Datasets:

WINGNUS
/

ACL-OCL

	{
	"paper_id": "Y16-2006",
	"header": {
	"generated_with": "S2ORC 1.0.0",
	"date_generated": "2023-01-19T13:47:16.706766Z"
	},
	"title": "Multiple Emotions Detection in Conversation Transcripts",
	"authors": [
	{
	"first": "Duc-Anh",
	"middle": [],
	"last": "Phan",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Nara Institute of Science and Technology",
	"location": {
	"postCode": "8916-5, 630-0192",
	"settlement": "Takayama, Ikoma, Nara",
	"country": "Japan"
	}
	},
	"email": "phan.ducanh.oq3@is.naist.jp"
	},
	{
	"first": "Hiroyuki",
	"middle": [],
	"last": "Shindo",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Nara Institute of Science and Technology",
	"location": {
	"postCode": "8916-5, 630-0192",
	"settlement": "Takayama, Ikoma, Nara",
	"country": "Japan"
	}
	},
	"email": "shindo@is.naist.jp"
	},
	{
	"first": "Yuji",
	"middle": [],
	"last": "Matsumoto",
	"suffix": "",
	"affiliation": {
	"laboratory": "",
	"institution": "Nara Institute of Science and Technology",
	"location": {
	"postCode": "8916-5, 630-0192",
	"settlement": "Takayama, Ikoma, Nara",
	"country": "Japan"
	}
	},
	"email": ""
	}
	],
	"year": "",
	"venue": null,
	"identifiers": {},
	"abstract": "In this paper, we present a method of predicting emotions from multi-label conversation transcripts. The transcripts are from a movie dialog corpus and annotated partly by 3 annotators. The method includes building an emotion lexicon bootstrapped from Wordnet following the notion of Plutchik's basic emotions and dyads. The lexicon is then adapted to the training data by using a simple Neural Network to fine-tune the weights toward each basic emotion. We then use the adapted lexicon to extract the features and use them for another Deep Network which does the detection of emotions in conversation transcripts. The experiments were conducted to confirm the effectiveness of the method, which turned out to be nearly as good as a human annotator.",
	"pdf_parse": {
	"paper_id": "Y16-2006",
	"_pdf_hash": "",
	"abstract": [
	{
	"text": "In this paper, we present a method of predicting emotions from multi-label conversation transcripts. The transcripts are from a movie dialog corpus and annotated partly by 3 annotators. The method includes building an emotion lexicon bootstrapped from Wordnet following the notion of Plutchik's basic emotions and dyads. The lexicon is then adapted to the training data by using a simple Neural Network to fine-tune the weights toward each basic emotion. We then use the adapted lexicon to extract the features and use them for another Deep Network which does the detection of emotions in conversation transcripts. The experiments were conducted to confirm the effectiveness of the method, which turned out to be nearly as good as a human annotator.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Abstract",
	"sec_num": null
	}
	],
	"body_text": [
	{
	"text": "Along with the trend of \"Affective Computing\", the task of Emotion Detection in text has received much attention in the recent years. However, very little research has been working on the detection of multiple emotion simultaneously. Instead, most of them make simple assumption that emotions are mutually exclusive and focus on multi-class classification. In fact, the nature of human emotion is complicated: emotions have connections, some are opposite of each other, while some occur together at the same time, resonate and create another emotional statedyads (Plutchik, 1980) The survey by Dave and Diwanji (2015) predicted the need for Emotion Detection in streaming data and the study of emotion flow during chatting. In this paper, we tackle the simplified version of this task by detecting the emotions in conversation. The corpus we used is made of conversations among movie characters, who take turns in the conversation. Those turns are called utterances, which are then manually annotated in a multi-label manner.",
	"cite_spans": [
	{
	"start": 563,
	"end": 579,
	"text": "(Plutchik, 1980)",
	"ref_id": "BIBREF19"
	},
	{
	"start": 594,
	"end": 617,
	"text": "Dave and Diwanji (2015)",
	"ref_id": "BIBREF10"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Emotion detection in conversation is essentially different from identifying emotions in news headlines (Strapparava and Mihalcea, 2007) or Tweets (Bollen et al., 2011) where each instance is independent of each other. Generally, the expression of Emotion in general depends on the words being used. However, it also quite depends on the grammar structure and syntactic variables such as: negations, embedded sentence, and the type of sentencequestion, exclamation, command or statement (Collier, 2014) . Therefore, similar to the detection of emotions of sentences in a paragraph, the context information of the whole conversation and what is said in the previous utterance should be taken into consideration. The extraction of context features will be further explained in sub-section 4.3.1",
	"cite_spans": [
	{
	"start": 103,
	"end": 135,
	"text": "(Strapparava and Mihalcea, 2007)",
	"ref_id": "BIBREF20"
	},
	{
	"start": 146,
	"end": 167,
	"text": "(Bollen et al., 2011)",
	"ref_id": "BIBREF4"
	},
	{
	"start": 486,
	"end": 501,
	"text": "(Collier, 2014)",
	"ref_id": "BIBREF8"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Unlike other works where small sets of basic emotions are used, we annotated the dataset using the notion of Plutchik's basic emotion and dyads (1980) . This eases the annotators' task since it offers annotators with wider range of emotion labels (8 basic and 23 combinations) to choose from.",
	"cite_spans": [
	{
	"start": 109,
	"end": 150,
	"text": "Plutchik's basic emotion and dyads (1980)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Previous research often relied on a list of 6 basic emotions (Ekman et al., 1987) with some variants. However, this notion fails to show conflict side of some emotions. For example, people should not feel happiness and sadness from the same incident altogether. Furthermore, Ekman's basic emotions are the result of observation made on human facial expressions so applying such notion in text classification task seems irrelevant. Newer works relies on dimensional representation using valence-arousal space (Calvo and Mac Kim, 2013; Yu et al., 2015) Plutchik (1980) suggested 4 axes of bipolar basic emotions: Joy -Sadness, Fear -Anger, Trust -Disgust, Surprise -Anticipatation. These primary emotions may blend to form the full spectrum of human emotional experience. The new complex emotions formed by them are called dyads (Figure 1 ). Plutchik's notion reasonably explains the connection between emotions. Some emotions will not occur at the same time since they are on the opposite side of the axis. Complex emotions can also be viewed as combinations of primary ones. The idea enables us to approach emotion detection in a more comprehensive manner. In the future, we may address not only complicating mixture of emotions but also the intensity of each of them.",
	"cite_spans": [
	{
	"start": 61,
	"end": 81,
	"text": "(Ekman et al., 1987)",
	"ref_id": "BIBREF12"
	},
	{
	"start": 508,
	"end": 533,
	"text": "(Calvo and Mac Kim, 2013;",
	"ref_id": "BIBREF6"
	},
	{
	"start": 534,
	"end": 550,
	"text": "Yu et al., 2015)",
	"ref_id": "BIBREF24"
	},
	{
	"start": 560,
	"end": 566,
	"text": "(1980)",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 827,
	"end": 836,
	"text": "(Figure 1",
	"ref_id": "FIGREF0"
	}
	],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "In this paper, we propose a three steps method for the detection of emotions in conversation: 1)Building Emotion Lexicon from Wordnet (Miller, 1995) . 2) Using simple Neural Network to adapt the lexicon to the training data. 3) Using Deep Network with features extracted from adapted lexicon and classify the multi-label corpus.",
	"cite_spans": [
	{
	"start": 134,
	"end": 148,
	"text": "(Miller, 1995)",
	"ref_id": "BIBREF17"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "The remainder of the paper is organized as follows. Section 2 summarizes related work on emotion detection. Section 3 discusses the nature of our dataset and explains the annotating scheme. Section 4 proposes our approach which includes the 3 steps mentioned above. Section 5 evaluates the lexicon, the effectiveness of the adapted lexicon and the proposed method in general. Section 6 gives the conclusion and discusses future work.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Introduction",
	"sec_num": "1"
	},
	{
	"text": "Most of the work in the field tried to define a small set of emotions (D'Mello et al., 2006; Yang et al., 2007) which involved only 3 and 4 emotional states respectively. Another work by Hasegawa et al. (2013) performed a multi-class classification on dialog data from Twitter in Japanese. They automatically labeled the obtained dialogs by using emotional expression clues, which is similar to our collocation list explained in sub-section 3.3. We propose a more comprehensive approach by exploiting Plutchik's notion which covers the full spectrum of human emotions to work on challenging multi-label conversation corpus.",
	"cite_spans": [
	{
	"start": 70,
	"end": 92,
	"text": "(D'Mello et al., 2006;",
	"ref_id": "BIBREF11"
	},
	{
	"start": 93,
	"end": 111,
	"text": "Yang et al., 2007)",
	"ref_id": "BIBREF23"
	},
	{
	"start": 187,
	"end": 209,
	"text": "Hasegawa et al. (2013)",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "Having the same notion, Buitinck et al. (2015) proposed a simple Bag of Words approach and tuned RAKEL for multi-label classification for movie reviews. We go further and work on conversation data where the exchange between characters and the context of the whole dialog are of great importance. The closest to our work is on paragraphs and documents which tried to improve the sentencelevel prediction of some special emotions which, due to data sparseness and inherent multi-label classification, were very hard to predict. They incorporated label dependency among labels and context dependency into the graph model to achieve such goal. However, their work is for paragraphs in Chinese. In our case, we take advantages of Deep Neural Network to capture the abstract representation of context information.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "Our system is different from previous methods in four main ways:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "\u2022 Plutchik's notion of primary emotions and dyads is incorporated in our system and provides scalability to address more than just primary emotions if needed in the future.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "\u2022 We bootstrapped the lexicon and then adapted it to the training data which improved the classification result",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "\u2022 The proposed method includes 2 neural networks, one for adapting the lexicon and the other for multi-label classification of emotions.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "\u2022 We use a set of manually constructed features instead of word-embedding directly for the Neural Network. The reason for that is further discussed in sub-section 4.3.1",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "3 Corpus, Dataset & Annotation Scheme",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Related Work",
	"sec_num": "2"
	},
	{
	"text": "The Cornell Movie Dialog dataset 1 was originally used for understanding the coordination of linguistic style in dialogs (Danescu-Niculescu-Mizil and Lee, 2011). It includes in total 304,713 utterances (turns in conversation) out of 220,579 conversational exchanges between 10,292 pairs of 9,035 movie characters from 617 movies. The annotating scheme is as follows:",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Movie Dialog Corpus",
	"sec_num": "3.1"
	},
	{
	"text": "\u2022 One utterance may hold zero, one or more emotions at the same time. The list of emotions to assign includes Plutchik's 8 basic emotions and 23 dyads. The sytem will treat the dyads as combination of basic emotions. In case an utterance holds no emotion, it should be annotated with \"None\"",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Movie Dialog Corpus",
	"sec_num": "3.1"
	},
	{
	"text": "\u2022 The annotators need to assign the whole utterance which may have two or more sentences inside with a set of all emotions expressed inside it. There may be cases where confict emotions according to Plutchik's notion appear simultaneously in the same utterance.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Movie Dialog Corpus",
	"sec_num": "3.1"
	},
	{
	"text": "The followings are some statistics of the corpus: total of 11,610 utterances, 10,008 of which are in the training data , 1,602 others are in the testing data, the average number of label per utterance is 1.29. We separated the training data which was annotated by only one annotator and the testing data which was annotated by all three annotators.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Movie Dialog Corpus",
	"sec_num": "3.1"
	},
	{
	"text": "1 http://www.mpi-sws.org/\u02dccristian/Cornell_ Movie-Dialogs_Corpus.html",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Movie Dialog Corpus",
	"sec_num": "3.1"
	},
	{
	"text": "One of the most common Inter-Annotator Agreement measurement is the Kappa statistics (Cohen, 1960) . Bhowmick et al. 2008suggested a Kappabased measurement for multi-class classification. However, none of them are applicable to our multilabel corpus because their ways of computing causes hypothetical probability of chance agreement P e to be greater than 1 since there are cases where two or more labels are annotated to a given instance. Therefore, we measure the Kappa statistics for each emotion class and then average them as shown in Table 1 . The survey by Artstein and Poesio (2008) suggested that low kappa scores are often observed in multi-label annotating tasks even when the annotators do not make much use of the ability to assign multiple tags. Some strong emotions: \"Anger\", \"Fear\", \"Surprise\" have better agreement scores as they have indicators such as question marks and excalmation forms. Nevertheless, they are easier for human to identify because they are the basic emotions that we -human inherits from animals. They are the emotions that trigger the \"fight or flight\" and \"stop and examine\" response. (Plutchik, 1980) Due to the time constraint, we had neither the time to show annotators the movies footage nor an adequate amount of sessions to work together and seek a better degree of agreement. Because the annotators only worked with the text data, it was very difficult for them to visualize the situation and make correct judgment.",
	"cite_spans": [
	{
	"start": 85,
	"end": 98,
	"text": "(Cohen, 1960)",
	"ref_id": "BIBREF7"
	},
	{
	"start": 566,
	"end": 592,
	"text": "Artstein and Poesio (2008)",
	"ref_id": "BIBREF0"
	},
	{
	"start": 1127,
	"end": 1143,
	"text": "(Plutchik, 1980)",
	"ref_id": "BIBREF19"
	}
	],
	"ref_spans": [
	{
	"start": 541,
	"end": 549,
	"text": "Table 1",
	"ref_id": "TABREF1"
	}
	],
	"eq_spans": [],
	"section": "Inter-Annotator Agreement",
	"sec_num": "3.2"
	},
	{
	"text": "We also use ISEAR dataset 2 for the process of producing collocation features. In the ISEAR datase, student respondents, both psychologists and nonpsychologists, were asked to report situations in which they had experienced 7 major emotions. Five out of them are completely identical to the basic emotions of Plutchik's . In each case, the questions covered the way they had appraised the situation and how they reacted. Therefore, to our belief, this dataset would provide good collocation features for the 5 identical emotions of our corpus. We mine this dataset for words which frequently appear together with one emotion. If a word also appears in other emotions situation, it loses its place as the indicator toward one specific emotion and we discard it from the collocation list. The use of this collocation list in our work is closely similar to emotional expression clues in (Hasegawa et al., 2013) .",
	"cite_spans": [
	{
	"start": 884,
	"end": 907,
	"text": "(Hasegawa et al., 2013)",
	"ref_id": "BIBREF15"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "ISEAR dataset for Collocation features",
	"sec_num": "3.3"
	},
	{
	"text": "Using Lexicon is proven to provide significant improvement in identifying the emotion conveyed by a word (Mohammad, 2012) . Therefore, in our case, we built a new lexicon, each lexical item of which displays not only its association with Plutchik's basic emotions but also how strong the association is.",
	"cite_spans": [
	{
	"start": 105,
	"end": 121,
	"text": "(Mohammad, 2012)",
	"ref_id": "BIBREF18"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Bulding Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "We define the primary emotions and dyads in Plutchik's theories as the seeds of our lexicon. Throughout Wordnet, we search for synonyms, hypernyms, hyponyms of the seeds. A reverse lemmatisation is necessary to retrieve related verbs, adjectives and adverbs and their derived forms (verb forms and comparative, superlative adjectives) of the seeds. We keep tracks of the original nouns and the seeds where the new words were derived from (Table 2). Note that sometimes a word was derived from different nouns and seeds, which suggests mixed emotional states.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Bulding Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "Each lexical item in the lexicon has a vector of values on each axis of the basic emotions: Joy -Sadness, Fear -Anger, Trust -Disgust, Surprise -",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Bulding Lexicon",
	"sec_num": "4.1"
	},
	{
	"text": "Original Nouns -Seeds joy (primary)-joy sadness (primary)-sadness fear (primary)-fear love (dyad)-love benevolent benevolence-love worship worship-fear , worship-love Anticipatation . We manually assign the primary emotions with a value vector of 1, 0 or -1 and the dyads with 0.5, 0 or -0.5, depending on the axes they belong. For example, \"joy\" came from the axis of Joy-Sadness, thus, its vector is [1,0,0,0] while the vector for \"sadness\" is [-1,0,0,0] ( Table 3 ). The dyad \"love\" came from primary emotions \"joy\" and \"trust\", hence its vector is [0.5,0.5,0,0]. It is to be noted that the minus sign only indicates that the emotion is on the other side of the axis. It is not a suggestion of negative emotion in any case.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 459,
	"end": 466,
	"text": "Table 3",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Words",
	"sec_num": null
	},
	{
	"text": "In addition, we calculate the wup similarity (Wu and Palmer, 1994 ) between a new word and the seed it came from, based on the depth of the two senses in the Wordnet taxonomy and that of their Least Common Subsumer.",
	"cite_spans": [
	{
	"start": 45,
	"end": 65,
	"text": "(Wu and Palmer, 1994",
	"ref_id": "BIBREF22"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Words",
	"sec_num": null
	},
	{
	"text": "wup(word, seed) = 2 * dep(lcs) dep(word) + dep (seed) (1) We assumed that the higher the similarity, the closer emotional state of the word to the seed. Thus, the value vector of a word is the sum of the products of each seed vector and the similarity between the word and such seed.",
	"cite_spans": [
	{
	"start": 47,
	"end": 53,
	"text": "(seed)",
	"ref_id": null
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Words",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "vector(word) = n k=1 vector(seed k ) \u00d7wup(word, seed k ) .",
	"eq_num": "(2)"
	}
	],
	"section": "Words",
	"sec_num": null
	},
	{
	"text": "For example, in the case of the word \"worship\", we first calculate the wup scores between the word and its two seeds: fear and love (Table 2) . Next, they are multiplied by the vectors of the seeds fear-[0,0,1,0] and love-[0.5,0.5,0,0], and then summed up to get the result (Table 3) .",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 132,
	"end": 141,
	"text": "(Table 2)",
	"ref_id": "TABREF2"
	},
	{
	"start": 274,
	"end": 283,
	"text": "(Table 3)",
	"ref_id": null
	}
	],
	"eq_spans": [],
	"section": "Words",
	"sec_num": null
	},
	{
	"text": "We understand that a lexicon bootstrapped from a general domain resource such as Wordnet has its effectiveness limited when it is applied on a specified domain. In order to partly solve this problem, we built a simple neural network with one input layer and one output softmax layer. The input to the network is the Bag-of-Words features of the training data. We then steps by steps, try to do binary classification on the basic emotion e j . Let each node of the network be corresponding to each lexical item in the Lexicon. The biases a node b i are initialized according to the value of each lexical items in the lexicon while the weights w i are randomly initialized. After each step, we update the biases and weights and then repeat the process for the whole 8 basic emotions. Figure 2 shows the structure of the network. We use the log-likelihood as the cost function for the network input: C = \u2212 ln(a L y ) where a L is the output of the final layer and y is the desired output. In the end, we updated the lexicon with new values from the network. We will discuss about the improvements made by the network later in section 5.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 782,
	"end": 790,
	"text": "Figure 2",
	"ref_id": "FIGREF2"
	}
	],
	"eq_spans": [],
	"section": "Adapting Lexicon to Training data",
	"sec_num": "4.2"
	},
	{
	"text": "The process of feature selection for the network is an heuristic one. We initially used a lot of features and then through logistic regression, unimportant features such as the genre of the movie or ngrams features were filtered out.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "The core part of the extraction process is to take advantages of the lexicon to transform an utterance to a vector of values expressing the tendency towards each emotion state. This task is done in a rule-based manner (Algorithm 1 and Figure 3) . Each word in the utterance is mapped to the lexicon to retrieve the value vector. The representation vector of an utterance is the sum vector of all the word inside it. The negation and word dependency are also taken into account when we calculate the sum with the help of NLTK (Bird et al., 2009) 2. The sum vector of all the utterances in the lexicon that appear in the conversation which provides the context of the conversation.",
	"cite_spans": [
	{
	"start": 525,
	"end": 544,
	"text": "(Bird et al., 2009)",
	"ref_id": null
	}
	],
	"ref_spans": [
	{
	"start": 235,
	"end": 244,
	"text": "Figure 3)",
	"ref_id": "FIGREF3"
	}
	],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "3. The sum vector of the previous utterance in the conversation which also provides the context of previous exchange (of what triggered the current emotion).",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "4. The polarity (negative/ positive) score of the sentence.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "5. Features such as: length, is it a question, is it an exclamatory sentence.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "6. Collocation features which indicate the number of appearances of words inside the ISEAR collocations list.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "The reason for us to use extracted features is that it is very hard to capture the context of both the conversation and previous exchange using direct wordembedding. While using a recurrent neural network can solve the latter, it is a challenge to address the first. Each conversation has different number of utterances, it may hurt the performance of the system and result in network architecture complexity if we use a non-fixed size window to monitor all the utterances in a same conversation.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Features Extraction",
	"sec_num": "4.3.1"
	},
	{
	"text": "The structure of the network is built as shown in Figure 4 . The raw input is generalized to produce a small set of features. These features are fed to the network as input layer. We have 2 fully connected hidden layers and an output layer. Since the task is a multi-label classification problem where softmax cannot be used,the output layer is change into sigmoid we add a set of threshold values (one for each basic emotions). Only the labels, whose output values greater than the threshold are considered valid. The thresholds are randomly initialized and then updated after each epochs the same way we updated the biases and weights. In our implementation of the network, Theano (Bastien et al., 2012) was used to take advantages of GPU computing power. The global cost function, similar to Zhang and Zhou (2006) , is defined to reward the system for right predictions and severely punish for wrong ones in equation 3.",
	"cite_spans": [
	{
	"start": 683,
	"end": 705,
	"text": "(Bastien et al., 2012)",
	"ref_id": null
	},
	{
	"start": 795,
	"end": 816,
	"text": "Zhang and Zhou (2006)",
	"ref_id": "BIBREF25"
	}
	],
	"ref_spans": [
	{
	"start": 50,
	"end": 58,
	"text": "Figure 4",
	"ref_id": "FIGREF4"
	}
	],
	"eq_spans": [],
	"section": "Building the Deep Network",
	"sec_num": "4.3.2"
	},
	{
	"text": "E = m i = 1 \|Y i \|\|\u0232 i \| (k,l)\u2208Y i \u00d7\u0232 i exp (\u2212(c i k \u2212 c i l ))",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Building the Deep Network",
	"sec_num": "4.3.2"
	},
	{
	"text": "(3) Let X be the set of all m instances. Let Y = {1, 2, .., Q} be the set of all possible labels, Y i is the set of true labels for ith instance x i and\u0232 i is the set of the labels not belong to x i . Obviously, Y i \u222a\u0232 i = Y . We define E as the global cost function of the network. c i is the set of actual outputs of the model for input x i , each label has its own output. c i k is the output of label k belongs to the set of true labels, k \u2208 Y i . Meanwhile, c i l is the output of label l for l \u2208\u0232 i . The difference c i k \u2212 c i l measure the output of the system between the labels, which an instance belong to and which it doesn't. Naturally, we want this difference to be as big as possible.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Building the Deep Network",
	"sec_num": "4.3.2"
	},
	{
	"text": "As mention above, we used the annotated movie dialog corpus for testing our method. For the gold standard of the test data, we applied the majority rules on the annotation. If one emotion is annotated by two or more annotators, we accept it as a true label for the utterance.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Corpus",
	"sec_num": null
	},
	{
	"text": "Evaluation Metrics In our study, 4 common evaluation metrics which have been popularly used in multi-label classification problems (Godbole and Sarawagi, 2004; are employed to measure the performance of our system to the baselines. Let Y i be set of true labels for a given sample, then Y i is the set the labels predicted by a system. Let m be the total number of samples.",
	"cite_spans": [
	{
	"start": 131,
	"end": 159,
	"text": "(Godbole and Sarawagi, 2004;",
	"ref_id": "BIBREF14"
	}
	],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Corpus",
	"sec_num": null
	},
	{
	"text": "Hamming score or accuracy in multi-label classification, gives the degree of similarity between the ground truth set of labels and the predicted set of labels.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "Hammingscore = 1 N N i \|Y i \u2229 Y i \| \|Y i \u222a Y i \|",
	"eq_num": "(4)"
	}
	],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "2. Precision: the fraction of correctly predicted labels over all the predicted labels in the set.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "P recision = 1 N N i \|Y i \u2229 Y i \| \|Y i \|",
	"eq_num": "(5)"
	}
	],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "3. Recall: the fraction of correctly predicted labels over all the true labels in the set.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "EQUATION",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [
	{
	"start": 0,
	"end": 8,
	"text": "EQUATION",
	"ref_id": "EQREF",
	"raw_str": "Recall = 1 N N i \|Y i \u2229 Y i \| \|Y i \|",
	"eq_num": "(6)"
	}
	],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "4. F1-measure: the harmonic mean of Precision and Recall. In our study, we gave equal importance to Precision and Recall.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "F 1 = 2 * precision * recall precision + recall (7)",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "1.",
	"sec_num": null
	},
	{
	"text": "To evaluate the system, we tried to replicate other works and applied them on our new corpus. A similar work is Buitinck et al. (2015) which use the same Plutchik's basic emotions and work on multi-label data. We used similar Meka's 3 RAkEL method and Bag-of-Words approach as the first baseline. We understand that Buitinck et al. (2015) 's system is finetuned for their corpus, therefore, it is a little unfair to apply it to our corpus and make comparison. Therefore, the second baseline is Meka's DBPNN which 3 http://meka.sourceforge.net/#about is reported as generally having better accuracy than RAkEL (Fernandez-Gonzalez et al., 2015) . We decided that the most important baseline is the human annotation. We calculated the evaluation metrics based on the annotation made by each annotator against the gold standard and averaged the result by the total number of annotators. Another baseline is our own system using the lexicon before adaptation. Figure 5 compares the performance of our system to the baselines.",
	"cite_spans": [
	{
	"start": 112,
	"end": 134,
	"text": "Buitinck et al. (2015)",
	"ref_id": "BIBREF5"
	},
	{
	"start": 316,
	"end": 338,
	"text": "Buitinck et al. (2015)",
	"ref_id": "BIBREF5"
	},
	{
	"start": 609,
	"end": 642,
	"text": "(Fernandez-Gonzalez et al., 2015)",
	"ref_id": "BIBREF13"
	}
	],
	"ref_spans": [
	{
	"start": 955,
	"end": 963,
	"text": "Figure 5",
	"ref_id": "FIGREF5"
	}
	],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "vs. Bag-of-Words Approaches: Our system, with and without lexicon adaptation performed remarkably better than the simple approaches using Meka's DBPNN and RAkEL. It exceeded the better DBPNN significantly in Hamming Score by 7.28 and 7.19, Recall by 12.85 and 5.95, F1-measure by 7.33 and 4.33 respectively. We argue that the context features played as an important factor here.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "Lexicon Adaptation vs. No Adaptation: We can clearly see the improvements made by the adaptation on our system in Recall and F1-measure,which are increased by 6.9 and 3.0. This confirmed the necessity of the adaptation step.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "vs. Human Annotator: This is the most important baseline, which explains how well our system performs in comparison with a Human Annotator. Please note that these values are averaged by the total number of annotators after the judgment made by each annotator are compared to the gold standard. Our system is slightly worse than a Human Annotator in all 4 metrics by 0.43 in Hamming Score, 0.79, 1.67, 1.69 in Precision, Recall and F1-measure respectively.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "These results confirmed the performance of our method which is slightly worse than such of an human annotator. On the other hand, our method is more efficient than simple Bag-of-Words approaches. We also confirmed the improvement made by the Lexicon Adaptation step to our system.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "Classification result for each emotion class: Table 4 shows the distribution of emotion classes and reports the classification result of each emotion class in the corpus. Imbalance can be seen among classes in the corpus. We observed the expected \"All-No-Recurrence\" problem for minority classes of Joy and Sadness (high accuracy and near zero F1) as the corpus is unbalanced. \"Surprise\" is the class with the highest Agreement score (Table 1) , it also achieves the highest F1 among other classes. While \"Anticipation\" class is a dominant class in the corpus, it suffers from low Agreement score. As a consequence, the classification result for this class is also not high. From this result, we can hope that in the future when movie footage are included, not only the agreement score but the system performance will also go up as well.",
	"cite_spans": [],
	"ref_spans": [
	{
	"start": 46,
	"end": 53,
	"text": "Table 4",
	"ref_id": "TABREF5"
	},
	{
	"start": 434,
	"end": 443,
	"text": "(Table 1)",
	"ref_id": "TABREF1"
	}
	],
	"eq_spans": [],
	"section": "Experimental Results",
	"sec_num": "5.2"
	},
	{
	"text": "In this paper, we propose our method of detecting and classifying emotions from a conversation corpus. The corpus is a set of movie dialogs annotated with multi-label emotions following Plutchik's notion of basic emotions and dyads. Our method involves building a lexicon from Wordnet using some seed emotion words, adapting the lexicon to the corpus, extracting a feature set from the input and classifying the emotions accordingly with the help of a deep neural network. The experiments show that our method's power to detecting emotion is comparable to that of a human annotator. However, one may argue that the disagreement among annotators may have affected the result. As discussed above, we hope to solve this problem by including the movies' footage in our annotating scheme. At the time of the submission, we are adding the footages as well as improving annotating scheme to have higher Kappa statistics and evaluate again our method. Once finished, the corpus will be published for other researchers to use. In the future, we also want to further exploit the method by incorporating emotion detection on voices and images and monitoring complex emotions other than the basic ones and their intensity.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Conclusion & Future Work",
	"sec_num": "6"
	},
	{
	"text": "30th Pacific Asia Conference on Language, Information and Computation (PACLIC 30)Seoul, Republic of Korea, October 28-30, 2016",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	},
	{
	"text": "http://www.affective-sciences.org/ researchmaterial PACLIC 30 Proceedings",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "",
	"sec_num": null
	}
	],
	"back_matter": [
	{
	"text": "This research was supported by CREST project of Japan Science and Technology Agency. We are grateful to our colleagues from Computational Linguistics Lab, NAIST, Japan who provided insights and expertise that greatly assisted the research. We also thank the reviewers for their valuable comments that further improved our work.",
	"cite_spans": [],
	"ref_spans": [],
	"eq_spans": [],
	"section": "Acknowledgments",
	"sec_num": null
	}
	],
	"bib_entries": {
	"BIBREF0": {
	"ref_id": "b0",
	"title": "Inter-coder agreement for computational linguistics",
	"authors": [
	{
	"first": "Ron",
	"middle": [],
	"last": "Artstein",
	"suffix": ""
	},
	{
	"first": "Massimo",
	"middle": [],
	"last": "Poesio",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Computational Linguistics",
	"volume": "34",
	"issue": "4",
	"pages": "555--596",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Ron Artstein and Massimo Poesio. 2008. Inter-coder agreement for computational linguistics. Computa- tional Linguistics, 34(4):555-596.",
	"links": null
	},
	"BIBREF1": {
	"ref_id": "b1",
	"title": "Arnaud Bergeron, Nicolas Bouchard, and Yoshua Bengio. 2012. Theano: new features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop",
	"authors": [
	{
	"first": "Fr\u00e9d\u00e9ric",
	"middle": [],
	"last": "Bastien",
	"suffix": ""
	},
	{
	"first": "Pascal",
	"middle": [],
	"last": "Lamblin",
	"suffix": ""
	},
	{
	"first": "Razvan",
	"middle": [],
	"last": "Pascanu",
	"suffix": ""
	},
	{
	"first": "James",
	"middle": [],
	"last": "Bergstra",
	"suffix": ""
	},
	{
	"first": "Ian",
	"middle": [
	"J"
	],
	"last": "Goodfellow",
	"suffix": ""
	}
	],
	"year": null,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Fr\u00e9d\u00e9ric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian J. Goodfellow, Arnaud Berg- eron, Nicolas Bouchard, and Yoshua Bengio. 2012. Theano: new features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop.",
	"links": null
	},
	"BIBREF2": {
	"ref_id": "b2",
	"title": "An agreement measure for determining interannotator reliability of human judgements on affective text",
	"authors": [
	{
	"first": "Pabitra",
	"middle": [],
	"last": "Plaban Kr Bhowmick",
	"suffix": ""
	},
	{
	"first": "Anupam",
	"middle": [],
	"last": "Mitra",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Basu",
	"suffix": ""
	}
	],
	"year": 2008,
	"venue": "Proceedings of the Workshop on Human Judgements in Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "58--65",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Plaban Kr Bhowmick, Pabitra Mitra, and Anupam Basu. 2008. An agreement measure for determining inter- annotator reliability of human judgements on affec- tive text. In Proceedings of the Workshop on Human Judgements in Computational Linguistics, pages 58- 65. Association for Computational Linguistics.",
	"links": null
	},
	"BIBREF4": {
	"ref_id": "b4",
	"title": "Modeling public mood and emotion: Twitter sentiment and socio-economic phenomena",
	"authors": [
	{
	"first": "Johan",
	"middle": [],
	"last": "Bollen",
	"suffix": ""
	},
	{
	"first": "Huina",
	"middle": [],
	"last": "Mao",
	"suffix": ""
	},
	{
	"first": "Alberto",
	"middle": [],
	"last": "Pepe",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Johan Bollen, Huina Mao, and Alberto Pepe. 2011. Modeling public mood and emotion: Twitter sentiment and socio-economic phenomena.",
	"links": null
	},
	"BIBREF5": {
	"ref_id": "b5",
	"title": "Multi-emotion detection in user-generated reviews",
	"authors": [
	{
	"first": "Lars",
	"middle": [],
	"last": "Buitinck",
	"suffix": ""
	},
	{
	"first": "Jesse",
	"middle": [],
	"last": "Van Amerongen",
	"suffix": ""
	},
	{
	"first": "Ed",
	"middle": [],
	"last": "Tan",
	"suffix": ""
	},
	{
	"first": "Maarten",
	"middle": [],
	"last": "De Rijke",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "Advances in Information Retrieval",
	"volume": "",
	"issue": "",
	"pages": "43--48",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Lars Buitinck, Jesse Van Amerongen, Ed Tan, and Maarten de Rijke. 2015. Multi-emotion detection in user-generated reviews. In Advances in Information Retrieval, pages 43-48. Springer.",
	"links": null
	},
	"BIBREF6": {
	"ref_id": "b6",
	"title": "Emotions in text: dimensional and categorical models",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "Rafael",
	"suffix": ""
	},
	{
	"first": "Sunghwan Mac",
	"middle": [],
	"last": "Calvo",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Kim",
	"suffix": ""
	}
	],
	"year": 2013,
	"venue": "Computational Intelligence",
	"volume": "29",
	"issue": "3",
	"pages": "527--543",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Rafael A Calvo and Sunghwan Mac Kim. 2013. Emo- tions in text: dimensional and categorical models. Computational Intelligence, 29(3):527-543.",
	"links": null
	},
	"BIBREF7": {
	"ref_id": "b7",
	"title": "Kappa: Coefficient of concordance",
	"authors": [
	{
	"first": "",
	"middle": [],
	"last": "Cohen",
	"suffix": ""
	}
	],
	"year": 1960,
	"venue": "Educ. Psych. Measurement",
	"volume": "20",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "J Cohen. 1960. Kappa: Coefficient of concordance. Educ. Psych. Measurement, 20:37.",
	"links": null
	},
	"BIBREF8": {
	"ref_id": "b8",
	"title": "Emotional expression",
	"authors": [
	{
	"first": "Gary",
	"middle": [],
	"last": "Collier",
	"suffix": ""
	}
	],
	"year": 2014,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Gary Collier. 2014. Emotional expression. Psychology Press.",
	"links": null
	},
	"BIBREF9": {
	"ref_id": "b9",
	"title": "Chameleons in imagined conversations: A new approach to understanding coordination of linguistic style in dialogs",
	"authors": [
	{
	"first": "Cristian",
	"middle": [],
	"last": "Danescu-Niculescu-Mizil",
	"suffix": ""
	},
	{
	"first": "Lillian",
	"middle": [],
	"last": "Lee",
	"suffix": ""
	}
	],
	"year": 2011,
	"venue": "Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, ACL",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Cristian Danescu-Niculescu-Mizil and Lillian Lee. 2011. Chameleons in imagined conversations: A new ap- proach to understanding coordination of linguistic style in dialogs. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, ACL 2011.",
	"links": null
	},
	"BIBREF10": {
	"ref_id": "b10",
	"title": "Trend analysis in social networking using opinion mining a survey",
	"authors": [
	{
	"first": "Saurin",
	"middle": [],
	"last": "Dave",
	"suffix": ""
	},
	{
	"first": "Hiteishi",
	"middle": [],
	"last": "Diwanji",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Saurin Dave and Hiteishi Diwanji. 2015. Trend analysis in social networking using opinion mining a survey.",
	"links": null
	},
	"BIBREF11": {
	"ref_id": "b11",
	"title": "Predicting affective states expressed through an emote-aloud procedure from autotutor's mixed-initiative dialogue",
	"authors": [
	{
	"first": "K D'",
	"middle": [],
	"last": "Sidney",
	"suffix": ""
	},
	{
	"first": "Scotty",
	"middle": [
	"D"
	],
	"last": "Mello",
	"suffix": ""
	},
	{
	"first": "Jeremiah",
	"middle": [],
	"last": "Craig",
	"suffix": ""
	},
	{
	"first": "Arthur C",
	"middle": [],
	"last": "Sullins",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Graesser",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "International Journal of Artificial Intelligence in Education",
	"volume": "16",
	"issue": "1",
	"pages": "3--28",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Sidney K D'Mello, Scotty D Craig, Jeremiah Sullins, and Arthur C Graesser. 2006. Predicting affective states expressed through an emote-aloud procedure from autotutor's mixed-initiative dialogue. Interna- tional Journal of Artificial Intelligence in Education, 16(1):3-28.",
	"links": null
	},
	"BIBREF12": {
	"ref_id": "b12",
	"title": "Universals and cultural differences in the judgments of facial expressions of emotion",
	"authors": [
	{
	"first": "Paul",
	"middle": [],
	"last": "Ekman",
	"suffix": ""
	},
	{
	"first": "V",
	"middle": [],
	"last": "Wallace",
	"suffix": ""
	},
	{
	"first": "Maureen",
	"middle": [
	"O"
	],
	"last": "Friesen",
	"suffix": ""
	},
	{
	"first": "Anthony",
	"middle": [],
	"last": "Sullivan",
	"suffix": ""
	},
	{
	"first": "Irene",
	"middle": [],
	"last": "Chan",
	"suffix": ""
	},
	{
	"first": "Karl",
	"middle": [],
	"last": "Diacoyanni-Tarlatzis",
	"suffix": ""
	},
	{
	"first": "Rainer",
	"middle": [],
	"last": "Heider",
	"suffix": ""
	},
	{
	"first": "William",
	"middle": [
	"Ayhan"
	],
	"last": "Krause",
	"suffix": ""
	},
	{
	"first": "Tom",
	"middle": [],
	"last": "Lecompte",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Pitcairn",
	"suffix": ""
	},
	{
	"first": "E",
	"middle": [],
	"last": "Pio",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Ricci-Bitti",
	"suffix": ""
	}
	],
	"year": 1987,
	"venue": "Journal of personality and social psychology",
	"volume": "53",
	"issue": "4",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Paul Ekman, Wallace V Friesen, Maureen O'Sullivan, Anthony Chan, Irene Diacoyanni-Tarlatzis, Karl Hei- der, Rainer Krause, William Ayhan LeCompte, Tom Pitcairn, Pio E Ricci-Bitti, et al. 1987. Universals and cultural differences in the judgments of facial expres- sions of emotion. Journal of personality and social psychology, 53(4):712.",
	"links": null
	},
	"BIBREF13": {
	"ref_id": "b13",
	"title": "Multidimensional classifiers for neuroanatomical data",
	"authors": [
	{
	"first": "Pablo",
	"middle": [],
	"last": "Fernandez-Gonzalez",
	"suffix": ""
	},
	{
	"first": "Concha",
	"middle": [],
	"last": "Bielza",
	"suffix": ""
	},
	{
	"first": "Pedro",
	"middle": [],
	"last": "Larranaga",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "ICML Workshop on Statistics, Machine Learning and Neuroscience",
	"volume": "",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Pablo Fernandez-Gonzalez, Concha Bielza, and Pedro Larranaga. 2015. Multidimensional classifiers for neuroanatomical data. In ICML Workshop on Statis- tics, Machine Learning and Neuroscience (Stamlins 2015).",
	"links": null
	},
	"BIBREF14": {
	"ref_id": "b14",
	"title": "Discriminative methods for multi-labeled classification",
	"authors": [
	{
	"first": "Shantanu",
	"middle": [],
	"last": "Godbole",
	"suffix": ""
	},
	{
	"first": "Sunita",
	"middle": [],
	"last": "Sarawagi",
	"suffix": ""
	}
	],
	"year": 2004,
	"venue": "Advances in Knowledge Discovery and Data Mining",
	"volume": "",
	"issue": "",
	"pages": "22--30",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Shantanu Godbole and Sunita Sarawagi. 2004. Discrim- inative methods for multi-labeled classification. In Advances in Knowledge Discovery and Data Mining, pages 22-30. Springer.",
	"links": null
	},
	"BIBREF15": {
	"ref_id": "b15",
	"title": "Predicting and eliciting addressee's emotion in online dialogue",
	"authors": [
	{
	"first": "Takayuki",
	"middle": [],
	"last": "Hasegawa",
	"suffix": ""
	},
	{
	"first": "Nobuhiro",
	"middle": [],
	"last": "Kaji",
	"suffix": ""
	},
	{
	"first": "Naoki",
	"middle": [],
	"last": "Yoshinaga",
	"suffix": ""
	},
	{
	"first": "Masashi",
	"middle": [],
	"last": "Toyoda",
	"suffix": ""
	}
	],
	"year": 2013,
	"venue": "ACL (1)",
	"volume": "",
	"issue": "",
	"pages": "964--972",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Takayuki Hasegawa, Nobuhiro Kaji, Naoki Yoshinaga, and Masashi Toyoda. 2013. Predicting and eliciting addressee's emotion in online dialogue. In ACL (1), pages 964-972.",
	"links": null
	},
	"BIBREF16": {
	"ref_id": "b16",
	"title": "Sentence-level emotion classification with label and context dependence",
	"authors": [
	{
	"first": "Shoushan",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "Lei",
	"middle": [],
	"last": "Huang",
	"suffix": ""
	},
	{
	"first": "Rong",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "Guodong",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
	"volume": "1",
	"issue": "",
	"pages": "1045--1053",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Shoushan Li, Lei Huang, Rong Wang, and Guodong Zhou. 2015. Sentence-level emotion classification with label and context dependence. In Proceedings of the 53rd Annual Meeting of the Association for Com- putational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1045-1053, Beijing, China, July. Association for Computational Linguistics.",
	"links": null
	},
	"BIBREF17": {
	"ref_id": "b17",
	"title": "Wordnet: a lexical database for english",
	"authors": [
	{
	"first": "A",
	"middle": [],
	"last": "George",
	"suffix": ""
	},
	{
	"first": "",
	"middle": [],
	"last": "Miller",
	"suffix": ""
	}
	],
	"year": 1995,
	"venue": "Communications of the ACM",
	"volume": "38",
	"issue": "11",
	"pages": "39--41",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41.",
	"links": null
	},
	"BIBREF18": {
	"ref_id": "b18",
	"title": "Portable features for classifying emotional text",
	"authors": [
	{
	"first": "Saif",
	"middle": [],
	"last": "Mohammad",
	"suffix": ""
	}
	],
	"year": 2012,
	"venue": "Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
	"volume": "",
	"issue": "",
	"pages": "587--591",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Saif Mohammad. 2012. Portable features for classifying emotional text. In Proceedings of the 2012 Confer- ence of the North American Chapter of the Associa- tion for Computational Linguistics: Human Language Technologies, pages 587-591. Association for Compu- tational Linguistics.",
	"links": null
	},
	"BIBREF19": {
	"ref_id": "b19",
	"title": "A general psychoevolutionary theory of emotion",
	"authors": [
	{
	"first": "Robert",
	"middle": [],
	"last": "Plutchik",
	"suffix": ""
	}
	],
	"year": 1980,
	"venue": "Theories of emotion",
	"volume": "1",
	"issue": "",
	"pages": "3--31",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Robert Plutchik. 1980. A general psychoevolutionary theory of emotion. Theories of emotion, 1:3-31.",
	"links": null
	},
	"BIBREF20": {
	"ref_id": "b20",
	"title": "Semeval-2007 task 14: Affective text",
	"authors": [
	{
	"first": "Carlo",
	"middle": [],
	"last": "Strapparava",
	"suffix": ""
	},
	{
	"first": "Rada",
	"middle": [],
	"last": "Mihalcea",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Proceedings of the 4th International Workshop on Semantic Evaluations",
	"volume": "",
	"issue": "",
	"pages": "70--74",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Carlo Strapparava and Rada Mihalcea. 2007. Semeval- 2007 task 14: Affective text. In Proceedings of the 4th International Workshop on Semantic Evaluations, pages 70-74. Association for Computational Linguis- tics.",
	"links": null
	},
	"BIBREF21": {
	"ref_id": "b21",
	"title": "Emotion detection in codeswitching texts via bilingual and sentimental information",
	"authors": [
	{
	"first": "Zhongqing",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "Sophia",
	"middle": [],
	"last": "Lee",
	"suffix": ""
	},
	{
	"first": "Shoushan",
	"middle": [],
	"last": "Li",
	"suffix": ""
	},
	{
	"first": "Guodong",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing",
	"volume": "2",
	"issue": "",
	"pages": "763--768",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhongqing Wang, Sophia Lee, Shoushan Li, and Guodong Zhou. 2015. Emotion detection in code- switching texts via bilingual and sentimental informa- tion. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 763-768, Beijing, China, July. Association for Computational Linguistics.",
	"links": null
	},
	"BIBREF22": {
	"ref_id": "b22",
	"title": "Verbs semantics and lexical selection",
	"authors": [
	{
	"first": "Zhibiao",
	"middle": [],
	"last": "Wu",
	"suffix": ""
	},
	{
	"first": "Martha",
	"middle": [],
	"last": "Palmer",
	"suffix": ""
	}
	],
	"year": 1994,
	"venue": "Proceedings of the 32nd annual meeting on Association for Computational Linguistics",
	"volume": "",
	"issue": "",
	"pages": "133--138",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Zhibiao Wu and Martha Palmer. 1994. Verbs semantics and lexical selection. In Proceedings of the 32nd an- nual meeting on Association for Computational Lin- guistics, pages 133-138. Association for Computa- tional Linguistics.",
	"links": null
	},
	"BIBREF23": {
	"ref_id": "b23",
	"title": "Emotion classification using web blog corpora",
	"authors": [
	{
	"first": "Changhua",
	"middle": [],
	"last": "Yang",
	"suffix": ""
	},
	{
	"first": "Kevin",
	"middle": [],
	"last": "Hsin-Yih Lin",
	"suffix": ""
	},
	{
	"first": "Hsin-Hsi",
	"middle": [],
	"last": "Chen",
	"suffix": ""
	}
	],
	"year": 2007,
	"venue": "Web Intelligence, IEEE/WIC/ACM International Conference on",
	"volume": "",
	"issue": "",
	"pages": "275--278",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Changhua Yang, Kevin Hsin-Yih Lin, and Hsin-Hsi Chen. 2007. Emotion classification using web blog corpora. In Web Intelligence, IEEE/WIC/ACM Inter- national Conference on, pages 275-278. IEEE.",
	"links": null
	},
	"BIBREF24": {
	"ref_id": "b24",
	"title": "Predicting valence-arousal ratings of words using a weighted graph method",
	"authors": [
	{
	"first": "Liang-Chih",
	"middle": [],
	"last": "Yu",
	"suffix": ""
	},
	{
	"first": "Jin",
	"middle": [],
	"last": "Wang",
	"suffix": ""
	},
	{
	"first": "Robert",
	"middle": [],
	"last": "Lai",
	"suffix": ""
	},
	{
	"first": "Xue-Jie",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	}
	],
	"year": 2015,
	"venue": "",
	"volume": "2",
	"issue": "",
	"pages": "",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Liang-Chih Yu, Jin Wang, K Robert Lai, and Xue-jie Zhang. 2015. Predicting valence-arousal ratings of words using a weighted graph method. Volume 2: Short Papers, page 788.",
	"links": null
	},
	"BIBREF25": {
	"ref_id": "b25",
	"title": "Multilabel neural networks with applications to functional genomics and text categorization. Knowledge and Data Engineering",
	"authors": [
	{
	"first": "Min-Ling",
	"middle": [],
	"last": "Zhang",
	"suffix": ""
	},
	{
	"first": "Zhi-Hua",
	"middle": [],
	"last": "Zhou",
	"suffix": ""
	}
	],
	"year": 2006,
	"venue": "IEEE Transactions on",
	"volume": "18",
	"issue": "10",
	"pages": "1338--1351",
	"other_ids": {},
	"num": null,
	"urls": [],
	"raw_text": "Min-Ling Zhang and Zhi-Hua Zhou. 2006. Multilabel neural networks with applications to functional ge- nomics and text categorization. Knowledge and Data Engineering, IEEE Transactions on, 18(10):1338- 1351.",
	"links": null
	}
	},
	"ref_entries": {
	"FIGREF0": {
	"uris": null,
	"type_str": "figure",
	"text": "Plutchik's basic emotion and dyads -image taken from http://twinklet8.blogspot.jp",
	"num": null
	},
	"FIGREF1": {
	"uris": null,
	"type_str": "figure",
	"text": "vector of some words. (J-S: Joy-Sadness, T-D: Trust-Disgust, F-A: Fear-Anger, S-A: Surprise-Anticipation )",
	"num": null
	},
	"FIGREF2": {
	"uris": null,
	"type_str": "figure",
	"text": "Adapting lexicon for emotional state e j",
	"num": null
	},
	"FIGREF3": {
	"uris": null,
	"type_str": "figure",
	"text": "Extracting tendency featuresEach utterance in the dataset is presented by the following compact set of 22 features: 1. The sum vector of the current utterance which suggest the local tendency.",
	"num": null
	},
	"FIGREF4": {
	"uris": null,
	"type_str": "figure",
	"text": "Structure of the Deep Network",
	"num": null
	},
	"FIGREF5": {
	"uris": null,
	"type_str": "figure",
	"text": "Evaluation of the system vs four baselines: 1) Human Annotators, 2) The system without lexicon adaptation, 3) DBPNN 4) RAkEL",
	"num": null
	},
	"TABREF1": {
	"html": null,
	"text": "Kappa Agreement score.",
	"type_str": "table",
	"content": "<table/>",
	"num": null
	},
	"TABREF2": {
	"html": null,
	"text": "Wordnet expansion.",
	"type_str": "table",
	"content": "<table/>",
	"num": null
	},
	"TABREF5": {
	"html": null,
	"text": "Accuracy and F1 for each emotion class.",
	"type_str": "table",
	"content": "<table/>",
	"num": null
	}
	}
	}
	}