| { | |
| "paper_id": "M91-1009", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T03:15:25.228414Z" | |
| }, | |
| "title": "HUGHES TRAINABLE TEXT SKIMMER : MUC-3 TEST RESULTS AND ANALYSI S", | |
| "authors": [ | |
| { | |
| "first": "Charles", | |
| "middle": [ | |
| "P" | |
| ], | |
| "last": "Dolan", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Seth", | |
| "middle": [ | |
| "R" | |
| ], | |
| "last": "Goldman", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Thomas", | |
| "middle": [ | |
| "V" | |
| ], | |
| "last": "Cuda", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Alan", | |
| "middle": [ | |
| "M" | |
| ], | |
| "last": "Nakamura", | |
| "suffix": "", | |
| "affiliation": {}, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "M91-1009", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "Malibu, CA 90265", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Hughes Research Laboratorie s 3011 Malibu Canyon Road M/S RL9 6", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Test results Figure 1 gives the official results for the Hughes Trainable Text Skimmer used for MUC 3 (TTS-MUC3) . TTS is a largely statistical system, using a K-Nearest Neighbor classifie r with the output of a shallow parser as features. (See the System Summary section of thi s volume for a detailed description of TTS-MUC3) . The performance, on a slot by slot basi s is, therefore, what one might expect : the pure set fills such as \"Incident Type\" and \"Category\" have much better performance than the string fills such as \"Human Target .\" In addition, we can see that \"Incident Date\" and \"Incident Location,\" for which special code was written, have performance above that of the string fills. One calendar month and approximately three (3) person months were spent on MUC3 . Before MUC3, we had constructed a text database facility and the pattern matcher used fo r shallow parsing . Therefore much of the time for MUC3 was spend evaluating alternative s for the statistical engine. Approximately 45% of the time was spent developing code for ideas that were not used in the final system . Of the remaining time, 30% was spen t developing code to extract and format information for MUC3 templates (including code to parse the templates of the DEV corpus), 15% was spent coding and tuning the K-Neares t Neighbor classifier, and 10% was spent creating phrasal patterns, either by hand o r extracting them automatically from the templates for the DEV corpus .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 13, | |
| "end": 21, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Hughes Research Laboratorie s 3011 Malibu Canyon Road M/S RL9 6", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "-----------------------------------", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SLOT REC PRE OVG FAL", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "----------------------------------- MATCHED", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "SLOT REC PRE OVG FAL", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The test settings for TTS-MUC3 were tuned to maximize recall . This resulted in roughly equal recall and precision . Some results in the companion paper in the System Summar y section of this volume indicate that we might tune TTS-MUC3 for higher precision at th e expense of recall . However, we believe that there are enough different algorithms tha t might substantially improve the performance of TTS that evaluating such trade-offs i s premature . For the official test, we used K=12 in the pattern classifier . The pattern classifier returns a set of hypotheses for various set and string fills . The hypotheses are returned with strengths between 0 .0 and 1 .0 which are then compared to a threshold ; all the thresholds on the feature extraction were extremely low (e .g., 0.1).", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Test setting s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The limiting factor for the Hughes TTS-MUC3 system was time . The K-Nearest Neighbo r classifier is surprisingly effective, but there are many variations that we did not have time to try. With a small amount of extra time we could make small improvements there . In addition, we suspect that our algorithm for grouping sentences into topics was responsibl e for many of our errors . However, improving this portion of the system will take muc h more time and, we believe, will require the addition of domain knowledge into the processing . Training", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Limiting factor s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The training regimen was extremely simple . A word frequency analysis was performed on the DEV corpus, and we selected those words that occurred between 10 and 105 times a s our content bearing words, resulting in about 1000 such words . These words were the n grouped by hand into approximately 400 conceptual classes . In addition, words were added to the lexicon for numbers, ordinals, roman numerals, compass points, etc . The lexicon and the DEV templates were used to drive the construction of phrases . Phrases were created from string fills by substituting conceptual classes for words . For example, \"SIX JESUITS\" would drive the creation of the phrase, ( : N UMBER -W : RELEGIOUS-ORDER-W) . The type of the string fill served as the semantic feature fo r the phrase.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Limiting factor s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For some phrases there where conflicts, for example, many phrases that might be mappe d to :ACTIVE -MILITARY as a human target, might also be mapped to :STATE -SPONSORED-VIOLENCE-INDIV as a perpetrating individual . For these phrases, the most frequent usage was chosen . After creating a large number of phrases automaticall y (approximately 1000), a set of hand constructed phrases was added to augment and repai r that set (approximately 200) .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Limiting factor s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "All the stories in the DEV corpus were used to build the case memory, however, the number of cases per different \"Type of Incident\" was limited to 35 . This means that once 35 cases of a particular incident type (i .e., Murder) had been seen, future cases of this type were ignored . This attempt to balance the training data was necessary because the numbe r of stories for each type of incident varied greatly . By restricting the maximum stories per topic, we tended to ignore many of the later stories in the training set.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Limiting factor s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "All the modules in TTS-MUC3 are domain independent. However, all the modules except the date extraction module, require some amount of training . Besides the training describe d above, the location extraction module requires a location database, including what location s contain what other locations . The overhead for constructing such training sets an d databases is quite large, but we feel that for applications of sufficient leverage, good use r interface design will ease the burden of constructing the training set and reduce the time for deploying TTS in new domains . In addition, integration with on-line data sources such a s map databases will eliminate the burden of creating special data files for natural languag e processing .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Domain independent modules", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "num": null, | |
| "uris": null, | |
| "text": "Official TST2 Score report Distribution of labo r", | |
| "type_str": "figure" | |
| } | |
| } | |
| } | |
| } |