method in Ermakov A. E.

[4]). It should be mentioned, however, thatin all the previous cases results had only been evaluated by ourselves. Participationin SentiRuEval gave us a chance to have an independent evaluation of our method andcompare our results with other participants’.In this paper we present results of applying a linguistics-based approach involving syntactic and semantic analysis to the task of automatic object-oriented sentimentanalysis. We confined ourselves to a linguistic method only, having excluded machinelearning, because it was interesting to see what results a pure linguistic approachwithout machine learning methods would provide.The task was to find sentiment-oriented opinions (positive and negative) abouttelecom companies in tweets.2. Related WorkUsually object-oriented or aspect-oriented approaches either rely only on statistics-based algorithms, word distance count, machine learning, etc.

to find opiniontargets (starting with the first work on opinion target extraction by Hu and Liu [5]);or they may use shallow parsing to segment a sentence, find significant conjunctions,negations, and modifiers (ex., Kan D. [7]). Other approaches are looking for syntacticdependency between a sentiment term and its target (ex., Popescu A., Etzioni O. [9]),ignoring sentiment-bearing words which are not syntactically related to any targetobject. The distinctive feature of our approach is that using a deep linguistic methodwe take into account not only syntactically related sentiment terms (which provideshigh precision) but also independent sentiment-bearing words and phrases (whichprovides high recall).Automatic Object-oriented Sentiment AnalysisSome researchers try combine statistical and linguistic methods in order to achievethe best results; for example, in Jakob N., Gurevych I. [6] authors use, among other, thedependency parse tree to link opinion expressions and the corresponding targets; andthe experiments show that adding the dependency path based feature yields significant improvement to their method.

However, their algorithm is searching for short anddirect dependency relations only; therefore, their approach has difficulties with morecomplex sentences. Furthermore, they do not distinguish between a target object (ex.,camera), its attributes or parts (ex., lens cap, strap), and its qualities (ex., usability); and,hence, they label the closest noun phrase as a target of the opinion. In contrast, we usea very basic ontology to distinguish between a target object, attributes, and qualities;and having found a sentiment related to an attribute or quality our algorithm goesdown the dependency parse tree searching for a target object. If not found syntactically,the target object is being searched for by a heuristic, based on the clause distance.

Whenthe target object is found, the sentiment labeled to its attribute is assigned to the object.3. MethodsTo perform the task we based on our previous researches and solutions. Detaileddescription of these methods can be found in Ermakov A. E., Pleshko V. V. [3] and Ermakov A. E. [4]. New to the approaches described in [3] and [4] was adding so-called‘Free Sentiment Detection’, which will be described in Section 3.2.The text analysis algorithm has the following stages in regard to the sentimentdetection task:1)Tokenization;2) Morphological analysis;3) Object extraction;4) Syntactic analysis;5) Fact extraction (use of semantic templates);6) Free sentiment detection.Stages 1, 2, and 4 were implemented by standard RCO tools for general textanalysis.

At stage 3 we paid more attention to the objects concerning the given subject(names of mobile companies, telecom terminology, etc.). Stages 5 and 6 were coreto the sentiment detection task and, therefore, will be described in detail.3.1.Semantic TemplatesThe main method of sentiment analysis involved usage of semantic templates.Semantic template is a directed graph representing a fragment of a syntactic treewith certain restrictions applied to its nodes.

The syntactic tree of a sentence containssemantic and syntactic relations between words, which are defined by the syntacticparser. The restrictions in the templates can be applied to a part of speech, name,semantic type, syntactic relations, morphological forms, etc. Fact extraction is performed by finding a subgraph in the syntactic tree of a sentence which is isomorphicto the template (with all restrictions applied).Polyakov P. Yu., Kalinina M.

V., Pleshko V. V. RCO syntactic analyzer, based on the dependency tree approach, has been used.The semantic network built by the syntactic parser is invariant to the word order andvoice; for example, sentences (1) Оператор украл деньги со счета and (2) Деньгиукрадены оператором со счета will have the same semantic net. Such semantic network constitutes an intermediate representation level between the semantic schemeof a situation and its verbal expression, that is, a deep-syntactic representation, abstracted from the surface syntax.Settings of the semantic interpreter allow filtering negative and ‘unreal’ (imperative, conditional, etc.) statements, which don’t correspond to real events and shouldnot be analyzed.

As a result, examples like (3) если Билайн будет плохо работать;сеть якобы падает; связь бы обрывалась; не Билайн плохо работает can be excluded from the sentiment detection.To decrease the number of templates describing semantic frames, we have socalled auxiliary templates, which add new nodes and relations into the semantic network. In the process of semantic analysis and fact extraction auxiliary templates workbefore all other templates, so that semantic templates can base on the net built by boththe syntactic analyzer and the auxiliary templates. For example, if we interpret phraseslike (4) Х does Y, X begins to do Y, and (5) X decides to do Y as equal for a particular semantic frame, instead of creating a semantic template for each example we can haveone auxiliary template, which will mark the subject of the main verb as the subjectof the subordinate verb, and one simple semantic template—(4) X does Y.Semantic templates can have so-called ‘forbidding nodes’ which impose restrictions on the context, defining in which context the template should not match.