{ "paper_id": "I08-1034", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:42:31.814468Z" }, "title": "The Telling Tail: Signals of Success in Electronic Negotiation Texts", "authors": [ { "first": "Marina", "middle": [], "last": "Sokolova", "suffix": "", "affiliation": { "laboratory": "IRO", "institution": "Universit\u00e9 de Montr\u00e9al Montr\u00e9al", "location": { "region": "Qu\u00e9bec", "country": "Canada" } }, "email": "sokolovm@iro.umontreal.ca" }, { "first": "Vivi", "middle": [], "last": "Nastase", "suffix": "", "affiliation": { "laboratory": "EML Research, gGmbH", "institution": "", "location": { "settlement": "Heidelberg", "country": "Germany" } }, "email": "nastase@eml-research.de" }, { "first": "Stan", "middle": [], "last": "Szpakowicz", "suffix": "", "affiliation": { "laboratory": "", "institution": "University of Ottawa", "location": { "settlement": "Ottawa", "region": "Ontario", "country": "Canada, Poland" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "We analyze the linguistic behaviour of participants in bilateral electronic negotiations, and discover that particular language characteristics are in contrast with face-to-face negotiations. Language patterns in the later part of electronic negotiation are highly indicative of the successful or unsuccessful outcome of the process, whereas in face-toface negotiations, the first part of the negotiation is more useful for predicting the outcome. We formulate our problem in terms of text classification on negotiation segments of different sizes. The data are represented by a variety of linguistic features that capture the gist of the discussion: negotiationor strategy-related words. We show that, as we consider ever smaller final segments of a negotiation transcript, the negotiationrelated words become more indicative of the negotiation outcome, and give predictions with higher Accuracy than larger segments from the beginning of the process.", "pdf_parse": { "paper_id": "I08-1034", "_pdf_hash": "", "abstract": [ { "text": "We analyze the linguistic behaviour of participants in bilateral electronic negotiations, and discover that particular language characteristics are in contrast with face-to-face negotiations. Language patterns in the later part of electronic negotiation are highly indicative of the successful or unsuccessful outcome of the process, whereas in face-toface negotiations, the first part of the negotiation is more useful for predicting the outcome. We formulate our problem in terms of text classification on negotiation segments of different sizes. The data are represented by a variety of linguistic features that capture the gist of the discussion: negotiationor strategy-related words. We show that, as we consider ever smaller final segments of a negotiation transcript, the negotiationrelated words become more indicative of the negotiation outcome, and give predictions with higher Accuracy than larger segments from the beginning of the process.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "We use language every day to convince, explain, manipulate and thus reach our goals. This aspect of language use is even more obvious in the context of negotiations. The parties must reach an agreement on the partitioning or sharing of a resource, while each party usually wants to leave the negotiation table with the larger piece of the pie. These tendencies become stronger when negotiators use only electronic means to communicate, that is to say, participate in electronic negotiations. In face-to-face contact, prosody and body language often have a crucial role in conveying attitudes and feelings. Enegotiators, on the other hand, must rely only on texts. We perform automatic analysis of the textual data in e-negotiations. We identify linguistic expressions of such negotiation-specific behaviour that are indicative of the final outcome of the process -success or failure -and observe how powerful a tool language is in helping people get what they want.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper we focus on the negotiation as an ongoing process. We analyze the linguistic features of messages exchanged at various points in the course of the negotiation, to determine the time frame in which the outcome becomes decided. From our experimental point of view, we determine the segment of the negotiation which is most predictive of the outcome. There is an imposed three-week deadline in the electronic negotiations that we analyze. We hypothesize that the pressure of the deadline is reflected in the messages exchanged. The messages written later in the process are more indicative of the outcome of the process. Our empirical results support this hypothesis; an analysis of the linguistic features that make this prediction possible shows what the negotiators' main concerns are as the deadline draws near.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Here is what our results contribute to the field of text analysis. Research on text records of faceto-face negotiations suggests that the language patterns used in the first half of a negotiation predict the negotiation outcome better than those in the second half (Simons, 1993) . The explanation was that in the first phase people establish contact, exchange personal information and engage in general polite conversation, creating a foundation of trust between partners. No numerical data, however, supported this diagnosis, and there was no distinction between the prediction of successful and unsuccessful outcomes. When it comes to text classification, our hypothesis says that the classification of the second parts of e-negotiation texts is more accurate with respect to the outcome than the classification of the first parts. This makes e-negotiation texts different from newsgroup messages, newspaper articles and other documents classified by Blatak et al. (2004) , where texts showe d better classification Accuracy on their initial parts. We report the results of several sets of Machine Learning (ML) experiments. Performed on varying-size text data segments, they support our hypothesis.", "cite_spans": [ { "start": 265, "end": 279, "text": "(Simons, 1993)", "ref_id": "BIBREF12" }, { "start": 954, "end": 974, "text": "Blatak et al. (2004)", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We worked with a collection of transcripts of negotiations conducted over the Internet using the Web-based negotiation support system Inspire (Kersten and Zhang, 2003) . Kersten and Zhang (2003) and Nastase (2006) classified e-negotiation outcomes using non-textual data. Classification based on texts is discussed in Sokolova and Szpakowicz, 2006) . None of those experiments considered segmenting the data, although Sokolova and Szpakowicz (2006) analyzed the importance of the first part of e-negotiations. The work we present here is the first attempt to investigate the effect of parts of e-negotiation textual data on classification quality. In this study we do not report types of expressions that are relevant to success and failure of negotiations. These expressions have been presented and analyzed in .", "cite_spans": [ { "start": 142, "end": 167, "text": "(Kersten and Zhang, 2003)", "ref_id": "BIBREF6" }, { "start": 170, "end": 194, "text": "Kersten and Zhang (2003)", "ref_id": "BIBREF6" }, { "start": 199, "end": 213, "text": "Nastase (2006)", "ref_id": "BIBREF8" }, { "start": 318, "end": 348, "text": "Sokolova and Szpakowicz, 2006)", "ref_id": "BIBREF14" }, { "start": 418, "end": 448, "text": "Sokolova and Szpakowicz (2006)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In section 2 we take a brief look at other work on the connection between behaviour and language. In section 3 we present our data and their representation for ML experiments, and we further motivate our work. Section 4 describes the experiments. We discuss the results in Section 5. Section 6 draws conclusions and discusses a few ideas for future work.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "2 Background Review Young (1991) discusses the theory that the situation in which language is used affects the way in which it is used. This theory was illustrated with a particular example of academic speech.", "cite_spans": [ { "start": 20, "end": 32, "text": "Young (1991)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The field of neuro-linguistic programming investigates how to program our language (among other things) to achieve a goal. In the 1980s, Rodger Bailey developed the Language and Behaviour Profile based on 60 meta-programs. Charvet (1997) presents a simplified approach with 14 metaprograms. This profile proposes that people's language patterns are indicators of behavioural preferences. In the study of planning dialogues (Chu-Carroll and Carberry, 2000), Searle's theory of speech acts used through the discourse analysis also supports the fact that language carries much of people's behaviour and emotions. Reitter and Moore (2007) studied repetitions in task-oriented conversations. They demonstrated that a speaker's shortterm ability to copy the interlocutor's syntax is autonomous from the success of the task, whereas long-term adaptation varies with such success.", "cite_spans": [ { "start": 223, "end": 237, "text": "Charvet (1997)", "ref_id": "BIBREF3" }, { "start": 610, "end": 634, "text": "Reitter and Moore (2007)", "ref_id": "BIBREF11" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We consider a negotiation to be a communication in which the participants want to reach an agreement relative to the splitting/sharing of resources. Language is one of the tools used to reach the goal. We propose that not all messages exchanged throughout a negotiation have the same effect on the negotiation outcome. To test this hypothesis, we take an ever smaller segment of the negotiation, and see how well we can predict the outcome of the process, based only on the messages in this fragment.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We encountered several challenges in predicting e-negotiation outcomes using the messages exchanged. First, electronic negotiations usually do not have a sequential-stage model of behaviour (Koeszegi et al, 2007) , which is common in faceto-face negotiations (Adair and Brett, 2005) . Here is an example of behavioural phases in face-to-face negotiations: Perform Relational Positioning \u2192 Identify the Problem \u2192 Generate Solutions \u2192 Reach Agreement. Unexpected turns and moves -typical of human behaviour -make prediction of the negotiation outcome difficult. In case of electronic negotiation, the absence of the usual negotiation structure further complicates the outcome prediction. This distinguishes e-negotiations from agentcustomer phone conversations studied in (Takeuchi et al, 2007) , where an agent follows the call flow pre-defined by his company's policy.", "cite_spans": [ { "start": 190, "end": 212, "text": "(Koeszegi et al, 2007)", "ref_id": "BIBREF7" }, { "start": 259, "end": 282, "text": "(Adair and Brett, 2005)", "ref_id": "BIBREF0" }, { "start": 770, "end": 792, "text": "(Takeuchi et al, 2007)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The longer an e-negotiation takes, the more elaborate the structure of the e-negotiation process becomes. Simpler e-negotiation may involve an exchange of well-structured business documents such as pre-defined contract or retail transactions. A more complex process comprises numerous offers and counter-offers and has a high degree of uncertainty because of the possible unpredictability of negotiation moves.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The next challenge stems from the limitations imposed by the use of electronic means. This overloads text messages with various tasks: negotiation issues themselves, introductions and closures traditional in negotiations, and even socializing. On the other hand, electronic means make the contacts less formal, allowing people to communicate more freely. As a result, the data have a high volume of informality such as abbreviations or slang.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "The last challenge is specific to text analysis. Enegotiations usually involve a multi-cultural audience of varied background, many of whom are not native English speakers. While communicating in English, they introduce a fair amount of spelling and grammatical mistakes.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Participants in a negotiation assume well-defined roles (such as buyers/sellers in some business negotiations, or facilitators in legal disputes), have goals, and adopt specific behaviour to achieve those goals (Koeszegi et al, 2007) . These circumstances are reflected in the language of texts exchanged in negotiations, and distinguish this type of texts from casual e-mail exchange and postings on discussion groups and chat boards. We claim that the language captured in e-negotiation textual data changes as a negotiation progresses, and that this is clearly detectable, even though it does not follow a sequential-stage model common in face-to-face-negotiations (Adair and Brett, 2005) or an agent-customer interaction call flow recommended by a company (Takeuchi et al, 2007) . To support the language change hypothesis, we have conducted a series of ML experiments on negotiation segments of varying size and position, using the largest available data of electronic negotiations.", "cite_spans": [ { "start": 211, "end": 233, "text": "(Koeszegi et al, 2007)", "ref_id": "BIBREF7" }, { "start": 668, "end": 691, "text": "(Adair and Brett, 2005)", "ref_id": "BIBREF0" }, { "start": 760, "end": 782, "text": "(Takeuchi et al, 2007)", "ref_id": "BIBREF15" } ], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "Our data come from the Web-based negoti-ation support system Inspire. Inspire has been used in business courses to teach students about e-negotiations and give them a chance to practice bilateral business negotiations conducted in a lightly controlled environment. For many users, conducting negotiations has been a business/ course assignment. Other users wanted to develop their English skills by participating in an Inspire-enabled negotiation. A negotiation would last up to three weeks, after which, if an agreement has not been reached, the systems would terminate the negotiation and record it as unsuccessful. The following is an example of a negotiation message (with the original spelling): Among the wealth of data gathered by Inspire, we have focussed on the accompanying text messages, extracted from the transcripts of 2557 negotiations. Each negotiation had two different participants, and one person participated in only one negotiation. The total number of contributors was over 5000; most of them were not native English speakers. The data contain 1, 514, 623 word tokens and 27, 055 types. Compared with benchmark corpora, for example the Brown or the Wall Street Journal corpus (Francis and Kucera, 1997; Paul and Baker, 1992) , this collection has a lower type-token ratio and a higher presence of content words among the most frequent words (this is typical of texts on a specific topic), and a high frequency of singular first-and second-person pronouns (this is typical of dialogues).", "cite_spans": [ { "start": 1198, "end": 1224, "text": "(Francis and Kucera, 1997;", "ref_id": null }, { "start": 1225, "end": 1246, "text": "Paul and Baker, 1992)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "Dear", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "We considered all messages from one negotiation to be a single negotiation text. We concatenated the messages in chronological order, keeping the punctuation and spelling unedited. Each negotiation had a unique label, either positive or negative, and was a training example in one of two classes -success- ful or unsuccessful. Inspire assigned a negotiation to the right class automatically. 55% of negotiations in our data set were successful, i.e. ended up with agreement.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "We represented a complete negotiation, or text as we consider it, as a combined bag of words. We matched the tokens in the messages with an inventory of domains from Longman Dictionary of Contemporary English (Procter, 1978) . This allowed us to select those terms that refer to negotiation specific issues -we call them negotiation-related words. We select strategic words based on words and patterns that literature shows to express the intentions, influence, self-obligations and motivations of the negotiation participants. In classifying successful and unsuccessful negotiations, subsets of these two types of features provided better Accuracy than statistically selected features, e.g. most frequent unigrams and unigrams with a higher log-likelihood values calculated between positive and negative classes .", "cite_spans": [ { "start": 209, "end": 224, "text": "(Procter, 1978)", "ref_id": "BIBREF10" } ], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "We halved each text, that is to say, the complete record of a negotiation. For each half we built a bag of 123 negotiation-related words -more on this in section 4. The binary attributes represented the presence or absence of the word in its half of the text. We concatenated the two bags, and labelled the resulting bag by the outcome of the whole negotiation: positive if the negotiation was successful, negative otherwise. We repeated this procedure for the split of the negotiation text into 3 4 and 1 4 . Our ML tools were Weka's (Witten and Frank, 2005) NAIVE BAYES (NB), the sequential minimal optimization (SVM) version of SUPPORT VECTOR MACHINE, and DECISION TREE (DT). In Table 1 we report Accuracy and Precision (P ), Recall (R) and F \u2212 score (F ). P , R, F are calculated on the positive class. For every classifier, the best Accuracy and corresponding P , R, F are reporte d; we performed an exhaustive search on adjustable parameters; the evaluation method was tenfold cross-validation. Our Accuracy results are comparable with those reported in previous studies (Kersten and Zhang, 2003; Nastase, 2006; Sokolova and Szpakowicz, 2006) .", "cite_spans": [ { "start": 535, "end": 559, "text": "(Witten and Frank, 2005)", "ref_id": "BIBREF16" }, { "start": 1077, "end": 1102, "text": "(Kersten and Zhang, 2003;", "ref_id": "BIBREF6" }, { "start": 1103, "end": 1117, "text": "Nastase, 2006;", "ref_id": "BIBREF8" }, { "start": 1118, "end": 1148, "text": "Sokolova and Szpakowicz, 2006)", "ref_id": "BIBREF14" } ], "ref_spans": [ { "start": 682, "end": 689, "text": "Table 1", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "We used the paired t-test to generalize the results on both splits. 1 The two-tailed P value was 0.0102. By conventional criteria, this difference is considered to be statistically significant.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "Accuracy and, especially, Precision results show that DECISION TREE is sensitive to the positions of words in different parts of the negotiations. SUPPORT VECTOR MACHINE and NAIVE BAYES change Accuracy only slightly. The Precision and Recall results give a better picture of the performance. The presence/absence of words recorded for different splits of negotiations influences the identification of true positive examples (successful negotiations) and true negative examples (unsuccessful negotiations). Recall displays that DT classifies successful negotiations better when the negotiations are split 1 2 and 1 2 . Precision and Recall together imply that unsuccessful negotiations have a higher rate of true classification achieved by NB, when the split is 3 4 and 1 4 . This split lets us improve the worst rates of true classifications -unsuccessful negotiations for DT and successful negotiations for NB. Generally, the unequal split al lows us to reduce the difference between true positive and true negative classification results, and thus makes the classification of negotiations more balanced than the equal split. For all the three classifiers, Accuracy and F \u2212 score are better on the 3 4 and 1 4 split.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Textual Data in Electronic Negotiations", "sec_num": "3" }, { "text": "We wanted to determine the placement of the segment of a negotiation most important in deciding whether the outcome is positive: at the beginning or at the end of the process. To do that, we split each negotiation in half, and built two parallel data sets, corresponding to the two halves. We classified each part using various ML tools. Next, we repeated the same classification tasks using smaller and smaller final segments, in order to monitor the variation in performance. Thus each negotiation text N consisted of the head segment (h) and the tail segment (t): N = h t, h t = \u2205, where |t| = |N | i and t was the segment at the end of N , and |h| = (i\u22121)|N | i covering the beginning of the negotiation. We stopped when for two consecutive splits two classifiers had better Accuracy on the head than on the tail. Each segment got the same class label as the whole negotiation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "For these experiments, as briefly explained in section 3, we took the textual negotiation data represented as bags of words. Because of the large number of word features (27, 055 tokens), we performed lexical feature selection.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "Statistical analysis of the corpus built from the Inspire negotiation messages has revealed that the issues discussed in these messages can be grouped into a small set of topics. The particular topic or domain to which a word belongs derives from the most frequent bigram and trigram meanings; for instance, the second most frequent trigram with the word delivery is payment upon delivery, so we assign delivery to the domain negotiation process. The data come from negotiations on a specific topic (sale/purchase of bicycle parts), so a likely candidate subset would be words related to it. We select such negotiation-related words as the first set of features. We show a text sample with the negotiation-related words in bold:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "Dear Georg, I hope you are doig well. I send you this message to ask you what happened to our offer. Just be aware that we will not be indifinitely waiting on your response. As I told you during our last meeting, Itex Manufacturing needs a partnership. So it is important to me to know if you are ready to negotiate with us. We can not afford losing so much precious time. We give you now five more days to answer our offer (1st of december 1997, 2400 Swiss time). After this dead line, we will propose our services to your concurrence. Strategies which the negotiators adopt (promises, threats, exchange of information, argumentation, and so on) affect the outcome (Sokolova and Szpakowicz, 2006) . Since the messages are dense, short and grammatically simple, the expression of strategies through language is straightforward and concentrates on communicating the main goal. The word categories that convey negotiators' strategies are modals, personal pronouns, volition verbs, mental verbs; we refer to them as strategic words. Strategic words constitute the second set of features. Our text sample with strategic words in bold looks as follows:", "cite_spans": [ { "start": 666, "end": 697, "text": "(Sokolova and Szpakowicz, 2006)", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "Dear Georg, I hope you are doig well. I send you this message to ask you what happened to our offer. Just be aware that we will not be indifinitely waiting on your response. As I told you during our last meeting, Itex Manufacturing needs a partnership. So it is important to me to know if you are ready to negotiate with us. We can not afford losing so much precious time. We give you now five more days to answer our offer (1st of december 1997, 2400 Swiss time). After this dead line, we will propose our services to your concurrence. I still believe in a good partnership and relationship between our two societies. Let me know if you think so.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "For Itex Manufacturing. Rookie.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "We work with kernel (SVM), decision-based (DT) and probabilistic (NB) classifiers. Applying classifiers with different working paradigms allow us to capture and understand different aspects of the data, as the results and our discussion in section 5 will show. For each classifier, we used tenfold crossvalidation and exhaustive search on adjustable parameters in model selection. The best results, in particular with high overall Accuracy, appear in Figure 1.", "cite_spans": [], "ref_spans": [ { "start": 451, "end": 457, "text": "Figure", "ref_id": null } ], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "When the data are represented using negotiationrelated words, the tail segments give more accurate outcome classification than the head segments. This holds for all splits and all classifiers; see Figure 1 . The increase in Accuracy when the head segments grow was to be expected, although it does not happen with DT and SVM -only with NB. At the same time, there is no monotonic decline in Accuracy when the length of the tail segments decreases. On the contrary, NB constantly improves the Accuracy of the classification. We note the fact that NB increases the Accuracy on both head and tail segments and makes the basic assumption of the conditional independence of features. We explain the NB results by the decreased dependence between the pres- ence/absence of negotiation-related words when the negotiations move to the second part of the process. The results on the strategic-word representation are slightly different for the three classifiers; see Table 2 : The Accuracy of the negotiation outcome classification on 2 and 3 splits of the second half of the negotiation -the tail segment. Classifying all negotiations as successful gives a baseline Accuracy of 55%. Figure 1 . SVM classifies all tail segments better than head segments, DT classifies tail segments better than head segments up to the 4 5 / 1 5 split, and NB classifies the tail segment better than the head segment only for the half-and-half split. The Accuracy results are unstable for all three classifiers, with the Accuracy on the head segments decreasing when the segments grow and the Accuracy on the tail segments increasing when the tail segments shrink. The performance of the classifiers indicate that, as the deadline approaches, negotiation-related words reflect the negotiation process better than strategic words.", "cite_spans": [], "ref_spans": [ { "start": 197, "end": 205, "text": "Figure 1", "ref_id": "FIGREF2" }, { "start": 958, "end": 965, "text": "Table 2", "ref_id": null }, { "start": 1175, "end": 1183, "text": "Figure 1", "ref_id": "FIGREF2" } ], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "To investigate which part of the tail segments is more important for classifying the outcomes, we introduced additional splits in the tail segments. We divided the second half of each text into 2 and 3 parts and repeated the classification procedures for every new split. The results appear at the top of Table 2 , where tail shows the classification results when the second half of the text was classified, and the other columns report the results on the tail splits; both splits satisfy the conditions tail = i s i , where", "cite_spans": [], "ref_spans": [ { "start": 305, "end": 312, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "s i s j = \u2205 for every i = j.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "The results show that adding splits in the tail segments emphasizes the importance of the last part of a negotiation. For negotiation-related word representation, the classification of the outcome on the last part of the tail is more accurate than on its other parts. This holds for all three classifiers. For the strategic-word representation the same is true for SVM and partially for DT, but not for NB; see the bottom of Table 2 . NB classifies the negotiation outcomes more accurately on s 1 than on s 2 and on s 2 rather than s 3 . ", "cite_spans": [], "ref_spans": [ { "start": 425, "end": 432, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "The Empirical Set-up", "sec_num": "4" }, { "text": "Taking into account the results reported in section 4, we chose negotiation-related words as the feature set. We selected for further analysis the half that performed better for a majority of the tools used. We focussed on the last part of the negotiation, and we extracted a gradually smaller fragment ( 1 2 \u2212 1 9 ; 9 is the average number of text messages in one negotiation). Figure 2 plots the results of the experiments performed with decreasing segment sizes. As we see, the tail segment of the length 1 7 gave a decline of the Accuracy for SVM and NB, with a slight improvement on smaller tail segments.", "cite_spans": [], "ref_spans": [ { "start": 379, "end": 387, "text": "Figure 2", "ref_id": "FIGREF3" } ], "eq_spans": [], "section": "Segmentation Results", "sec_num": "5" }, { "text": "A more detailed analysis comes from considering the Precision and Recall results on the segments; see Table 3 . On 1 7 and 1 9 tail segments a higher Precision indicates that all classifiers have improved the identification of true negatives (unsuccessful negotiations). This means that the trends in the class of unsuccessful negotiations become more noticeable for the classifiers when the deadline approaches. The 1 8 split is an exception, with the abrupt drop of true negative classification by DE-CISION TREE. The correct classification of positive examples (successful negotiations), however, diminishes when splits become smaller; this applies to the performance of all three classifiers. This means that at the end of the negotiations the class of successful negotiations becomes more diverse and, subsequently, multi-modal, and the trends are more difficult to capture by the classifiers.", "cite_spans": [], "ref_spans": [ { "start": 102, "end": 109, "text": "Table 3", "ref_id": "TABREF4" } ], "eq_spans": [], "section": "Segmentation Results", "sec_num": "5" }, { "text": "As in the previous experiments, NB's Accuracy on the tail segments is higher than on the complete data. The opposite is true for SVM and DT: their Accuracy on the tail segments is lower than on the complete data. We explain this by the fact that the sizes of tail segments in the last splits do not give these two classifiers sufficient information.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Segmentation Results", "sec_num": "5" }, { "text": "We have analyzed textual messages exchanged in the course of electronic negotiations. The results support our hypothesis that texts of electronic negotiation have different characteristics than records of face-to-face negotiation. In particular, messages exchanged later in the process are more informative with regard to the negotiation outcome than messages exchanged at the beginning.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Future Work", "sec_num": "6" }, { "text": "We represented textual records of negotiations by two types of word features. These features capture the important aspects of the negotiation process -negotiation-related concepts and indicators of the strategies employed. We performed extensive experiments with different types of ML algorithms and segments of varying sizes from the beginning and the end of the negotiation, on a collection of over 2500 electronic negotiations. Our study shows that words expressing negotiation-related concepts are more useful for distinguishing successful and failed negotiations, especially towards the end of negotiations. We also have shown that there is no linear dependency between the segment sizes and Accuracy of classification of the negotiation success and failure.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Future Work", "sec_num": "6" }, { "text": "Our research plans include a continuation of the investigation of the negotiators' behaviour in electronic negotiations and its reflection in language. To see whether dialogue analysis improves prediction of the negotiation outcomes, we will look at negotiations as dialogues between participants and take into account their roles, e.g. buyer and seller. We will split a negotiation at message boundaries to avoid arbitrary splits of the negotiation process.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Discussion and Future Work", "sec_num": "6" }, { "text": "Results on the same data require the paired version of t-test.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [ { "text": "Partial support for this work came from the Natural Sciences and Engineering Research Council of Canada.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgments", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "The negotiation dance: time, culture, and behavioral sequences in negotiation", "authors": [ { "first": "J", "middle": [ "M" ], "last": "Adair", "suffix": "" }, { "first": "", "middle": [], "last": "Brett", "suffix": "" } ], "year": 2005, "venue": "Organization Science", "volume": "16", "issue": "1", "pages": "33--51", "other_ids": {}, "num": null, "urls": [], "raw_text": "Adair, J. M. Brett. 2005. The negotiation dance: time, culture, and behavioral sequences in negotiation. Or- ganization Science. 16(1): 33-51.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Fragments and Text Categorization", "authors": [ { "first": "J", "middle": [], "last": "Blat\u00e1k", "suffix": "" }, { "first": "E", "middle": [], "last": "Mr\u00e1kov\u00e1", "suffix": "" }, { "first": "L", "middle": [], "last": "Popel\u00ednsky", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 42th Annual Meeting of the Association of Computational Linguistics (ACL", "volume": "", "issue": "", "pages": "226--229", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Blat\u00e1k, E. Mr\u00e1kov\u00e1, L. Popel\u00ednsky. 2004. Fragments and Text Categorization. Proceedings of the 42th An- nual Meeting of the Association of Computational Lin- guistics (ACL 2004). Companion Volume: 226-229, Association for Computational Linguistics.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Negotiating Globally", "authors": [ { "first": "J", "middle": [ "M" ], "last": "Brett", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. M. Brett. 2001. Negotiating Globally. Jossey-Bass.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "Words that Change Minds: Mastering the Language of Influence", "authors": [ { "first": "S", "middle": [ "R" ], "last": "Charvet", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. R. Charvet. 1997. Words that Change Minds: Master- ing the Language of Influence. Kendall/Hunt.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Conflict Resolution in Collaborative Planning Dialogues", "authors": [ { "first": "J", "middle": [], "last": "Chu-Carroll", "suffix": "" }, { "first": "S", "middle": [], "last": "Carberry", "suffix": "" } ], "year": 2000, "venue": "International Journal of HumanComputer Studies", "volume": "53", "issue": "6", "pages": "969--1015", "other_ids": {}, "num": null, "urls": [], "raw_text": "J. Chu-Carroll, S. Carberry. 2000. Conflict Resolution in Collaborative Planning Dialogues. International Jour- nal of HumanComputer Studies. 53(6): 969-1015.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Computational Analysis of Present-Day American English", "authors": [ { "first": "W", "middle": [ "N" ], "last": "Francis", "suffix": "" }, { "first": "H", "middle": [], "last": "Ku\u010dera", "suffix": "" } ], "year": 1967, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "W. N. Francis, H. Ku\u010dera. 1967. Computational Analysis of Present-Day American English, Brown University Press.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Mining Inspire Data for the Determinants of Successful Internet Negotiations", "authors": [ { "first": "G", "middle": [ "E" ], "last": "Kersten", "suffix": "" }, { "first": "G", "middle": [], "last": "Zhang", "suffix": "" } ], "year": 2003, "venue": "Central European Journal of Operational Research", "volume": "11", "issue": "3", "pages": "297--316", "other_ids": {}, "num": null, "urls": [], "raw_text": "G. E. Kersten, G. Zhang. 2003. Mining Inspire Data for the Determinants of Successful Internet Negotia- tions. Central European Journal of Operational Re- search. 11(3): 297-316.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Data-driven Episodic Phase Analysis of E-negotiation. Group Decision and Negotiation", "authors": [ { "first": "S", "middle": [], "last": "Koeszegi", "suffix": "" }, { "first": "E.-M", "middle": [], "last": "Pesendorfer", "suffix": "" }, { "first": "R", "middle": [], "last": "Vetschera", "suffix": "" } ], "year": 2007, "venue": "", "volume": "2", "issue": "", "pages": "11--130", "other_ids": {}, "num": null, "urls": [], "raw_text": "S. Koeszegi, E.-M. Pesendorfer, R. Vetschera. 2007. Data-driven Episodic Phase Analysis of E-negotiation. Group Decision and Negotiation 2007. 2: 11-130.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Concession curve analysis for Inspire negotiations. Group Decision and Negotiation", "authors": [ { "first": "V", "middle": [], "last": "Nastase", "suffix": "" } ], "year": 2006, "venue": "", "volume": "15", "issue": "", "pages": "18--193", "other_ids": {}, "num": null, "urls": [], "raw_text": "V. Nastase. 2006. Concession curve analysis for Inspire negotiations. Group Decision and Negotiation. 15: 18-193.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "The Design for the Wall Street Journal-based CSR Corpus", "authors": [ { "first": "D", "middle": [ "B" ], "last": "Paul", "suffix": "" }, { "first": "J", "middle": [ "M" ], "last": "Baker", "suffix": "" } ], "year": 1992, "venue": "Proceedings of the 2nd International Conference on Spoken Language Processing", "volume": "", "issue": "", "pages": "357--361", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. B. Paul and J. M. Baker 1992 The Design for the Wall Street Journal-based CSR Corpus. Proceedings of the 2nd International Conference on Spoken Lan- guage Processing (ICSLP'92), 357-361.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Longman Dictionary of Contemporary English", "authors": [ { "first": "P", "middle": [], "last": "Procter", "suffix": "" } ], "year": 1978, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "P. Procter. 1978. Longman Dictionary of Contemporary English. Longman Group Ltd. Essex, UK.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Predicting Success in Dialogue", "authors": [ { "first": "D", "middle": [], "last": "Reitter", "suffix": "" }, { "first": "J", "middle": [], "last": "Moore", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL 2007", "volume": "", "issue": "", "pages": "808--815", "other_ids": {}, "num": null, "urls": [], "raw_text": "D. Reitter, J. Moore. 2007. Predicting Success in Dia- logue. Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics (ACL 2007), 808-815, Association for Computational Linguistics.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Speech Patterns and the Concept of Utility in Cognitive Maps: the Case of Integrative Bargaining", "authors": [ { "first": "T", "middle": [], "last": "Simons", "suffix": "" } ], "year": 1993, "venue": "Academy of Management Journal", "volume": "38", "issue": "1", "pages": "139--156", "other_ids": {}, "num": null, "urls": [], "raw_text": "T. Simons. 1993. Speech Patterns and the Concept of Utility in Cognitive Maps: the Case of Integrative Bar- gaining. Academy of Management Journal. 38(1): 139-156.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Feature Selection in Electronic Negotiation Texts", "authors": [ { "first": "M", "middle": [], "last": "Sokolova", "suffix": "" }, { "first": "V", "middle": [], "last": "Nastase", "suffix": "" }, { "first": "M", "middle": [], "last": "Shah", "suffix": "" }, { "first": "S", "middle": [], "last": "Szpakowicz", "suffix": "" } ], "year": 2005, "venue": "Proceedings of Recent Advances in Natural Language Processing (RANLP'2005)", "volume": "", "issue": "", "pages": "518--524", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Sokolova, V. Nastase, M. Shah, S. Szpakowicz. 2005. Feature Selection in Electronic Negotiation Texts. Proceedings of Recent Advances in Natural Language Processing (RANLP'2005), 518-524, In- coma Ltd, Bulgaria.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Language Patterns in the Learning of Strategies from Negotiation Texts", "authors": [ { "first": "M", "middle": [], "last": "Sokolova", "suffix": "" }, { "first": "S", "middle": [], "last": "Szpakowicz", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the 19th Canadian Conference on Artificial Intelligence (AI'2006)", "volume": "", "issue": "", "pages": "288--299", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Sokolova, S. Szpakowicz. 2006. Language Patterns in the Learning of Strategies from Negotiation Texts. Proceedings of the 19th Canadian Conference on Ar- tificial Intelligence (AI'2006), 288-299, Springer.", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Automatic Identification of Important Segments and Expressions for Mining of Business-Oriented Conversations at Contact Centers", "authors": [ { "first": "M", "middle": [], "last": "Sokolova", "suffix": "" }, { "first": "S", "middle": [], "last": "Szpakowicz", "suffix": "" }, { "first": "H", "middle": [], "last": "Springer", "suffix": "" }, { "first": "L", "middle": [], "last": "Takeuchi", "suffix": "" }, { "first": "T", "middle": [], "last": "Subramaniam", "suffix": "" }, { "first": "S", "middle": [], "last": "Nasukawa", "suffix": "" }, { "first": "", "middle": [], "last": "Roy", "suffix": "" } ], "year": 2005, "venue": "Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning", "volume": "", "issue": "", "pages": "458--467", "other_ids": {}, "num": null, "urls": [], "raw_text": "M. Sokolova and S. Szpakowicz. 2005 Analysis and Classification of Strategies in Electronic Negotiations. Proceedings of the 18th Canadian Conference on Ar- tificial Intelligence (AI'2005), 145-157, Springer. , H. Takeuchi, L. Subramaniam, T. Nasukawa, S. Roy. 2007 Automatic Identification of Important Segments and Expressions for Mining of Business-Oriented Conversations at Contact Centers. Proceedings of the 2007 Joint Conference on Empirical Methods in Natu- ral Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), 458-467, As- sociation for Computational Linguistics.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Data Mining", "authors": [ { "first": "I", "middle": [], "last": "Witten", "suffix": "" }, { "first": "E", "middle": [], "last": "Frank", "suffix": "" } ], "year": 2005, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "I. Witten, E. Frank. 2005. Data Mining, 2nd ed..", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Language as Behaviour, Language as Code: A Study of Academic English", "authors": [ { "first": "L", "middle": [], "last": "Young", "suffix": "" } ], "year": 1991, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "L. Young. 1991. Language as Behaviour, Language as Code: A Study of Academic English. John Benjamins.", "links": null } }, "ref_entries": { "FIGREF0": { "type_str": "figure", "num": null, "uris": null, "text": "still believe in a good partnership and relationship between our two societies. Let me know if you think so. For Itex Manufacturing. Rookie." }, "FIGREF1": { "type_str": "figure", "num": null, "uris": null, "text": "negotiation-related words tail part, negotiation-related words head part, strategic words tail part, strategic words" }, "FIGREF2": { "type_str": "figure", "num": null, "uris": null, "text": "The classification Accuracy with DT, SVM and NB, for negotiation-related and strategic words." }, "FIGREF3": { "type_str": "figure", "num": null, "uris": null, "text": "The evolution of the success and failure classification Accuracy with decreasing segment sizes." }, "TABREF0": { "num": null, "html": null, "type_str": "table", "content": "", "text": "Georg, I hope you are doig well. I send you this message to ask you what happened to our offer. Just be aware that we will not be indifinitely waiting on your response. As I told you during our last meeting, Itex Manufacturing needs a partnership. So it is important to me to know if you are ready to negotiate with us. We can not afford losing so much precious time. We give you now five more days to answer our offer (1st of december 1997, 2400 Swiss time). After this dead line, will propose our services to your concurrence. I still believe in a good partnership and relationship between our two societies. Let me know if you think so. For Itex Manufacturing. Rookie." }, "TABREF1": { "num": null, "html": null, "type_str": "table", "content": "
FeaturesSplitNBSVMDT
AccFPR AccFPR AccFPR
negotiation-related 1/2 and 1
", "text": "/2 68.1 70.4 73.0 68.0 73.6 76.8 75.4 78.2 73.9 78.8 72.1 86.8 negotiation-related 3/4 and 1/4 69.1 71.3 74.1 68.7 73.7 77.0 75.5 78.5 75.4 79.4 73.8 86.0 Accuracy and corresponding F \u2212 score, Precision and Recall . Classifying all negotiations as successful gives a baseline Accuracy of 55%." }, "TABREF3": { "num": null, "html": null, "type_str": "table", "content": "
Classifier1/31/41/51/61/71/81/9
PRPRPRPRPRPRPR
DT74.2
", "text": "85.3 74.2 84.3 75.2 82.3 73.61 83.0 74.5 82.4 72.1 81.6 74.0 81.3 SVM 76.1 78.1 76.3 76.3 77.0 75.3 78.3 75.3 77.2 73.4 76.9 72.3 77.6 71.6 NB 73.8 71.8 71.8 73.9 74.8 71.9 74.9 72.0 71.3 72.2 70.8 72.5 70.5 74.3" }, "TABREF4": { "num": null, "html": null, "type_str": "table", "content": "", "text": "Precision and Recall on the tail segments; negotiation-related words. Precision and Recall are calculated on the positive class." } } } }