{
"paper_id": "M91-1014",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T03:15:21.027942Z"
},
"title": "PRC PAKTUS : MUC-3 TEST RESULTS AND ANALYSI S",
"authors": [
{
"first": "Cheryl",
"middle": [],
"last": "Kariya",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "PRC Inc",
"location": {
"addrLine": "1500 Planning Research Drive McLean",
"postCode": "22102",
"region": "VA"
}
},
"email": "kariya_cheryl@po.gis.prc.com"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "For natural language understanding researchers at PRC, preparing for MUC-3 was a usefu l endeavor in many respects. The value of having a well-defined task, a large corpus, and a n automated scoring program can hardly be overestimated. This exercise has pointed out the strengths of our system, confirmed our feelings about which aspects most need work, and taugh t us how to approach a task of this magnitude efficiently. In this paper, we will discuss our test results, the ways in which we prepared for the task, and lessons learned .",
"pdf_parse": {
"paper_id": "M91-1014",
"_pdf_hash": "",
"abstract": [
{
"text": "For natural language understanding researchers at PRC, preparing for MUC-3 was a usefu l endeavor in many respects. The value of having a well-defined task, a large corpus, and a n automated scoring program can hardly be overestimated. This exercise has pointed out the strengths of our system, confirmed our feelings about which aspects most need work, and taugh t us how to approach a task of this magnitude efficiently. In this paper, we will discuss our test results, the ways in which we prepared for the task, and lessons learned .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Overall, we were pleased with PAKTUS's capabilities and performance, and we ar e excited about planned improvements to the system . It was particularly satisfying to observe that only a few changes had to be made to the grammar, and that the lexicon structure and tools easil y accomodated new additions .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RESULT S",
"sec_num": null
},
{
"text": "It came as no surprise to us that the discourse module was the component most in need o f development, as it is the most recent and least-developed part of the system . Pieces for a fairl y major restructuring are already in place, and discourse development will be one of our main effort s in the coming year. To a large extent, our overgeneration of templates was due to the unfinishe d state of the discourse component. Currently, PAKTUS proposes one template for each sentenc e which contains an interesting-looking incident. In discourse processing, the incident templates are merged for the same incident type, if the date and location are not missing and not different . This was a temporary solution designed to prevent wild overgeneration . As we achieve greater recall in the other slots, those fills can be used in a more sophisticated merging strategy ; for the moment, a simple strategy was all that was feasible . The main limiting factor for PRC was the availability of people for development . CPU cycles was also a factor, insofar as it limited the number of development tests that could be run in a given amount of time . It took from 4 to 6 hours to process 100 documents, depending on trac e options .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "RESULT S",
"sec_num": null
},
{
"text": "We directed most of our energies to linguistic development, as shown in Table 2 below , and except for the discourse component, the linguistic aspects of the task have essentially been completed. Because we had less time remaining to devote to engineering issues, much of th e information that PAKTUS produced has not yet been used in template-filling . In using the linguistic output to fill templates, we began with the two aspects of the template-filling task w e considered the most important: identification of relevant incidents (template id) and identificatio n of incident type . Our recall and precision scores for those slots reflect the amount of time we spent on them (see Table 1 ) . We believe that, given more time to convert linguistic output to templat e fills, we will be able to achieve comparable scores for the other slots as well .",
"cite_spans": [],
"ref_spans": [
{
"start": 72,
"end": 79,
"text": "Table 2",
"ref_id": "TABREF1"
},
{
"start": 685,
"end": 692,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "RESULT S",
"sec_num": null
},
{
"text": "Five PRC researchers participated in MUC-3 development . ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "TIME SPENT ON DEVELOPMEN T",
"sec_num": null
},
{
"text": "In adapting PAKTUS (PRC's Adaptive Knowledge-based Text Understanding System) to any new domain, adjustments to each of the core system components must be made . We briefly discuss the types and amounts of adjustments that were necessary in each component, and outlin e some of the new features that were added as we worked on MUC-3 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "PAKTUS DEVELOPMEN T",
"sec_num": null
},
{
"text": "As for any new message type, a new template specifying the format of the input stream had to be built, so that the input could be separated into messages . Methods for handling corpusspecific use of punctuation --for example, [square] brackets and double dashes --were developed . Header information (time and location) was saved for later processing . Also, because time (date of incident) played such an important role in the MUC-3 task, a new feature was added to th e preprocessor which bracketed time expressions in the text, calculated a date, and passed it on to th e parser as a symbol . In general, adaptations to the preprocessor are only a small proportion of th e entire development effort .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Preprocessor",
"sec_num": null
},
{
"text": "Adding the domain-specific lexicon required a considerable amount of time . This is primarily due to the volume of new items to be entered . Especially time-consuming was identifying the terrorist organizations (due to variation in names, acronyms, translations) and their associated countries. PAKTUS's automated tools for entering words, synonyms and compounds , together with feature inheritance in the lexical and semantic networks, made actually enterin g individual items easy . Because the MUC-3 corpus contains large numbers of similar items (e .g. , locations, terrorist organizations), facilities for batch-entering words were developed . In addition to adding lexical items, a few features had to be added to categories of words ; for example, a slot was added to words in the PERSON category to aid in classifying the TYPE of human target . Again, inheritance in the network made this process straightforward and fast.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexicon",
"sec_num": null
},
{
"text": "Another new feature for the lexicon is the use of heuristics for guessing at unknown word s whose roots are unknown . (PAKTUS uses inflectional and derivational morphology if the root s erg in the lexicon .) Word regularities are identified using forward and reverse concordances of th e MUC-3 corpus, and exceptions to those regularities are entered in the lexicon . An example of a heuristic used for MUC-3 is the guess that an unknown word ending in \"-z\" or \"-o\" is a Spanis h name . We have thus far developed about two dozen heuristics.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Lexicon",
"sec_num": null
},
{
"text": "The only significant change made to the core grammar for MUC-3 involved tim e expressions . A new arc was added to handle bracketed time expressions as adverbs, and the caseframe-applier was also adjusted. Some additional work was done on apposition, but this was no t corpus-specific development . Altogether the grammar was well-suited to handle a MUC-3-lik e task, and changes were minimal .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Grammar",
"sec_num": null
},
{
"text": "PAKTUS's discourse component is currently fairly application-specific . Many of the patterns which identify important information (for this task) had to be written from scratch , although a few were recycled from previous applications, such as MUC-2 . The discourse component, both for MUC-3 and in general, is, as we said above, the least-developed aspect o f PAKTUS, and the area in which we expect to make the most improvement in the coming year . One new feature already implemented is the addition of \"word patterns,\" which use only lexical information (but include conceptual associations), to supplement the discourse patterns, which us e parse output . These word patterns accounted for roughly one-third of our recall .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discourse",
"sec_num": null
},
{
"text": "We put some effort into developing a pattern-based filter, to be used at the pre-parse stage , for identifying relevant sentences and/or documents . This filter was not used in the tests, but w e expect to complete development and implement it for MUC-4 . In addition, we plan to expand th e bracketing capabilities to include at least names, titles, and locations. The major thrusts in the coming year, however, will be : -to modify the discourse component to include more broadly-based, linguistic information; and -to develop more tools for analyzing our output .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Future direction s",
"sec_num": null
},
{
"text": "One of the best aspects of the MUC-3 corpus is its generality . Nearly all of th e development we did for this effort can be reused for another application . The exceptions would b e some domain-specific lexicon and some of the specific discourse patterns .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Reusabilit y",
"sec_num": null
},
{
"text": "We relied almost exclusively on the 100 messages in Test 1 for training after the February interim conference, as it had a reliable, consistent key . Using Test 1, we ran approximately 30 tests. A few of these were run to determine the effect of different timing strategies on output; the others, to test improvements in slot-filling from using lexical patterns, partial parses, run-ons, etc . The development corpus was used primarily for lexical development . Two large tests were run o n the entire development corpus to locate a few coding errors and measure linguistic performance.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "TRAINING AND IMPROVEMEN T",
"sec_num": null
},
{
"text": "The improvement in PAKTUS's linguistic performance between February and May 199 1 can be seen in the following tables, derived from the test runs on the development corpus . Total time (%) 100 17.7 34.5 47 .9 N/A Avg time for parse 6.3 2.4 9.5 11 .1 N/A (in seconds) Avg time for preprocessor 2.5 2.1 2.9 3 .0 N/A (in seconds)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "TRAINING AND IMPROVEMEN T",
"sec_num": null
},
{
"text": "The most significant changes is that the number and percentage of sentences that were full y parsed went up nearly 8% between February and May. This reflects changes in all PAKTUS components except the discourse components . Comparison between PRC's scores for the Test 1 and Test 2 reflects the discourse improvements as well .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "TRAINING AND IMPROVEMEN T",
"sec_num": null
},
{
"text": "As PAKTUS had never before been exercised on a scale as large as the MUC-3 corpus , this was an opportunity to find out how it would hold up in a life-sized scenario . We learned that the system architecture, knowledge representation and algorithms were more than adequate for th e task, that it was possible to do all we needed to do, with no major changes . Further, we discovered just how robust PAKTUS is . We were pleasantly surprised, for example, tha t PAKTUS was able to parse 47% of the sentences in the development corpus, particularly sinc e relatively little time was spent on grammar development for MUC-3 .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WHAT WE LEARNE D About PAKTU S",
"sec_num": null
},
{
"text": "The MUC-3 corpus and automatic scoring tool made it possible for us to do extensiv e experimentation on PAKTUS's timer. We discovered several points in the parsing where speed could be improved, while losing little or no important information .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WHAT WE LEARNE D About PAKTU S",
"sec_num": null
},
{
"text": "Although we had been working on extensions to the preprocessor (e .g., document/sentence filtering, word patterns) before the MUC-3 Phase 1 conference, participation in MUC-3 helpe d crystallize our ideas on how best to use those extensions . Further work showed us tha t preprocessor output could be used in many profitable ways (e .g., helping to resolve anaphora) , without compromising linguistic principles .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WHAT WE LEARNE D About PAKTU S",
"sec_num": null
},
{
"text": "Finally, as mentioned above, our participation in MUC-3 gave added impetus to the development of an improved discourse component .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "WHAT WE LEARNE D About PAKTU S",
"sec_num": null
},
{
"text": "Working on MUC-3 provided the PRC team of researchers valuable experience in largescale system development . Basically, we learned what needed to be done and how to divide th e work among ourselves efficiently . We also, albeit somewhat belatedly, learned to make maximu m use of the tools provided us . For example, rather late in the game, it surprised us that we had no t thought to use the development corpus as a source of information about perpetrator organizations !",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About the task",
"sec_num": null
},
{
"text": "Task design : Complete specification of a complex task to be performed is a non-trivial undertaking, requiring multiple iterations to resolve outstanding issues . MUC-3 has clearl y demonstrated that, in spite of excellent initial task design, unforeseen issues inevitably arise, and an ongoing mechanism to provide clarification is indispensable for such complex problems .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About evaluation",
"sec_num": null
},
{
"text": "Training: A valid evaluation must be representative of the domain and type of text o n which the system was trained. The training corpus must therefore be sufficiently large and varie d to cover most of the issues which are likely to arise in a test . (This is one of many areas where MUC-3 was vastly superior to MUC-2 .) Furthermore, for complex tasks such as MUC-3, it i s critical to have an authority available to judge the correctness of a system's response to the trainin g corpus. The development corpus keys, to which each site contributed 100 messages' worth , provided such an \"authority\" . The process of manually generating keys pointed out many issue s which required task clarification, and thus reduced the risk of misinterpreting some aspects of th e task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About evaluation",
"sec_num": null
},
{
"text": "Test Set: Selection of test messages should be done, as it was in MUC-3, from the corpu s before development begins . How to ensure that the test set is representative of the corpus , linguistically and in terms of content, is a question which it may be useful to address for MUC-4 , as is the issue of the appropriate size of a test set. The \"answers\" to the test set should be produce d independently of the test, by an unbiased party (j the system developers). Agreement by all participants to abide by the decisions of the unbiased party must be obtained before the test, an d must be adhered to after the test .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About evaluation",
"sec_num": null
},
{
"text": "Metrics and Scoring : The automatic scoring program is a valuable tool for bot h development and testing . It ensures a base level of impartiality in scoring, although that impartiality is mitigated by the fact that participants could assign themselves credit interactively , possibly using different standards . A possible solution, ultimately adopted for MUC-3, is to have all results scored blindly by an unbiased party .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About evaluation",
"sec_num": null
},
{
"text": "The MUC-3 scoring program has been useful not only as a means for standardization i n testing, but also as a development tool . The fact that different metrics (e,g ., recall, precision , overgeneration) were provided on a slot-by-slot basis gave us a clear picture of our progress . For viewing test results, this means of reporting highlights the strengths and weaknesses of differen t aspects of the systems, rather than providing a single numeric score which masks the details . It thus provides a larger, fairer picture of the systems than would otherwise be possible .",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "About evaluation",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {},
"ref_entries": {
"FIGREF0": {
"text": "------------------------------------------------------------------------",
"type_str": "figure",
"num": null,
"uris": null
},
"TABREF0": {
"content": "
SLOT | REC | PRE | OVG | FAL |
template-id | 88 | 61 | 39 | |
incident-date | 51 | 60 | 0 | |
incident-type | 69 | 78 | 0 | 1 |
category | 68 | 58 | 24 | 30 |
indiv-perps | 3 | 30 | 1 0 | |
org-perps | 1 | 100 | 0 | |
perp-confidence | 20 | 13 | 47 | 1 2 |
phys-target-ids | 9 | 32 | 1 2 | |
phys-target-num | 26 | 66 | 6 | |
phys-target-types | 10 | 67 | 0 | 0 |
human-target-ids | 2 | 42 | 0 | |
human-target-num | 4 | 100 | 0 | |
human-target-types | 2 | 70 | 0 | 0 |
target-nationality | 6 | 100 | 0 | 0 |
instrument-types | 0 | * | * | 0 |
incident-location | 47 | 53 | 0 | |
phys-effects | 18 | 24 | 41 | 2 |
human-effects | 5 | 21 | 14 | 1 |
------------------------- | | | | |
MATCHED ONLY | 32 | 53 | 20 | |
MATCHED/MISSING | 28 | 53 | 20 | |
ALL TEMPLATES | 28 | 36 | 46 | |
SET FILLS ONL Y | 28 | 47 | 24 | 1 |
Table 1 : PRC Scores | | | | |
",
"type_str": "table",
"text": "PRC's scores for MUC-3 Phase 2 .",
"num": null,
"html": null
},
"TABREF1": {
"content": "",
"type_str": "table",
"text": "",
"num": null,
"html": null
},
"TABREF2": {
"content": "",
"type_str": "table",
"text": "Breakdown of MUC-related hour s",
"num": null,
"html": null
}
}
}
}