Sentiment analysis often relies on a semantic orientation lexicon of positive and negative words. Determining the semantic orientation of words is necessary for correct estimation of the content of statements in the media, Internet, in the writings and speech. Qualitative adverbs expressing evaluation, intensity, direction of action are important as the modifiers of the main sentence predicate. In this paper we propose a method for extracting seed set of adverbs from a collection of pairs of antonym. A model based on the representation of a set of synonyms from the Russian lexicons as a graph, and determination the semantic orientation of the adverbs concerning three main dimensions of the semantic differential also demonstrated. The assessment of performance of the method in comparison with the dictionary data shows effectiveness of the method obtained.
Nowadays, the availability of resources for Natural Language Processing (NLP)
remains a hot topic, in particular for Russian especially due to the lack of
comprehensive semantic resources, despite efforts made to provide a freely-available Russian
WordNet [
The aim of this paper is to extract a list of semantically oriented adverbs and develop the measure of proximity based on dictionaries of synonyms. The article is structured as follows. In Section 1 the problem of extracting the seed set of semantically oriented adverbs from the lexicon of Russian antonyms is discussed. In Section 2 we describe the previously proposed measures of semantic distance between words, as well as an elementary way to map synonyms onto a graph. In Section 3 the basic characteristics of the subjective understanding of the meaning and the measures based on the distance in a graph of synonyms are discussed. Finally, Section 4 presents some results and conclusions. Additionally, we explore the use of visualization techniques to gain insight into the results obtained. 2
A number of approaches have been proposed for creating semantic orientation
lexicons in English, most of them are computationally expensive and rely on
significant manual annotation and large corpora. Particularly, the General Inquirer [
First of all we have analyzed a list of antonyms collected from published
dictionaries of antonyms [
Next step was to extend our seed lexicon to obtain a broad coverage of different
texts under consideration concerning sentiment analysis. Automatic approaches to
create (English) semantic orientation lexicon and, more generally, approaches for
word-level sentiment annotation can be grouped into two kinds: (1) those that rely on
manually created lexical resources—most of which use WordNet; and (2) those that
rely on text corpora [
A number of distance or similarity measures exist for English based (completely or
partially) on WordNet. In particular, such measure is defined as the number of edges
of the path through the taxonomic relations (IS-A, Part-of, or WordNet’s hyponymy
relation). In [
The semantic distance between the words could be determined in the similar way as the definition adopted in graph theory [10]. The simplest approach is just to gather all the words from the Dictionary of synonyms and to link each member of a synonymous group with its dominant word as indicated in the Dictionary. Let G(W,S) be the undirected graph, with W the set of nodes being all the words from the Dictionary with associated part-of-speech, S - the set of edges connecting each member of synonymous group with its dominant word. Every group of synonymous words could be connected to each other and form a clique in G graph. A path P is the sequence of nodes connected by edges of G and geodesic is the shortest path between two nodes. Geodesic distance, D(wi,wj) between two words wi and wj is the length (number of edges) of the shortest path between wi and wj. If there is no path between wi and wj, the distance between them is infinity. The minimal path-length defines a metric on the set of synonyms. All axioms of the metric space are fulfilled in this case. Usually synonymous groups comprises the words of the same grammatical category and entire graph G is decomposed into disjoint sub-graphs or networks for nouns, verbs, adjectives and adverbs. (Fig. 1). In each network exists a maximal connected component that contains 70-90% of all nodes of the graph constructed from the Dictionary of synonyms. Maximum component in the class of Russian adverbs contains about 8500 words. The words in this connected component could be analyzed using the metric defined by the length of geodesics. 4
Classical work on the measurement of emotional or affective values in texts is the theory of semantic differential by Charles Osgood. Word meaning in cognitive psychology, is “a strictly psychological one: those cognitive states of human language users which are necessary antecedent conditions for selective encoding of lexical signs and necessary subsequent conditions in selective decoding of signs in messages.” [11]. Semantic differential method was applied mainly to the adjectives measured in such dimensions as active/passive, good/bad, positive/negative, beautiful/ugly, etc. Each pair of bipolar adjectives is a factor or an axis in the method of semantic differential. Application of factor analysis to extensive empirical material gave an unexpected result: most of the variance in judgment could be explained by only three major factors including the evaluative factor (e.g., positive/negative); the potency factor (e.g., strong/weak); and the activity factor (e.g., active/passive). Among these three factors, the evaluative factor has the strongest relative weight for determining the semantic orientation.
Turning to the selected Russian adverbs, we note that the vast majority of adverbs is matched with the words of other parts of speech primarily with the adjectives (cheerful – cheerfully // бодрый - бодро, brutal – brutally // жестокий - жестоко), so that the semantic differential can be naturally extended to motivated adverbs, which bear semantic meaning and, accordingly, deliver the information on their semantic orientation. All three pairs of bipolar adverbs negatively/positively (плохо/хорошо); weakly/strongly (слабо/сильно), passively/actively (пассивно/активно) are contained in the maximal component of the sub-graph of synonymous adverbs Gadv. One can assume that the distance to positively (хорошо) is a measure of positive assessment of an adverb. However, it is easy to show that this measure is in fact rather controversial.
A striking example of this is that the words positively (хорошо) and negatively (плохо) are closely related through the path of synonyms. There is a sequence of only 5 words in English (negatively, hardly, tightly, thoroughly, comprehensively, soundly, positively), and 6 words in Russian (плохо, дешево, легко, просто, совсем, очень, здорово, хорошо) – see Fig. 1 – connecting opposites, each pair of words in this sequence is certainly synonymous (at least in one of their meanings). Thus, we find that d(positively, negatively) = 6; d(хорошо, плохо) = 7. Despite the fact that the adverb positively (хорошо) and negatively (плохо) have opposite meanings, they are closely related by synonymy path. Of course, this is not due to any error in the Dictionary of synonyms. Partial explanation lies in the wide use of two Russian adverbs хорошо (625 ipm), плохо (187 ipm) [12]. Due to the fact that both words are members of the maximum connected component of Gadv sub-graph, we can consider not only the shortest distance from any adverb to "positively", but the shortest distance to its antonym, "negatively". This idea is concretized [13] in the definition of EVA function, which allows to measure the relative distance from the word of two opposites, "positively" and "negatively":
EVA (w) = (d (w, neg)-d (w, pos)) / d (neg, pos).
Under the assumption that there is no word "worse than negatively" or "better than
positively" the values of EVA lie in the interval [
An assumption on the boundary position of words negatively/positively is not entirely justified. Intuitively, perfectly (превосходно) is better than positively, disgustingly (отвратительно) is worse than negatively. Bearing this in mind and using the geometry of a triangle with vertices {w, pos, neg}, we redefine the function of EVA, namely:
EVA1(w)=(d(w, neg)-d(w, pos))*(d(w, neg)+d(w, pos))/d2(neg, pos).
The values of EVA1 sometimes are beyond the interval [
For English adjectives (and motivated adverbs) there exists the source for assessing the measure constructed above in comparison with the independently obtained answers to the "General Inquirer” [11], which contains a set of words to assess three Osgood’s factors. Word lists were obtained from the Stanford political dictionary, where each of the 3000 most frequent common words were assessed by three or more experts concerning each Osgood’s factor. Thus 765 positive and 873 negative words for the assessment factor were obtained, 1474 strong and 647 weak word for the potency factor and 1568 active and 732 passive words for activity factor. Comparison of results obtained with the General Inquirer gave the values of 70 - 80% of matches, depending on what words were considered as neutral in terms of EVA function.
In the absence of available data for content analysis we used the Russian dictionaries of antonyms as an independent source. Antonymous pair is a pair of words (or rather, the specific meanings of words), one opposed to the other on semantic grounds, such as hot - cold fast - slow, present - absent. We suggest that adverbs belonging to pair of antonyms lie on the "opposite sides" of the entire set of adverbs. Methods of multidimensional scaling deliver a mapping of multidimensional space with the defined distance between individual points d(wi, wj) onto a space of smaller dimension, namely the plane (Fig. 2). Figure 2 a, b shows that the pairs of antonyms lie near the diameters of the set of adverbs. For a more profound study of the structure of the space of adverbs we have constructed chains of synonyms connecting antonyms pairs within the sub-graph Gadv.
Fig. 2. Two chains of synonyms, joining antonymous pairs of adverbs.
a) up – a consistent path; b) down – an inconsistent result.
Chain in Fig. 2a is a consistent result, i.e. the chain of synonyms passes on the periphery of the set of adverbs and the distances between the synonyms do not exceed the distance between the antonyms. Unfortunately, the situation is not always as favorable. In Fig. 2b pair of antonyms is close to the diameter, but the chain of synonyms is not at the periphery of the set, but lays in the central part of the set, alternates its direction, and the distances between synonyms is often greater than the distance between antonyms. Probably it is necessary to determine more accurate distance between the words and to choose correctly the axes of the adverb space using the principal components method. These new axes should not coincide Osgood’s dimensions. 5
In this paper we define a measure of the distance between adverbs using synonyms graph. It seems obvious that the choice of similarity measure, or distance largely depends on the type of the problem. The choice of distance measure on the grounds of synonyms is connected with the goal of determining the semantic orientation of adverbs. In contrast to Osgood’s semantic differential associated with the reaction of people on the stimulus - words presented, or the possible emotional impact of words, this model is based solely on the lexical material and is intended to represent relatively objective meanings which are fixed in Dictionaries. Further studies will determine the semantic orientation of sentences or the whole text on the basis of the orientation of its constituent words. Our method allows to evaluate other classes of words such as nouns, adjectives and verbs, but this extension will require a significant increase of calculations and special methods for processing large data sets, since an algorithm for computing shortest paths requires O(n3) operations, where n is the number of words in graph G(W,S). 9. Budanitsky, A., Hirst, G. Semantic distance in WordNet: An experimental, applicationoriented evaluation of five measures. In: Workshop on WordNet and Other Lexical Resources. Second meeting of the NAACL, Pittsburgh, (2001) 10. Potemkin, S.B. Semantic Distance over Lexical Database and WordNet In: Proceedings of 10 International Conference «Cognitive Modeling in Linguistics », CML, Montenegro Bechichi (2008). 11. Osgood, C.E., Succi, G.J., Tannenbaum P.H., The Measurement of Meaning.. University of Illinois Press, Urbana IL (1957). 12. Sharov, S. Frequency Dictionary of Russian
http://www.artint.ru/projects/frqlist.asp (2003) 13. Kamps, J., Marx, M., Robert, J., Mokken, M. Using WordNet to Measure Semantic Orientations of Adjectives In: Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC'04, Vol. IV ) pp. 1115-1118. (2004)