File size: 18,124 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 |
{
"paper_id": "A92-1043",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T02:03:54.309451Z"
},
"title": "Learning a Scanning Understanding for \"Real-world\" Library Categorization",
"authors": [
{
"first": "Stefan",
"middle": [],
"last": "Wermter",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "University of Hamburg",
"location": {
"postCode": "2000",
"settlement": "Hamburg 50",
"country": "Federal Republic of Germany"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "This paper describes, compares, and evaluates three different approaches for learning a semantic classification of library titles: 1) syntactically condensed titles, 2) complete titles, and 3) titles without insignificant words are used for learning the classification in connectionist recurrent plausibility networks. In particular, we demonstrate in this paper that automatically derived feature representations and recurrent plausibility networks can scale up to several thousand library titles and reach almost perfect classification accuracy (>98%) compared to a real-world library classification.",
"pdf_parse": {
"paper_id": "A92-1043",
"_pdf_hash": "",
"abstract": [
{
"text": "This paper describes, compares, and evaluates three different approaches for learning a semantic classification of library titles: 1) syntactically condensed titles, 2) complete titles, and 3) titles without insignificant words are used for learning the classification in connectionist recurrent plausibility networks. In particular, we demonstrate in this paper that automatically derived feature representations and recurrent plausibility networks can scale up to several thousand library titles and reach almost perfect classification accuracy (>98%) compared to a real-world library classification.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Our goal is to examine hybrid symbolic/connectionist and connectionist approaches for classifying a substantial number of real-world title phrases. These approaches are embedded in the framework of SCAN (Wermter 92), a Symbolic Connectionist Approach for Natural language phrases, aimed towards a scanning understanding of natural language rather than focusing on an in-depth understanding. For our experiments we took an existing classification from the online catalog of the main library at Dortmund University and as a first subclassification we selected titles from three classes: \"computer science\" (CS), \"history/politics\" (HP), and \"materials/geology\" (MG).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "The first approach used syntactic condensation based on a chart parser and a headnoun extractor. The symbolic chart parser built a syntactic structure for a title using a context-free grammar and a syntactic lexicon. Then the headnoun extractor retrieved the sequence of headnouns for building a compound noun. For instance, the compound noun \"software access guidelines\" was generated from \"guidelines on subject access to microcomputer software\". This headnoun extractor was motivated *This research was supported in part by the Federal Secretary for Research and Technology under contract #01IV101AO and by the Computer Science Department of Dortmund University. by the close relationship between noun phrases and compound nouns and by the importance of nouns as content words (Finin 80).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Symbolic Syntactic Condensation",
"sec_num": "2.1"
},
{
"text": "Each noun in a compound noun was represented with 16 binary manually encoded semantic features, like measuring-event, changing-event, scientific-field, property, mechanism etc. The set of semantic features had been developed as a noun representation for a related scientific technical domain and had been used for structural disambiguation (Wermter 89). The first approach contained a relatively small set of 76 titles since for each noun 16 features had to be determined manually and for each word in the title the syntactic category had to be in the lexicon which contained 900 entries.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Symbolic Syntactic Condensation",
"sec_num": "2.1"
},
{
"text": "In our second approach, we used an automatically acquired significance vector for each word based on the occurrence of the words in certain classes. Each value v(w, ci) in a significance vector represented the frequency of occurrence of word w in class ci divided by the total frequency of word w in all classes. These significance vectors were computed for the words of 2493 library titles from the three classes CS, HP, and MG.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Unrestricted Complete Phrases",
"sec_num": "2.2"
},
{
"text": "In the third approach we analyzed the most frequent words in the 2493 titles of the second approach. We eliminated words that occured more than five times in our corpus and that were prepositions, conjunctions, articles, and pronouns. Words were represented with the same significance vectors as in the second approach. This elimination of frequently occuring domain-independent words was expected to make classification easier since many domain-independent insignificant words were removed from the titles.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Elimination of Insignificant Words",
"sec_num": "2.3"
},
{
"text": "The Architecture of the Recurrent Plausibility Network",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "3",
"sec_num": null
},
{
"text": "The semantic classification was learned by using a connectionist recurrent plausibility network. A recurrent plausibility network is similar to a simple recurrent network (Elman 89) but instead of learning to predict words, recurrent connections support the assignment of plausible classes (see figure 1) . The recurrent plausibility network was trained in a supervised mode using the backpropagation learning algorithm (Rumelhart et al. 86) . In each training step the feature representation of a word and its preceding context was presented to the network in the word bank and context bank together with the desired class. A unit in the output layer received the value 1 if the unit represented the particular class of the title, otherwise the unit received the value 0. The real-valued hidden layer represented the context of preceding words. At the beginning of a title the context bank was initialized with values of 0 since there was no preceding context. After the first word had been presented the context bank was initialized with the values of the hidden layer that encoded the reduced preceding context. ",
"cite_spans": [
{
"start": 420,
"end": 441,
"text": "(Rumelhart et al. 86)",
"ref_id": null
}
],
"ref_spans": [
{
"start": 295,
"end": 304,
"text": "figure 1)",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "3",
"sec_num": null
},
{
"text": "For the first approach the 76 titles were divided into 61 titles for training and 15 titles for testing. For the 2493 titles in the second and third approach we used 1249 titles for training and 1244 for testing. Using these training and test sets we examined different network architectures and training parameters. For the first approach a configuration with 6 hidden units and a learning rate of 0.001 showed the smallest number of errors on the training and test set. For the second and third approach 3 hidden units and the learning rate 0.0001 performed best.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Conclusions",
"sec_num": "4"
},
{
"text": "Below we show example titles, the titles after preprocessing, and their sequential class assignment. The first two titles illustrate that two titles with the same final headnoun (\"design\") are assigned to different classes due to their different learned preceding context. The third title illustrates the second approach of classifying an unrestricted complete phrase. The network first assigns the CS class for the initial phrase \"On the operating experience of the...\" since such initial representations have occurred in the CS class. However, when more specific knowledge is available (\"doppler sodar system...\") the assigned class is changed to the MG class. In the fourth example the same title is shown for the third approach which eliminates insignificant domainindependent words. In general, the second and third approach have the potential to deal with unanticipated grammatical and even ungrammatical titles since they do not rely on a predefined grammar.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Conclusions",
"sec_num": "4"
},
{
"text": "1. Title: Design of relational database schemes by deleting attributes in the canonical decomposition; Approachl: Compound noun: Decomposition (CS) attribute (CS) scheme (CS) design (CS) 2. Title: Design of bulkheads for controlling water in underground mines; Approach1: Compound noun: Mine (MG) water (MG) bulkhead (MG) design (MG) 3. Title: On the operating experience of the doppler sodar system at the Forschungszentrum Juelich; Approach2: Unrestricted complete title: On (CS) the (CS) operating (CS) experience (CS) of (CS) the (CS) doppler (MG) sodar (MG) system (MG) at (MG) the (MG) Forschungszentrum (MG) Juelich (MG) 4. Title: On the operating experience Of the doppler sodar system at the Forschungszentrum Juelich; Approach3: Unrestricted reduced title: operating (CS) experience (CS) doppler (MG) sodar (MG) system (MG) Forschungszentrum (MG) Juelich (MG)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Conclusions",
"sec_num": "4"
},
{
"text": "The overall performance of the three approaches as recorded in the best found configuration is summarized in table 1. The first approach performed worst for classifying new titles from the test set although the titles in the training set were learned completely. The second approach performed better on the test set for a much bigger training and test set of unrestricted phrases. The third approach demonstrated that the elimination of insignificant words from unrestricted phrases can improve performance for the big set of titles.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Conclusions",
"sec_num": "4"
},
{
"text": "Performance Approach1 Approach2 Approach3 Training 100% 98.4% 99.9% Testing",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Results and Conclusions",
"sec_num": "4"
},
{
"text": "97.7% 99.4% Table 1 : Performance for Semantic Classification",
"cite_spans": [],
"ref_spans": [
{
"start": 12,
"end": 19,
"text": "Table 1",
"ref_id": null
}
],
"eq_spans": [],
"section": "93%",
"sec_num": null
},
{
"text": "In conclusion, we described and evaluated three different approaches for semantic classification which use hybrid symbolic/connectionist and connectionist representations. Our results show that recurrent plausibility networks and automatically acquired feature representations can provide an efficient basis for learning and generalizing a scanning understanding of real-world library classifications.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "93%",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Structured representations and connectionist models. Proceedings o] the Eleventh Annual Conference o] the Cognitive Science Society, Ann Arbor. Finin T.W. 1980. The semantic Interpretation of Compound Nominals",
"authors": [
{
"first": "J",
"middle": [
"L"
],
"last": "Elman",
"suffix": ""
}
],
"year": 1989,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Elman J.L. 1989. Structured representations and connection- ist models. Proceedings o] the Eleventh Annual Conference o] the Cognitive Science Society, Ann Arbor. Finin T.W. 1980. The semantic Interpretation of Com- pound Nominals. PhD Thesis. University of Illinois at Urbana-Champaign.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Learning Internal Representations by Error Propagation",
"authors": [
{
"first": "D",
"middle": [
"E"
],
"last": "Rumelhart",
"suffix": ""
},
{
"first": "G",
"middle": [
"E"
],
"last": "Hinton",
"suffix": ""
},
{
"first": "R",
"middle": [
"J"
],
"last": "Williams",
"suffix": ""
}
],
"year": 1986,
"venue": "Integration of Semantic and Syntactic Constraints for Structural Noun Phrase Disambiguation",
"volume": "1",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Rumelhart D.E., Hinton G.E., Williams R.J. 1986. Learn- ing Internal Representations by Error Propagation. In: Rumelhart D.E., McClelland J.L. (Eds.) Parallel distributed Processing Vol. 1. MIT Press, Cambridge, MA. Wermter, S. 1989. Integration of Semantic and Syntac- tic Constraints for Structural Noun Phrase Disambiguation.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Eleventh International Joint Conlerence on Artificial Intelligence, Detroit. Wermter, S. 1992 (forthcoming). Scanning Understanding: A Symbolic Connectionist Approach .for Natural Language Phrases",
"authors": [],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Proceedings o] the Eleventh International Joint Conlerence on Artificial Intelligence, Detroit. Wermter, S. 1992 (forthcoming). Scanning Understand- ing: A Symbolic Connectionist Approach .for Natural Lan- guage Phrases. Technical Report. University of Hamburg.",
"links": null
}
},
"ref_entries": {
"FIGREF1": {
"uris": null,
"num": null,
"text": "Recurrent Plausibility Network for Titles",
"type_str": "figure"
}
}
}
} |