{ "paper_id": "I05-1026", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:26:22.057429Z" }, "title": "Using Multiple Discriminant Analysis Approach for Linear Text Segmentation", "authors": [ { "first": "Zhu", "middle": [], "last": "Jingbo", "suffix": "", "affiliation": { "laboratory": "Natural Language Processing Laboratory", "institution": "Northeastern University", "location": { "settlement": "Shenyang", "country": "P.R. China" } }, "email": "zhujingbo@mail.neu.edu.cn" }, { "first": "Ye", "middle": [], "last": "Na", "suffix": "", "affiliation": { "laboratory": "Natural Language Processing Laboratory", "institution": "Northeastern University", "location": { "settlement": "Shenyang", "country": "P.R. China" } }, "email": "yena@ics.neu.edu.cn" }, { "first": "Chang", "middle": [], "last": "Xinzhi", "suffix": "", "affiliation": { "laboratory": "Natural Language Processing Laboratory", "institution": "Northeastern University", "location": { "settlement": "Shenyang", "country": "P.R. China" } }, "email": "changxz@ics.neu.edu.cn" }, { "first": "Chen", "middle": [], "last": "Wenliang", "suffix": "", "affiliation": { "laboratory": "Natural Language Processing Laboratory", "institution": "Northeastern University", "location": { "settlement": "Shenyang", "country": "P.R. China" } }, "email": "chenwl@mail.neu.edu.cn" }, { "first": "Benjamin", "middle": [ "K" ], "last": "Tsou", "suffix": "", "affiliation": { "laboratory": "", "institution": "City University of Hong Kong", "location": { "region": "HK" } }, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Research on linear text segmentation has been an ongoing focus in NLP for the last decade, and it has great potential for a wide range of applications such as document summarization, information retrieval and text understanding. However, for linear text segmentation, there are two critical problems involving automatic boundary detection and automatic determination of the number of segments in a document. In this paper, we propose a new domain-independent statistical model for linear text segmentation. In our model, Multiple Discriminant Analysis (MDA) criterion function is used to achieve global optimization in finding the best segmentation by means of the largest word similarity within a segment and the smallest word similarity between segments. To alleviate the high computational complexity problem introduced by the model, genetic algorithms (GAs) are used. Comparative experimental results show that our method based on MDA criterion functions has achieved higher P k measure (Beeferman) than that of the baseline system using TextTiling algorithm.", "pdf_parse": { "paper_id": "I05-1026", "_pdf_hash": "", "abstract": [ { "text": "Research on linear text segmentation has been an ongoing focus in NLP for the last decade, and it has great potential for a wide range of applications such as document summarization, information retrieval and text understanding. However, for linear text segmentation, there are two critical problems involving automatic boundary detection and automatic determination of the number of segments in a document. In this paper, we propose a new domain-independent statistical model for linear text segmentation. In our model, Multiple Discriminant Analysis (MDA) criterion function is used to achieve global optimization in finding the best segmentation by means of the largest word similarity within a segment and the smallest word similarity between segments. To alleviate the high computational complexity problem introduced by the model, genetic algorithms (GAs) are used. Comparative experimental results show that our method based on MDA criterion functions has achieved higher P k measure (Beeferman) than that of the baseline system using TextTiling algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Typically a document is concerned with more than one subject, and most texts consist of long sequences of paragraphs with very little structural demarcation. The goal of linear text segmentation is to divide a document into topically-coherent sections, each corresponding to a relevant subject. Linear text segmentation has been applied in document summarization, information retrieval, and text understanding. For example, in recent years, passage-retrieval techniques based on linear text segmentation, are becoming increasingly popular in information retrieval as relevant text passages often provide better answers than complete document texts in response to user queries [1] .", "cite_spans": [ { "start": 676, "end": 679, "text": "[1]", "ref_id": "BIBREF0" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In recent years, many techniques have been applied to linear text segmentation. Some have used linguistic information [2, 3, 4, 5, 6, 9] such as cue phrases, punctuation marks, prosodic features, reference, and new words occurrence. Others have used statistical methods [7, 8, 10, 11, 12, 13, 14, 15] such as those based on word cooccurrence, lexical cohesion relations, semantic network, similarity between adjacent parts of texts, similarity between all parts of a text, dynamic programming algorithm, and HMM model.", "cite_spans": [ { "start": 118, "end": 121, "text": "[2,", "ref_id": "BIBREF1" }, { "start": 122, "end": 124, "text": "3,", "ref_id": "BIBREF2" }, { "start": 125, "end": 127, "text": "4,", "ref_id": "BIBREF3" }, { "start": 128, "end": 130, "text": "5,", "ref_id": "BIBREF4" }, { "start": 131, "end": 133, "text": "6,", "ref_id": "BIBREF5" }, { "start": 134, "end": 136, "text": "9]", "ref_id": "BIBREF8" }, { "start": 270, "end": 273, "text": "[7,", "ref_id": "BIBREF6" }, { "start": 274, "end": 276, "text": "8,", "ref_id": "BIBREF7" }, { "start": 277, "end": 280, "text": "10,", "ref_id": "BIBREF9" }, { "start": 281, "end": 284, "text": "11,", "ref_id": "BIBREF10" }, { "start": 285, "end": 288, "text": "12,", "ref_id": "BIBREF11" }, { "start": 289, "end": 292, "text": "13,", "ref_id": "BIBREF12" }, { "start": 293, "end": 296, "text": "14,", "ref_id": "BIBREF13" }, { "start": 297, "end": 300, "text": "15]", "ref_id": "BIBREF14" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In linear text segmentation study, there are two critical problems involving automatic boundary detection and automatic determination of the number of segments in a document. Some efforts have focused on using similarity between adjacent parts of a text to solve topic boundary detection. In fact, the similarity threshold is very hard to set, and it is very difficult to identify exactly topic boundaries only according to similarity between adjacent parts of a text. Other works have focused on the similarity between all parts of a text. Reynar [7] and Choi [13] used dotplots technique to perform linear text segmentation which can be seen as a form of approximate and local optimization. Yaari [16] has used agglomerative clustering to perform hierarchical segmentation. Others [10, 17, 18, 19] used dynamic programming to perform exact and global optimization in which some prior parameters are needed. These parameters can be obtained via uninformative prior probabilities [18] , or estimated from training data [19] .", "cite_spans": [ { "start": 548, "end": 551, "text": "[7]", "ref_id": "BIBREF6" }, { "start": 561, "end": 565, "text": "[13]", "ref_id": "BIBREF12" }, { "start": 699, "end": 703, "text": "[16]", "ref_id": "BIBREF15" }, { "start": 783, "end": 787, "text": "[10,", "ref_id": "BIBREF9" }, { "start": 788, "end": 791, "text": "17,", "ref_id": "BIBREF16" }, { "start": 792, "end": 795, "text": "18,", "ref_id": "BIBREF17" }, { "start": 796, "end": 799, "text": "19]", "ref_id": "BIBREF18" }, { "start": 980, "end": 984, "text": "[18]", "ref_id": "BIBREF17" }, { "start": 1019, "end": 1023, "text": "[19]", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we propose a new statistical model for linear text segmentation, which uses Multiple Discriminant Analysis (MDA) method to define a global criterion function for document segmentation. Our method focuses on within-segment word similarity and between-segment word similarity. This process can achieve global optimization in addressing the two aforementioned problems of linear text segmentation. Our method is domain-independent and does not use any training data.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In section 2, we introduce Multiple Discriminant Analysis (MDA) criterion functions in detail. In section 3, our statistical model of linear text segmentation is proposed. A new MDA criterion function revised by adding penalty factor is further discussed in section 4. Comparative experimental results are given in Section 5. At last, we address conclusions and future work in section 6.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In statistical pattern classification, MDA approach is commonly used to find effective linear transformations [20, 21] . The MDA approach seeks a projection that best separates the data in a least-squares sense. As shown in Figure 1 , using MDA method we could get the greatest separation over data space when average within-class distance is the smallest, and average between-class distance is the largest.", "cite_spans": [ { "start": 110, "end": 114, "text": "[20,", "ref_id": "BIBREF19" }, { "start": 115, "end": 118, "text": "21]", "ref_id": "BIBREF20" } ], "ref_spans": [ { "start": 224, "end": 232, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "Similarly, if we consider a document as data space, and a segment as a class, the basic idea of our approach for linear text segmentation is to find best segmentation of a document(greatest separation over data space) by focusing on within-segment word similarity and between-segment word similarity. It is clear that the smaller the average within-class distance or the average between-class distance, the larger the within-segment word similarity or the between-segment word similarity, and vice versa. In other words, we want to find the best segmentation of a document in which within-segment word similarity is the largest, and between-segment word similarity is the smallest. To achieve this goal, we introduce a criterion function to evaluate the segmentation of a document and assign a score to it. In this paper, we adopt the MDA approach to define a global criterion function of document segmentation, and called as MDA criterion function, which is described below. Fig. 1 . When average within-class distance is the smallest, and average between-class distance is the largest, the greatest separation over data space is shown Let W=w 1 w 2 \u2026w t be a text consisting of t words, and let S=s 1 s 2 \u2026s c be a segmentation of W consisting of c segments. We define W as data space, S as segmentation distribution over data space W. Because the lengths of paragraphs or sentences can be highly irregular, unbalanced comparisons can result in text segmentation process. Thus we adopt the block method that is used in the TextTiling algorithm [2, 3] , but we replace lexical word with block. In our model, we group blocksize words into a block which can be represented by a d-dimensional vector. In practice, we find that the value of blocksize=100 works well for many Chinese documents. Then W =w 1 w 2 \u2026w t can be redefined as B=b 1 b 2 \u2026b k . As illustrated in Figure 1 , a cross point can be defined as a d-dimensional block vector.", "cite_spans": [ { "start": 1546, "end": 1549, "text": "[2,", "ref_id": "BIBREF1" }, { "start": 1550, "end": 1552, "text": "3]", "ref_id": "BIBREF2" } ], "ref_spans": [ { "start": 976, "end": 982, "text": "Fig. 1", "ref_id": null }, { "start": 1867, "end": 1875, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "In this paper, we introduce MDA criterion function J d in the following form [20] ( ) ( ) ( )", "cite_spans": [ { "start": 77, "end": 81, "text": "[20]", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "B d W tr S J s tr S = (1)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "Where tr(A) is the trace of matrix A. S W and S B are within-segment scatter matrix and between-segment scatter matrix, respectively. S W is defined by", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 1 ( )( ) i c t W i i i i bs i S P b m b m n = \u2208 = \u2212 \u2212 \u2211 \u2211", "eq_num": "(2)" } ], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "Where b stands for blocks belonging to segment s i , P i is the a priori probability of segment s i , and is defined to be the ratio of blocks in segment s i divided by the total number of blocks of the document, n i is the number of blocks in the segment s i , m i is the d-dimensional block mean of the segment s i given by", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "1 . i i b s i m b n \u2208 = \u2211 (3) D 1 labeled class \u03c9 1 D 2 labeled class \u03c9 2 D 3 labeled class \u03c9 3", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "Suppose that a total mean vector m is defined by", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "1 1 1 c i i B i m b nm n n = = = \u2211 \u2211 (4)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "In equation 1, between-segment scatter matrix S B is defined by", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 ( )( ) c t B i i i i S P m m m m = = \u2212 \u2212 \u2211", "eq_num": "(5)" } ], "section": "MDA Criterion Function", "sec_num": "2" }, { "text": "Using the same definitions of text W, segmentation S and blocks B in section 2, we first discuss the statistical model for linear text segmentation. The key of statistical model for text segmentation is to find the segmentation with maximum-probability. This can be turned into another task of finding segmentation with highest J d score equally. The most likely segmentation is given by", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "$ arg max P( | ) arg max ( , ) def d S S S S W J WS = =", "eq_num": "(6)" } ], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "As mentioned above, because paragraph or sentence length can be highly irregular, it leads to unbalanced comparisons in text segmentation process. So W =w 1 w 2 \u2026w n could be redefined as B=b 1 b 2 \u2026b k , and the most likely segmentation is given by \u02c6a rg max P( | ) arg max ( , )", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "def d S S S S B J BS = =", "eq_num": "(7)" } ], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "The computational complexity for achieving the above solution is O(2 k ), where k is the number of blocks in a document. To alleviate the high computational complexity problem, we adopt the genetic algorithms (GAs) [22] . GAs provides a learning method motivated by an analogy to biological evolution. Rather than searching from general-to-specific hypotheses, or from simple-to-complex, GAs generate successor hypotheses by repeatedly mutating and recombining parts of the best currently known hypotheses. GAs have most commonly been applied to optimization problems outside machine learning, and are especially suited to tasks in which hypotheses are complex.", "cite_spans": [ { "start": 215, "end": 219, "text": "[22]", "ref_id": "BIBREF21" } ], "ref_spans": [], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "By adopting this methodology, we derive the following text segmentation algorithm, as illustrated in Figure 2 . In this paper, we focus our study on paragraphlevel linear text segmentation, in which the potential boundary mark between segments can be placed only between adjacent paragraphs.", "cite_spans": [], "ref_spans": [ { "start": 101, "end": 109, "text": "Figure 2", "ref_id": "FIGREF0" } ], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "Given a text W and blocks B, K max is the total number of paragraphs in the text. Initialization: S best = {}, J d (B,S best )=0.0 Segmentation:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "For k = 2 to K max Begin 1) Use genetic algorithms and equation 7to find the best segmentation S of k segments. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Statistical Model for Linear Text Segmentation", "sec_num": "3" }, { "text": "In the text segmentation process, adjacent boundary adjustment should be considered in cases when there are some very close adjacent but incorrect segment boundaries. In experiments we find that in these cases some single-sentence paragraphs are wrongly recognized as isolated segments. To solve the problem, we propose a penalty factor (PF) to prevent assignment of very short segment boundaries (such as a single-sentence segment) by adjusting very close adjacent boundaries, and therefore improve the performance of linear text segmentation system.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Penalty Factor", "sec_num": "4" }, { "text": "Suppose that we get a segmentation S=s 1 s 2 \u2026s c of the input document, let L be the length of the document, L i be the length of the segment s i . We know L=L 1 +L 2 +\u2026+L c . We define penalty factor as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Penalty Factor", "sec_num": "4" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 c i i L PF L = = \u220f", "eq_num": "(8)" } ], "section": "Penalty Factor", "sec_num": "4" }, { "text": "As can be seen, short-length segments would result in smaller penalty factor. We use penalty factor to revise the J d scores of segmentations. To incorporate the penalty factor PF, our MDA criterion function J d can be rewritten as", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Penalty Factor", "sec_num": "4" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "1 ( ) ( ) ( ) ( ) c i B d PF d i W L tr S J x PF J x L tr S \u2212 = = \u00d7 = \u00d7 \u220f", "eq_num": "(9)" } ], "section": "Penalty Factor", "sec_num": "4" }, { "text": "In the following experiments, we will evaluate effectiveness of using the two MDA criterion functions J d and J d-PF for linear text segmentation.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Penalty Factor", "sec_num": "4" }, { "text": "Precision and recall statistics are conventional means of evaluating the performance of classification algorithms. For the segmentation task, recall measures the fraction of actual boundaries that an automatic segmenter correctly identifies, and precision measures the fraction of boundaries identified by an automatic segmenter that are actual boundaries. The shortcoming is that every inaccurately estimated segment boundary is penalized equally whether it is near or far from a true segment boundary.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Methods", "sec_num": "5.1" }, { "text": "To overcome the shortcoming of precision and recall, we use a measure called P k , proposed by Beeferman et al. [8] . P k method measures the proportion of sentences which are wrongly predicted to belong in the same segment or sentences which are wrongly predicted to belong in different segments. More formally, given two segmentations ref(true segmentation) and hyp(hypothetical segmentation) for a document of n sentences, P k is formally defined by The function D\u03bc is a distance probability distribution over the set of possible distances between sentences chosen randomly from the document, and will in general depend on certain parameters \u00b5 such as the average spacing between sentences. In equation (10) , D\u03bcwas defined as an exponential distribution with mean 1/\u00b5, a parameter that we fix at the approximate mean document length for the domain [8] .", "cite_spans": [ { "start": 112, "end": 115, "text": "[8]", "ref_id": "BIBREF7" }, { "start": 706, "end": 710, "text": "(10)", "ref_id": "BIBREF9" }, { "start": 852, "end": 855, "text": "[8]", "ref_id": "BIBREF7" } ], "ref_spans": [], "eq_spans": [], "section": "Evaluation Methods", "sec_num": "5.1" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "( , ) i j D i j e \u00b5 \u00b5 \u00b5 \u03b3 \u2212 \u2212 =", "eq_num": "(11)" } ], "section": "Evaluation Methods", "sec_num": "5.1" }, { "text": "Where \u03d2 \u03bc is a normalization chosen so that D\u03bc is a probability distribution over the range of distance it can accept. From the above formulation, we could find one weakness of the metric: there is no principled way of specifying the distance distribution D\u03bc. In the following experiments, we use P k as performance measure, where the mean segment length in the test data was 1/\u03bc=11 sentences.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Evaluation Methods", "sec_num": "5.1" }, { "text": "We mainly focus our work on paragraph-level linear text segmentation techniques. The Hearst's TextTiling algorithm [2, 3] is a simple and domain-independent technique for linear text segmentation, which segments at the paragraph level. Topic boundaries are determined by changes in the sequence of similarity scores. This algorithm uses a simple cutoff function to determine automatically the number of boundaries.", "cite_spans": [ { "start": 115, "end": 118, "text": "[2,", "ref_id": "BIBREF1" }, { "start": 119, "end": 121, "text": "3]", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Quantitative Results", "sec_num": "5.2" }, { "text": "In our experiments, we use the TextTiling algorithm to provide the baseline system, and use the P k measure to evaluate and compare the performance of the TextTiling and our method. Our data set -NEU_TS, is collected manually, and it consists of 100 Chinese documents, all from 2004-2005 Chinese People's Daily newspaper. The number of segments per document varies from five to eight. The average number of paragraphs per document is 25.8 paragraphs. To build the ground truth for NEU_TS data set, five trained graduate students in our laboratory who are working on the analysis of Chinese document are asked to provide judgment on the segmentation of every Chinese document. We first use the toolkit CipSegSDK [23] for document preprocessing, including word segmentation, but with the removal of stopwords from all documents.", "cite_spans": [ { "start": 711, "end": 715, "text": "[23]", "ref_id": "BIBREF22" } ], "ref_spans": [], "eq_spans": [], "section": "Quantitative Results", "sec_num": "5.2" }, { "text": "In the first experiment, we assume the number of segments of an input document is known in advance. We use the NEU_TS data set and the P k measure to evaluate and compare the performance of TextTiling and our method. The purpose of this experiment is to compare the performance of boundary detection techniques of TextTiling algorithm and our model using MDA criterion functions. In the TextTiling algorithm, topic boundaries are determined by changes in the sequence of similarity scores. The boundaries are determined by locating the lowermost portions of valleys in the resulting plot. Therefore, it is not a global evaluation method. However, in our model, MDA criterion function provides a global evaluation method to text segmentation; it selects the best segmentation with the largest within-segment word similarity and the smallest between-segment word similarity. Results shown in Table 1 indicated that our boundary detection techniques based on two MDA criterion functions perform better than the TextTiling algorithm, and MDA criterion function J d-PF works the best. ", "cite_spans": [], "ref_spans": [ { "start": 890, "end": 897, "text": "Table 1", "ref_id": "TABREF0" } ], "eq_spans": [], "section": "1) Experiment 1", "sec_num": null }, { "text": "In this experiment, we assume the number of segments of a document is unknown in advance. In other words, Texttiling algorithm and our model should determine the number of segments of a document automatically. Similar to Experiment 1, the same data set is used and the P k measure is calculated for both TextTiling and our method using MDA criterion functions J d and J d-PF . The comparative results are shown in Table 2 . As mentioned above, how to determine the number of segments to be assigned to a document is a difficult problem. Texttiling algorithm uses a simple cutoff function method to determine the number of segments and it is sensitive to the patterns of similarity scores [2, 3] . The cutoff function is defined as a function of the average and standard deviations of the depth scores for the text under analysis. A boundary is drawn only if the depth score exceeds the cutoff value. We think that the simple cutoff function method is hard to achieve global optimization when solving these two key problems of linear text segmentation process. In our model, two MDA criterion functions J d and J d-PF are used to determine the number of segments and boundary detection by maximizing J d score of segmentations. Once the maximum-score segmentation is found, the number of segments of the document is produced automatically. Experimental results show that our MDA criterion functions are superior to the TextTiling's cutoff function in terms of automatic determination of the number of segments. It is also shown that the MDA criterion function J d-PF revised with Penalty Factor works better than J d . In implementation, we have adopted genetic algorithms (GAs) to alleviate the computational complexity of MDA, and have obtained good results.", "cite_spans": [ { "start": 688, "end": 691, "text": "[2,", "ref_id": "BIBREF1" }, { "start": 692, "end": 694, "text": "3]", "ref_id": "BIBREF2" } ], "ref_spans": [ { "start": 414, "end": 421, "text": "Table 2", "ref_id": "TABREF1" } ], "eq_spans": [], "section": "2) Experiment 2", "sec_num": null }, { "text": "In this paper, we studied and proposed a new domain-independent statistical model for linear text segmentation in which multiple discriminant analysis(MDA) approach is used as global criterion function for document segmentation. We attempted to achieve global optimization in solving the two fundamental problems of text segmentation involving automatic boundary detection and automatic determination of number of segments of a document, by focusing on within-segment word similarity and between-segment word similarity. We also applied genetic algorithms(GAs) to reduce the high computational complexity of MDA based method. Experimental results show that our method based on MDA criterion functions outperforms the TextTiling algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" }, { "text": "The solution to the high computational complexity problem will continue to be studied by using other effective optimization algorithm or near optimal solutions. In the next stage we plan to combine MDA criterion functions with other algorithms such as clustering to improve the performance of our text segmentation system, and apply the text segmentation technique to other text processing task, such as information retrieval and document summarization.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusions and Future Work", "sec_num": "6" } ], "back_matter": [ { "text": "We thank Keh-Yih Su and Matthew Ma for discussions related to this work. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Acknowledgements", "sec_num": null } ], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Automatic text decomposition using text segments and text themes", "authors": [ { "first": "Gerard", "middle": [], "last": "Salton", "suffix": "" }, { "first": "Amit", "middle": [], "last": "Singhal", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Buckley", "suffix": "" }, { "first": "Mandar", "middle": [], "last": "Mitra", "suffix": "" } ], "year": 1996, "venue": "In proceedings of the seventh ACM conference on Hypertext", "volume": "", "issue": "", "pages": "53--65", "other_ids": {}, "num": null, "urls": [], "raw_text": "Gerard Salton, Amit Singhal, Chris Buckley, and Mandar Mitra.: Automatic text decomposition using text segments and text themes. In proceedings of the seventh ACM conference on Hypertext, Bethesda, Maryland, United States (1996) 53-65", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Multi-paragraph segmentation of expository text", "authors": [ { "first": "M", "middle": [ "A" ], "last": "Hearst", "suffix": "" } ], "year": 1994, "venue": "proceedings of the 32th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "9--16", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hearst, M.A.: Multi-paragraph segmentation of expository text. In proceedings of the 32th Annual Meeting of the Association for Computational Linguistics, Las Cruces, New Mexico (1994) 9-16", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "TextTiling: segmenting text into multi-paragraph subtopic passages", "authors": [ { "first": "M", "middle": [ "A" ], "last": "Hearst", "suffix": "" } ], "year": 1997, "venue": "Computational Linguistics", "volume": "23", "issue": "1", "pages": "33--64", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hearst, M.A.: TextTiling: segmenting text into multi-paragraph subtopic passages. Computational Linguistics, Vol.23, No.1 (1997) 33-64", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "A new tool for discourse analysis: The vocabulary management profile. Language", "authors": [ { "first": "G", "middle": [], "last": "Youmans", "suffix": "" } ], "year": 1991, "venue": "", "volume": "67", "issue": "", "pages": "763--789", "other_ids": {}, "num": null, "urls": [], "raw_text": "Youmans, G.: A new tool for discourse analysis: The vocabulary management profile. Language, Vol.67, No.4 (1991) 763-789", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Lexical cohesion computed by thesauri relations as an indicator of the structure of text", "authors": [ { "first": "J", "middle": [], "last": "Morris", "suffix": "" }, { "first": "G", "middle": [], "last": "Hirst", "suffix": "" } ], "year": 1991, "venue": "Computational Linguistics", "volume": "17", "issue": "1", "pages": "21--42", "other_ids": {}, "num": null, "urls": [], "raw_text": "Morris, J. and Hirst, G.: Lexical cohesion computed by thesauri relations as an indicator of the structure of text. Computational Linguistics, Vol.17, No.1 (1991) 21-42", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Text segmentation based on similarity between words", "authors": [ { "first": "H", "middle": [], "last": "Kozima", "suffix": "" } ], "year": 1993, "venue": "proceedings of the 31th Annual Meeting of the Association for Computational Linguistics, Student Session", "volume": "", "issue": "", "pages": "286--288", "other_ids": {}, "num": null, "urls": [], "raw_text": "Kozima, H.: Text segmentation based on similarity between words. In proceedings of the 31th Annual Meeting of the Association for Computational Linguistics, Student Session (1993) 286-288", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "An automatic method of finding topic boundaries", "authors": [ { "first": "J", "middle": [ "C" ], "last": "Reynar", "suffix": "" } ], "year": 1994, "venue": "proceedings of the 32 nd Annual Meeting of the Association for Computational Linguistics, Student Session", "volume": "", "issue": "", "pages": "331--333", "other_ids": {}, "num": null, "urls": [], "raw_text": "Reynar, J.C.: An automatic method of finding topic boundaries. In proceedings of the 32 nd Annual Meeting of the Association for Computational Linguistics, Student Session, Las Cruces, New Mexico (1994) 331-333", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Text segmentation using exponential models", "authors": [ { "first": "D", "middle": [], "last": "Beeferman", "suffix": "" }, { "first": "A", "middle": [], "last": "Berger", "suffix": "" }, { "first": "J", "middle": [], "last": "Lafferty", "suffix": "" } ], "year": 1997, "venue": "proceedings of the Second Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "35--46", "other_ids": {}, "num": null, "urls": [], "raw_text": "Beeferman, D., Berger, A., and Lafferty, J.: Text segmentation using exponential models. In proceedings of the Second Conference on Empirical Methods in Natural Language Processing, pages, Providence, Rhode Island (1997) 35-46", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Intention-based segmentation: Human reliability and correlation with linguistic cues", "authors": [ { "first": "R", "middle": [], "last": "Passoneau", "suffix": "" }, { "first": "D", "middle": [ "J" ], "last": "Litman", "suffix": "" } ], "year": 1993, "venue": "proceedings of the 31st Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "148--155", "other_ids": {}, "num": null, "urls": [], "raw_text": "Passoneau, R. and Litman, D.J.: Intention-based segmentation: Human reliability and correlation with linguistic cues. In proceedings of the 31st Meeting of the Association for Computational Linguistics (1993) 148-155", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Text segmentation by topic", "authors": [ { "first": "M", "middle": [], "last": "Jay", "suffix": "" }, { "first": "", "middle": [], "last": "Ponte", "suffix": "" }, { "first": "W", "middle": [], "last": "Bruce", "suffix": "" }, { "first": "", "middle": [], "last": "Croft", "suffix": "" } ], "year": 1997, "venue": "proceeding of the first European conference on research and advanced technology for digital libraries", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jay M. Ponte and Bruce W. Croft.: Text segmentation by topic. In proceeding of the first European conference on research and advanced technology for digital libraries. U.Mass. Computer Science Technical Report TR97-18 (1997)", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Statistical models for topic segmentation", "authors": [ { "first": "J", "middle": [ "C" ], "last": "Reynar", "suffix": "" } ], "year": 1999, "venue": "proceedings of the 37th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "357--364", "other_ids": {}, "num": null, "urls": [], "raw_text": "Reynar, J.C.: Statistical models for topic segmentation. In proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (1999) 357-364", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Intentional features of local and global discourse", "authors": [ { "first": "J", "middle": [], "last": "Hirschberg", "suffix": "" }, { "first": "B", "middle": [], "last": "Grosz", "suffix": "" } ], "year": 1992, "venue": "proceedings of the Workshop on Spoken Language Systems", "volume": "", "issue": "", "pages": "441--446", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hirschberg, J. and Grosz, B.: Intentional features of local and global discourse. In proceedings of the Workshop on Spoken Language Systems (1992) 441-446", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Advances in domain independent linear text segmentation", "authors": [ { "first": "Y", "middle": [ "Y" ], "last": "Freddy", "suffix": "" }, { "first": "", "middle": [], "last": "Choi", "suffix": "" } ], "year": 2000, "venue": "Proc. of NAACL-2000", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Freddy Y. Y. Choi.: Advances in domain independent linear text segmentation. In Proc. of NAACL-2000 (2000)", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Latent semantic analysis for text segmentation", "authors": [ { "first": "F", "middle": [ "Y Y" ], "last": "Choi", "suffix": "" }, { "first": "P", "middle": [], "last": "Wiemer-Hastings", "suffix": "" }, { "first": "J", "middle": [], "last": "Moore", "suffix": "" } ], "year": 2001, "venue": "proceedings of the 6th Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "109--117", "other_ids": {}, "num": null, "urls": [], "raw_text": "Choi, F.Y.Y., Wiemer-Hastings, P. & Moore, J.: Latent semantic analysis for text segmentation. In proceedings of the 6th Conference on Empirical Methods in Natural Language Processing (2001) 109-117.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Topic segmentation with an aspect hidden Markov model", "authors": [ { "first": "D", "middle": [ "M" ], "last": "Blei", "suffix": "" }, { "first": "P", "middle": [ "J" ], "last": "Moreno", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Blei, D.M. and Moreno, P.J.: Topic segmentation with an aspect hidden Markov model. Tech. Rep. CRL 2001-07, COMPAQ Cambridge Research Lab (2001)", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "Segmentation of expository texts by hierarchical agglomerative clustering", "authors": [ { "first": "Y", "middle": [], "last": "Yaari", "suffix": "" } ], "year": 1997, "venue": "proceedings of the conference on recent advances in natural language processing", "volume": "", "issue": "", "pages": "59--65", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yaari, Y.: Segmentation of expository texts by hierarchical agglomerative clustering. In proceedings of the conference on recent advances in natural language processing (1997) 59-65", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Optimal multi-paragraph text segmentation by dynamic programming", "authors": [ { "first": "O", "middle": [], "last": "Heinonen", "suffix": "" } ], "year": 1998, "venue": "proceedings of 17 th international conference on computational linguistics", "volume": "", "issue": "", "pages": "1484--1486", "other_ids": {}, "num": null, "urls": [], "raw_text": "Heinonen, O.: Optimal multi-paragraph text segmentation by dynamic programming. In proceedings of 17 th international conference on computational linguistics (1998) 1484- 1486.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "A statistical model for domain-independent text segmentation", "authors": [ { "first": "M", "middle": [], "last": "Utiyama", "suffix": "" }, { "first": "H", "middle": [], "last": "Isahara", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "491--498", "other_ids": {}, "num": null, "urls": [], "raw_text": "Utiyama, M., and Isahara, H.: A statistical model for domain-independent text segmentation. In proceedings of the 9 th conference of the European chapter of the association for computational linguistics (2001) 491-498", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Linear Text Segmentation using a Dynamic Programming Algorithm", "authors": [ { "first": "A", "middle": [], "last": "Kehagias", "suffix": "" }, { "first": "", "middle": [], "last": "Fragkou", "suffix": "" }, { "first": "", "middle": [], "last": "Petridis", "suffix": "" } ], "year": 2003, "venue": "proceedings of 10th Conference of European chapter of the association for computational linguistics", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "A Kehagias, P Fragkou, V Petridis.: Linear Text Segmentation using a Dynamic Programming Algorithm. In proceedings of 10th Conference of European chapter of the association for computational linguistics (2003)", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Pattern Classification. Second Edition", "authors": [ { "first": "R", "middle": [], "last": "Duda", "suffix": "" }, { "first": "P", "middle": [], "last": "Hart", "suffix": "" }, { "first": "D", "middle": [], "last": "Stork", "suffix": "" } ], "year": 2001, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "R. Duda, P. Hart, and D. Stork.: Pattern Classification. Second Edition, John Wiley & Sons (2001)", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Pattern recognition principles", "authors": [ { "first": "Julius", "middle": [ "T" ], "last": "Tol", "suffix": "" }, { "first": "Rafael", "middle": [ "C" ], "last": "Gonzaiez", "suffix": "" } ], "year": 1974, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Julius T.Tol and Rafael C. Gonzaiez.: Pattern recognition principles. Addison-Wesley Publishing Company (1974)", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Machine Learning", "authors": [ { "first": "Tom", "middle": [ "M" ], "last": "Mitchell", "suffix": "" } ], "year": 1997, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tom M.Mitchell.: Machine Learning. McGraw-Hill (1997)", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Natural language processingresearch on making computers understand human languages", "authors": [ { "first": "Yao", "middle": [], "last": "Tianshun", "suffix": "" }, { "first": "Zhu", "middle": [], "last": "Jingbo", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Zhang Li", "suffix": "" }, { "first": "", "middle": [], "last": "Ying", "suffix": "" } ], "year": 2002, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Yao Tianshun, Zhu Jingbo, Zhang li, and Yang Ying.: Natural language processing- research on making computers understand human languages. Tsinghua university press (2002)", "links": null } }, "ref_entries": { "FIGREF0": { "text": "If J d (B,S best ) < J d (B,S) Then Begin S best = S and J d (B,S best ) = J d (B,S). Endif Endfor Output the best segmentation S best .", "type_str": "figure", "num": null, "uris": null }, "FIGREF1": { "text": "MDA-based text segmentation algorithm", "type_str": "figure", "num": null, "uris": null }, "FIGREF2": { "text": "Where \u03b4 ref (i,j) is an indicator function whose value is 1 if sentences i and j belong in the same segment in the true segmentation, and 0 otherwise. Similarly, \u03b4 hyp (i,j) is an indicator function which evaluates to 1 if sentences i and j belong in the same segment in the hypothetical segmentation, and 0 otherwise. The operator between \u03b4 ref (i,j) and \u03b4 hyp (i,j) in the above formula is the XNOR function on its two operands.", "type_str": "figure", "num": null, "uris": null }, "TABREF0": { "content": "
MeasureTextTiling algorithmMDA method using J dMDA method using J d-PF
P k value0.8250.8690.905
", "num": null, "text": "P k value with known number of document segments", "type_str": "table", "html": null }, "TABREF1": { "content": "
MeasureTextTiling algorithmMDA method using J dMDA method using J d-PF
P k value0.8080.8310.87
", "num": null, "text": "P k value with unknown number of document segments", "type_str": "table", "html": null } } } }