| { | |
| "paper_id": "M91-1003", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T03:15:37.627050Z" | |
| }, | |
| "title": "COMPARING MUCK-II AND MUC-3 : ASSESSING THE DIFFICULTY OF DIFFERENT TASK S", | |
| "authors": [ | |
| { | |
| "first": "Lynette", | |
| "middle": [], | |
| "last": "Hirschma", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "Spoken Language Systems Grou p MIT Laboratory for Computer Science Cambridge", | |
| "institution": "", | |
| "location": { | |
| "postCode": "0213 9", | |
| "region": "MA" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "", | |
| "pdf_parse": { | |
| "paper_id": "M91-1003", | |
| "_pdf_hash": "", | |
| "abstract": [], | |
| "body_text": [ | |
| { | |
| "text": "The natural language community has made impressive progress in evaluation over the last four years . However, as the evaluations become more sophisticated and more ambitious, a fundamental proble m emerges : how to compare results across changing evaluation paradigms . When we change domain , task, and scoring procedures, as has been the case from MUCK-I to MUCK-II to MUC-3, we los e comparability of results . This makes it difficult to determine whether the field has made progress sinc e the last evaluation . Part of the success of the MUC conferences has been due to the incremental approach taken to system evaluation . Over the four year period of the three conferences, the domain has becom e more \" realistic\", the task has become more ambitious and specified in much greater detail, and th e scoring procedures have evolved to provide a largely automated scoring mechanism . This process has been critical to demonstrating the utility of the overall evaluation process . However we still need som e way to assess overall progress of the field, and thus we need to compare results and task difficulty o f MUC-3 relative to MUCK-II .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "OVERVIE W", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This comparison is complicated by the absence of any generally agreed upon metrics for comparin g the difficulty of two natural language tasks . There is no real analog, for example, to perplexity in th e speech community, which provides a rough cross-task assessment of the difficulty of recognizing speech , given a corpus and some grammar or language model for that corpus . Natural language does not hav e a set of metrics whose affect on task difficulty is well-understood . In the absence of such metrics, thi s paper outlines a set of dimensions that capture some of the differences between the two tasks . It remain s a subject for further research to define appropriate metrics for cross-task comparison and to determin e how these metrics correlate with performance .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "OVERVIE W", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Clearly it is impossible to come up with a single number that characterizes the relative difficulty o f MUCK-II and MUC-3 . Nonetheless, we can characterize both qualitative and quantitative difference s along the following dimensions :", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "OVERVIE W", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "\u2022 Complexity of dat a In general, it is safe to say that MUC-3 was much harder than MUCK-II in all of these dimensions . It is also clear that the scores for MUC-3 were lower (even after adjusting for the difference in how score s * Parse difficulty may increase as the cube of sentenc e length for some parsers Table 1 : Complexity of Dat a were computed) . The best precision and recall figures for MUCK-II were in the range of 70-80% on th e final test (using a corrected score, calculated in the MUC-3 style) . For MUC-3, precision and recall were in the range of 45-65% for the final test . Although the error rates were about double for MUC-3, the tas k was many times harder . From this, we can conclude that the field has indeed made impressive progres s in the two years since MUCK-II .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 312, | |
| "end": 319, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "OVERVIE W", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Complexity of the Dat a", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "DIMENSIONS OF COMPARISO N", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The first respect in which MUCK-II and MUC-3 differ is in the type of messages chosen as input . MUCK-II used Navy message traffic (operational reports) as input, which are filled with jargon, bu t also cover a rather limited domain with a fairly small vocabulary (2000 words) . These messages wer e characterized by a telegraphic syntax and somewhat random punctuation ; run-on sentences, for example , occurred quite frequently. MUC-3 uses news wire reports as input, which are more general (in tha t they use less jargon), but cover a much wider range of subjects, with a much larger vocabulary (20,00 0 words) . Although the syntax is generally more \"standard\" in MUC-3, the richness of the corpus pose s new problems . Sentences can be extremely long and quite rich in their range of syntactic construction s (see Figure 1 for extreme examples of MUCK-II and MUC-3 sentences) .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 819, | |
| "end": 827, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "DIMENSIONS OF COMPARISO N", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In MUCK-II, there were four distinct message types ; in MUC-3, there are sixteen . These includ e \"text\", \"communique\" and \"editorial\" as well as translated transcriptions of radio communiques originall y in Spanish . The average sentence length is longer in MUC-3 (27 words compared to 12 words for MUCK -II), as is message length (12 sentences/message for MUC-3 compared to 3 sentences/message for MUCK -II) . These differences are summarized in Table 1 . Table 1 makes any attempt to measure grammatical complexity . Of th e various measures, perhaps sentence length correlates most closely. In general, longer sentences are harde r to parse ; for certain types of parser, the time to parse increases as the cube of sentence length . Even thi s does not begin to describe the rich range of syntactic constructions in MUC-3 . On the other hand, thi s richness is difficult to compare to the difficulties of handling telegraphic and heavily elliptical materia l in MUCK-II . Figure 1 contains a sentence from MUCK-II and a sentence from MUC-3, to illustrate th e different problems posed by the two corpora .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 448, | |
| "end": 455, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 458, | |
| "end": 465, | |
| "text": "Table 1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 976, | |
| "end": 984, | |
| "text": "Figure 1", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "DIMENSIONS OF COMPARISO N", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "In addition to the metrics reported in Table 1 , there are other metrics which would help to captur e the notion of data complexity . We could measure the rate of growth of the vocabulary, for example , to determine how frequently new words appear ; this would also give us insight into whether we had a", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 39, | |
| "end": 46, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "None of the measures in", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "He disclosed that, in August 1988, he was assigned the missions of planning the assassinatio n of the president of the republic and the beheading of the Honduran spiritual and political leadership through the physical elimination of Honduran Archbishop Msgr Hector Enrique Santo s and presidential candidates Carlo Roberto Flores and Rafael Leonardo Callejas Romero, o f the Liberal and National parties respectively. sufficient amount of training data . We could measure perplexity using a simple grammar (e .g ., a bigram language model) ; this is an information theoretic measure which is standard in the speech recognitio n community and gives an idea of how many words (on average) can follow a given word in the corpus . There are also some system-dependent measures which would be interesting : number of grammar rules used, number of co-occurrence patterns, number of inference rules, and size of the knowledge base . Investigation of what to measure and how these measures relate to other things (e .g ., accuracy, speed ) should be considered a subject of further research in its own right, if we expect to make meaningfu l comparisons across different domains .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "MUC-3", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The size of the MUC-3 corpus has had a profound impact on how the participating message understanding systems were built and debugged . MUC-3 provided one to two orders of magnitude more dat a than MUCK-II : 1300 training messages and some 400,000 words, compared to 105 messages and 3,00 0 words for MUC :-II . If hand-debugging was still (barely) possible in MUCK-II, it was clearly an overwhelming task for MUC-3 . In addition, system throughput became a major consideration in MUC-3 . If it takes a system one day to process 100 messages, then running through the development corpus (130 0 messages) would take two weeks -a serious impediment to system development . This has placed a greate r premium on system throughput and on automated procedures for detecting errors and evaluating overal l performance . It has also led some systems to explore methods for skimming and for distinguishing \"important\" or high-information passages from less important passages, as well as robust or partial parsin g techniques .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Corpus Dimensions", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The corpus dimensions differed for the test sets as well . The test set for MUC-3 consisted of 10 0 messages (33,000 words), compared to five messages for MUCK-II (158 words) . The larger trainin g corpus placed heavier processing demands on the systems, but it also meant that the test data was mor e representative of the training data . In MUCK-II, 16% of the total number of words in the test represente d previously unseen tokens . In MUC-3, this figure dropped to 1 .6% .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Corpus Dimensions", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "There was a slight change in task focus between MUCK-II and MUC-3 . In MUCK-II, the tas k was template-fill, where each message generated at least one template (although one type of templat e was an \"OTHER\" template, indicating that it did not describe any significant event) . All messages were considered to \"relevant\" . in the sense of generating a template ; only 5% of the training messages generate d an \" OTHER\" template . MUC-3 required both relevance assessment and template fill : only about 50% of the messages were relevant . Part of the task involved determining whether a message containe d relevant information, according to a complicated set of rules that distinguished current terrorist event s from military attacks and from reports on no-longer-current terrorist events . Thus relevance assessmen t turned out to be a complex task, requiring four pages of instructions and definitions in MUC-3 (compare d to half a page of instructions for MUCK-II) . Filling a template for an irrelevant message was penalize d -to a greater or lesser extent, depending on which metrics were used in the summary scoring .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Nature of the Tas k", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Although this represents a change between the two tasks, it is difficult to come up with any numerica l measures to quantify this difference . On the one hand, the participants reported that this shift did no t contribute substantially to the difficulty . On the other hand, most sites devoted substantial effort to creating a specialized set of rules to distinguish relevant from irrelevant messages . Understanding thes e rules for relevance was certainly one of the least portable and most domain-specific aspects of the task , so it undoubtedly did contribute to the greater difficulty of MUC-3 .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Nature of the Tas k", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Reflecting the change in application domains, the templates changed from MUCK-II to MUC-3 . Th e templates differ in how many types of template there are, in the number of slots, in the allowable range of slot fills, and in number of fills per slot (since more than one fill is required in certain cases) . MUCK-I I had 6 types of events and 10 slots per template, of which five were filled from small closed class lists , three from larger closed class lists, and two by string fills from the text . MUC-3 had 10 types of event s and 17 slots (not counting slots reserved for indexing messages and templates), of which eight were smal l closed classes, two were larger closed classes, and seven required numerical or string fills from the text .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Difficulty of the Template Fill Tas k", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "For the MUCK-II test, there were 5 templates generated for 5 messages with just over one fill pe r slot (55 fills for 50 slots) . For the MUC-3 test of 100 messages, 65 out of 100 were relevant . For th e relevant messages, there were 133 templates generated (roughly 2 templates per message, counting the 1 9 optional templates) . The ratio of slots to slot fillers was approximately 2500 answers for 2260 slots (1 . 1 answers/slot) . However, many answers in MUC-3 included cross-references to other slot fillers, whic h were required to get full credit for a correct answer . There were approximately 1000 of these crossreferences, so a more realistic estimate of number of \"fills\" was 3500 (1 .5 answers/slot) . This information * Answers for MUC-3 include cross-reference answer s ** Difficulty = \"perplexity\" counting number of possible fills for a slot as the branching factor at that poin t is summarized in Table 3 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 916, | |
| "end": 923, | |
| "text": "Table 3", | |
| "ref_id": "TABREF4" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Difficulty of the Template Fill Tas k", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "It is possible to enumerate the ways in which the two template fill tasks differ, but it is extremel y difficult to assess how this affects the overall difficulty of filling the template . One crude approach is t o compute a perplexity-like measure of the tasks, looking at the filled template as a string of answers, usin g the number of possible fills for each slot is an estimate of the branching factor at that point . This yields a \"difficulty\" figure of 17 for MUCK-II as opposed to a figure of 30 for MUC-3 . This number corresponde d to the perceived increase in difficulty between the two tasks by the participants : MUC-3 was definitel y viewed as harder, but not an order of magnitude harder .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Difficulty of the Template Fill Tas k", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Finally, the two tasks also scored the results differently . MUCK-II generally used a score based aroun d 1 : wrong = 0, no answer = 1, right = 2 . In MUC-3, the correct answer counted 1, the wrong answe r counted 0 . It is possible to recompute the scores for MUC-2 to make them comparable to MUC-3 . If w e do this, we find that the top-scoring systems in MUCK-II had precision and recall scores in the 70-80 % range . This compares to 45-65% for the top-scoring systems in MUC-3 for the run which maximized bot h precision and recall using the \"MATCHED-MISSING\" method of computing the score r . Since 100% is the upper bound, it is actually more meaningful to compare the \"shortfall\" from the upper bound ; fo r MUCK-II, this is 20-30% and for MUC-3, 35-55% . Thus MUC-3 performance is about half as good a s (has twice the shortfall) as MUCK-II . Table 4 attempts to summarize this discussion by providing a rough order of magnitude for th e different dimensions . We see from this that MUC-3 is many times harder than MUCK-II, in three o f I The term \"MATCHED-MISSING\" refers to the metric which penalized systems for each missing template the four dimensions, while performance has only been cut by a factor of two . Even though the relation between difficulty and precision/recall figures is certainly not linear (the last 10-20% is always much harder to get than the first 80%), the degree of difficulty has increased much more than the performanc e has deteriorated .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 852, | |
| "end": 859, | |
| "text": "Table 4", | |
| "ref_id": "TABREF6" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Scoring and Result s", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "This comparison is reassuring in several respects . First, it means that the field has made very substantial progress in the past two years . MUC-3 shows that current message understanding systems are able to handle a realistic corpus, with a realistic throughput with a reasonable degree of accuracy -highe r precision and recall than many information retrieval systems are likely to get . Secondly, it means that as a test, MUC-3 was well-designed . Part of the motivation in changing tasks and domains after MUCK-I I was to make the problem realistic and sufficiently challenging so that there would be no easy or trick solutions . MUC-3 has served that purpose admirably . It is realistic but current systems can achieve a reasonable level of performance . It is hard enough so that there is substantial room for improvement . This task can provide a reasonable challenge for message understanding systems over the next severa l years .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CONCLUSION S", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Finally, this comparison leads to an important conclusion about evaluation methodology . This pape r represents a tentative first step towards defining some ways of measuring the dimensions of an application . But it is clear that we need to do much more work in this area in order to gain insight into what dimension s really affect success and which ones are less critical . We need to run experiments, where we can vary on e set of parameters, while holding others constant . In short, to gain the maximum benefit from evaluatio n efforts such as the MUC conferences, we need to make evaluation methodology itself a legitimate topi c for research .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "CONCLUSION S", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "I would like to thank a number of people for their help in assembling the data and providing insight s into this problem . In particular, I would like to acknowledge the help that I received from Beth Sundhei m at NOSC, who furnished me with much of the data and some very helpful suggestions for possible ways t o compare the MUCK-II and MUC-3 tasks . I would also like to thank Ralph Grishman (NYU), Jerry Hobbs (SRI), Wendy Lehnert (University of Massachusetts) and Lisa Rau (GE) for feedback and suggestions . Any errors or controversial conclusions are, however, my sole responsibility .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "ACKNOWLEDGEMENT S", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": {}, | |
| "ref_entries": { | |
| "FIGREF1": { | |
| "type_str": "figure", | |
| "num": null, | |
| "uris": null, | |
| "text": "Sample Hard Sentences from MUCK-II and MUC-3" | |
| }, | |
| "TABREF2": { | |
| "content": "<table/>", | |
| "text": "", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null | |
| }, | |
| "TABREF4": { | |
| "content": "<table/>", | |
| "text": "Comparison of the Template Fill Tas k", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null | |
| }, | |
| "TABREF5": { | |
| "content": "<table><tr><td>e</td><td/></tr><tr><td>DIMENSION</td><td>FACTOR 1</td></tr><tr><td>1 . Complexity of Data</td><td>10x</td></tr><tr><td>2 . Corpus Dimensions</td><td>100x</td></tr><tr><td>3 . Nature of Task</td><td>-</td></tr><tr><td>4 . Difficulty of Template Fill</td><td>2x</td></tr><tr><td>5 . Overall Performance</td><td>0 .5x</td></tr></table>", | |
| "text": ", but counte d spurious templates wrong only in the template ID slot, not in each individual filled slo t", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null | |
| }, | |
| "TABREF6": { | |
| "content": "<table/>", | |
| "text": "Summary of Differences : MUCK-II vs . MUC-3", | |
| "num": null, | |
| "type_str": "table", | |
| "html": null | |
| } | |
| } | |
| } | |
| } |