Emotion Recognition From Text-a Survey

Ms. Pallavi D. Phalke , Dr. Emmanuel M.

ABSTRACT

Emotion is a very important facet of human behaviour which affect on the way people interact in the society. In recent year many methods on human emotions recognition have been published such as recognizing emotion from facial expression and gestures, speech and by written text. This paper focuses on classification of emotion expressed by the online text, based on pre–defined list of emotion. The collection of dataset is the basic step, which is collected from the various sources like daily used sentences, user status from various social networking websites such as facebook and twitter. Using this data set we target only on the keywords that show human emotions. The targeted keywords are extracted from the dataset and translated into the format which can be processed by the classifier to finally generate the Predicting model which is further compared by the test dataset to give the emotions in the input sentences or documents.

Keywords— Affective Computing, Classification, Document Categorization, Emotion Detections.

1. INTRODUCTION

Recently much research is going on in emotion recognition domain. Recognition of emotions is very useful to human-machine communication. Many kinds of the communication system can react properly for the human’s emotional actions by applying emotion recognition techniques on them. These systems include dialogue system, automatic answering system and robot. The recognition of emotion has been implemented in many kinds of media, such as image, speech, facial expressions, signal, textual data, and so on. Text is the most popular and main tool for the human to convey messages, communicate thoughts and express inclination. Textual data make it possible for people to exchange opinions, ideas, and emotions using text only. Therefore the research for recognizing from the textual data is valuable. Keyword-based approach to the proposed system since the keyword-based approach shows high recognizing accuracy for emotional keywords.

Interaction between humans and computers has been increased with increase in development of information technology. Recognizing emotion in text from document or sentences is the first step in realizing this new advanced communication which includes communication of information such as how the writer/speaker feels about the fact or how they want the reader/listener to feel. Analyzing text, detecting emotions is useful for many purposes, which includes identifying what emotion a newspaper headline is trying to evoke, identifying users emotion from their statuses of different social networking sites, devising dialogue systems that respond appropriately to different emotional states of the user and identifying blogs that express specific emotions towards the topic of interest. List of emotions and words that are indicative of each emotion is likely to be useful in identifying emotions in text because, many times different emotions are expressed by different words. For example cry and gloomy are indicative of sadness, boiling and shout are indicative of anger, yummy and delightful indicate the emotion of joy.

To capture emotion from text document we require the classification which aims at presume the emotion conveyed by the documents based on predefined lists of emotion, such as Joy, Anger, Fear, Disgust, Sad and Surprise. This emotion recognition approach is mainly focused on two main tasks.

1) The test data that is text document collected from any news articles, user statuses from different social networking sites etc. required for understanding the emotions evoked by words. This is because a different word arouses different emotions comprehended from our day to day experiences. For this purpose, need is to enhanced dictionary with emotion word from ISEAR, WorldNet Affect to improve in result.

2) Need for text normalization to handle negation, since the scope of words is larger in this scenario, the usage of words and their diverted form is large too. So these problems need to be solved properly.

The next part of this paper is organised as follows: Section II discusses a survey of emotion detection from text, Section III describes different algorithms on different datasets for emotion recognition, Section IV briefly compares proposed work followed by experimental study with result in section V and Section V concludes the paper.

2. THE SURVEY OF EMOTION DETECTION FROM TEXTS

Definitions about emotion, its categories, and their influences have been an important research issue long before computers emerged, so that the emotional state of a person may be inferred under different situations. In its most common formulation, the emotion detection from text problem is reduced to finding the relations between specific input texts and the actual emotions that drives the author to type/write in such styles. Intuitively, finding the relations usually relies on specific surface texts that are included in the input texts, and other deeper inferences that will be formally discussed below. Once the relations can be determined, they can be generalized to predict others’ emotions from their articles, or even single sentences.

At the first glance, it does not seem to involve so many difficulties. In real life, different people tend to use similar phrases (i.e. “Oh yes!”) to express similar feelings (i.e. joy) under similar circumstances (i.e. achieving a goal); even they native languages are different, the mapping of such phrases from each language may be obvious. More formally, the emotion detection from text problem can be formulated as follows: Let E be the set of all emotions, A be the set of all authors, and let T be the set of all possible representations of emotion-expressing texts. Let r be a function to reflect emotion e of author a from text t, i.e., r: A × T → E and the function r would be the answer to our problem.

The central problem of emotion detection systems lies in that, though the definitions of E and T may be straightforward from the macroscopic view, the definitions of individual element, even subsets in both sets of E and T would be rather confusing. On one hand, for the set T, new elements may add in as the languages are constantly evolving. On the other hand, currently there are no standard classifications of “all human emotions” due to the complex nature of human minds, and any emotion classifications can only be seen as “labels” annotated afterwards for different purposes.

As a result, before seeking the relation function r, all related research firstly define the classification system of emotion classifications, defining the number of emotions. Secondly, after finding the relation function r or equivalent mechanisms, they still need to be revised over time to adopt changes in the set T. In the following subsections, we will present a classification of emotion detection methods proposed in the literature, based on how detection are made. Although they can all be classified into content-based approaches from the point of view of information retrieval, their problem formulation differs from each other:

1. Keyword-based detection: Emotions are detected based on the related set(s) of keywords found in the input text;

2. Learning-based detection: Emotions are detected based on previous training result with respect to specific statistic learning methods;

3. Hybrid detection: Emotions are detected based on the combination of detected keyword, learned patterns, and other supplementary information;

Besides these emotion detection methods that infer emotions at sentence level, there has been work done also on detection from online blogs or articles [1][2]. For example, though each sentence in a blog article may indicate different emotions, the article as a whole may tend to indicate specific ones, as the overall syntactic and semantic data could strengthen particular emotion(s). However, this paper focuses on detection methods with respect to single sentences, because this is the foundation of full text detection.

A. KEYWORD-BASED METHODS

Keyword-based methods are the most intuitive ways to detect textual emotions. To approximate the set T, since all the names of emotions (emotion labels) are also meaningful texts, these names themselves may serve as elements in both sets of E and T. Similarly, those words with the same meanings of the emotion labels can also indicate the same emotions. The keywords of emotion labels constitute the subset EL in set T, where EL also classifies all the elements in E. The set EL is constructed and utilized based on the assumption of keyword independence, and basically ignores the possibilities of using different types of keywords simultaneously to express complicated emotions.

Keyword-based emotion detection serves as the starting point of textual emotion recognition. Once the set EL of emotion labels (and related words) is constructed, it can be used exhaustively to examine if a sentence contains any emotions.

However, while detecting emotions based on related keywords is very straightforward and easy to use, the key to increase accuracy falls to two of the pre-processing methods, which are sentence parsing to extract keywords, and the construction of emotional keyword dictionary. Parsers utilized in emotion detection are almost ready-made software packages, whereas their corresponding theories may differ from dependency grammar to theta role assignments. On the other hand, constructing emotional keyword dictionary would be naval to other fields [3]. As this dictionary collects not only the keywords, but also the relations among them, this dictionary usually exists in the form of thesaurus, or even ontology, to contain relations more than similar and opposite ones. Semi-automatic construction of EL based on WorldNet-like dictionaries is proposed in [4] and [5].

As was observed in [6], keyword-based emotion detection methods have three limitations described below.

1) AMBIGUITY IN KEYWORD

Though using emotion keywords is a straightforward way to detect associated emotions, the meanings of keywords could be multiple and vague. Except those words standing for emotion labels themselves, most words could change their meanings according to different usages and contexts. It is not feasible to include all possible combinations into the set EL. Moreover, even the minimum set of emotion labels (without all their synonyms) could have different emotions in some extreme cases such as ironic or cynical sentences.

2) INCAPABILITY OF RECOGNIZING SENTENCES WITHOUT KEYWORDS

As Keyword-based approach is totally based on the set of emotion keywords, sentences without any keywords would imply like they don’t contain any emotions at all, which is obviously wrong.

3) LACK OF LINGUISTIC DATA

Syntax structures and semantics also affect on expressed emotions. For example, “He laughed at me “and “I laughed at him” would suggest different emotions from the first person’s point of view. Therefore, ignoring linguistic information also create a problem to keyword-based methods.

B. LEARNING-BASED METHODS

Researchers using learning-based methods attempt to formulate the problem differently. The original problem that determining emotions from input texts has become how to classify the input texts into different emotions. Unlike keyword-based detection methods, learning-based methods try to detect emotions based on a previously trained classifier, which apply various theories of machine learning such as support vector machines [7] and conditional random fields [8], to determine which emotion category should the input text belongs.

However, comparing the satisfactory results in multimodal emotion detection [9], the results of detection from texts drop considerably. The reasons are addressed below:

1) DIFFICULTIES IN DETERMINING EMOTION INDICATORS

The first problem is, though learning-based methods can automatically determine the probabilities between features and emotions, learning-based methods still need keywords, but just in the form of features. The most intuitive features may be emoticons, which can be seen as author’s emotion annotations in the texts. The cascading problems would be the same as those in keyword-based methods.

2) OVER-SIMPLIFIED EMOTION CATEGORIES

Nevertheless, lacking of efficient features other than emotion keywords, most learning-based methods can only classify sentences into two categories, which are positive and negative. Although the number of emotion labels depends on the emotion model applied, we would expect to refine more categories in practical systems.

C. HYBRID METHODS

Since keyword-based methods with thesaurus and naïve learning-based methods could not acquire satisfactory results, some systems use a hybrid approach by combining both or adding different components, which help to improve accuracy and refine the categories. The most significant hybrid system so far is the work of Wu, Chuang and Lin [6], which utilizes a rule-based approach to extract semantics related to specific emotions, and Chinese lexicon ontology to extract attributes. These semantics and attributes are then associated with emotions in the form of emotion association rules. As a result, these emotion association rules, replacing original emotion keywords, serve as the training features of their learning module based on separable mixture models. Their method outperforms previous approaches, but categories of emotions are still limited.

D. SUMMARY AND CONCLUSIONS

As described in this section, much research has been done over the past several years, utilizing linguistics, machine learning, information retrieval, and other theories to detect emotions. Their experiments show that, computers can distinguish emotions from texts like humans, although in a coarse way. However, all methods have certain limitations, as described in the previous subsections, and they lack context analysis to refine emotion categories with existing emotion models, where much work has been done to put them computationalized in the domain of believable agents. On the other hand, applications of affective computing would expect more refined results of emotion detection to further interact with users. Therefore, developing a more advanced architecture based on integrating current approaches and psychological theories would be in a pressing need.

III. ALGORITHMS USED IN EMOTION RECOGNITION

A brief summary of the various works for emotion recognition discussed in this paper are presented in Table1.

Table 1: Results and feature-set comparison of algorithms

IV.EMOTION RECOGNITION IN SOCIAL COMMUNICATION

The block diagram of the emotion recognition system studied in this paper is depicted in Figure 1.It contains three main modules: Affective communication unit, Data Aggregator, Emotion Recognition Engine and recognized emotion class as an output.

Figure 1 : Block diagram of emotion recognition system for Affective communication

1. AFFECTIVE COMMUNICATION UNIT

Affective Communication Unit is nothing but the users account in any social networking site (tweeter or facebook). This system take input from these two social networking sites.

2. DATA AGGREGATOR

Data Aggregator collects user tweets and status from tweeter and facebook. These tweets/status serve as an input to Emotion Recognition Engine.

3. EMOTION RECOGNITION ENGINE

Emotion Recognition Engine including Bayesian Network classifier categorizes incoming data into 3 types of emotions: happiness,

sadness, and neutral, because this system mainly focuses on finding stress level of user. It is broken up into 2 major phase: Training Phase and Testing Phase. Training phase consist of five important parts: The Training Dataset, Keyword Extraction, Keyword conversion, Training Model and Predicting Model. Before it generate the predicting model or file, training phase get the training dataset from which it extracted the keyword from the emotion training date, and convert the keyword using keyword conversion into the format that can be processed by the classifier in the Training Model.

Testing phase which is also called predicting phase consist of Testing dataset, Keyword extraction, Keyword conversion and predict model. The testing phase extract the Keyword from the given sentence, which was the input from the keyboard and then translate the keyword (word of natural language) using the Keyword conversion into the format that can be processed and then we compare it with a predicting file in predict module and finally gives the output as appropriate emotion expressed by the text.

VI.CONCLUSION

The proposed system is able to recognize the happy and sad state of a person from his tweets posted on tweeter from his mobile. The experimental results Shows that the we get better accuracy using Naive Bayes classifier than that of Support Vector Machine.

VII. REFERENCES

[1] 2. Tim M.H. Li, Michael Chau, Paul W.C. Wong, and Paul S.F. Yip”A Hybrid System for Online Detection of Emotional Distress” PAISI 2012, LNCS 7299 Springer-Verlag Berlin Heidelberg 2012M, 73–80.

[2] Abbasi, A., Chen, H., Thoms, S., Fu, T.: “Affect Analysis of Web Forums and Blogs Using Correlation Ensembles.” IEEE Transactions on Knowledge and Data Engineering (2008) ,1168–1180.

[3] T. Wilson, J. Wiebe, and R. Hwa, “Just how mad are you? Finding strong and weak opinion clauses,” Proc. 21st Conference of the American Association for Artificial Intelligence Jul. 2007, 761-769.

[4] D. B. Bracewell, “Semi-Automatic Creation of an Emotion Dictionary Using WordNet and its Evaluation,” Proc. IEEE conference on Cybernetics and Intelligent Systems, IEEE Press, Sep. 2008, 21-24.

[5] J. Yang, D. B. Bracewell, F. Ren, and S. Kuroiwa, “The Creation of a Chinese Emotion Ontology Based on HowNet”, Engineering Letters, Feb. 2008,166-171.

[6] C.-H. Wu, Z.-J. Chuang, and Y.-C. Lin, “Emotion Recognition from Text Using Semantic Labels and Separable Mixture Models,” ACM Transactions on Asian Language Information Processing Jun. 2006, 165-183.

[7] Z. Teng, F. Ren, and S. Kuroiwa, “Recognition of Emotion with SVMs,” in Lecture Notes of Artificial Intelligence Eds.Springer, Berlin Heidelberg, 2006,701-710 .

[8] C. Yang, K. H.-Y. Lin, and H.-H. Chen, “Emotion classification using web blog corpora,” Proc. IEEE/WIC/ACM International Conference on Web Intelligence. IEEE Computer Society, Nov. 2007, 275-278.

[9] C. M. Lee, S. S. Narayanan, and R. Pieraccini, “Combining Acoustic and Language Information for Emotion Recognition,” Proc. 7^th International Conference on Spoken Language Processing (ICSLP 02), 2002, 873-876. [10]http://www.affectivesciences.org/reserachmaterial

[11] http://www.weka.net.nz/