François Mairesse, Milica Gašić, Filip Jurčíček, Simon Keizer, Blaise Thomson, Kai Yu, Steve Young
ACL 2010
http://aclweb.org/anthology-new/P/P10/P10-1157.pdf

Presented by Ondřej Dušek
Report by Honza Václ

The following exercises were sent to the mail conference few days before and were aimed to make the readers think about the semantic stack representation used in the paper. They were not answered in very much detail in the lecture, just went through to make sure we understand the basic concepts. Thus, the following solutions are mostly my own interpretation and are not guaranteed to be 100 percent correct.

Ex. 1) Try to think of a semantic stack representation and a dialogue act
for the following sentence:
For Chinese food, The Golden Palace restaurant is the best one in the
centre of the city. It is located on the side of the river near East
Road.
The solution could look something like this (inspired by Table 1):

2) Try to think of a surface realization for the following dialogue act:
reject(type=placetoeat,eattype=restaurant,pricerange=cheap,area=citycentre,food=Indian)

Note: Ondřej admitted that the syntax of the dialogue act in the Ex. 2 was not exactly the same as the one used throughout the paper, but rather taken out directly from the corpus used by the authors (publicly available).

The paper was widely discussed throughout the whole session. The report tries to divide the points discussed in correspondence to the sections of the paper.

The paper proposes a semi-automatic translation evaluation metric that is claimed to be both well correlated with human judgment (especially in comparison to BLEU) and less labour-intensive than HTER (which is claimed to be much more expensive).

Meant assumes that a good translation is one where the reader understands correctly “Who did what to whom, when, where and why” - which, as Martin noted, is rather adequacy than fluency, and therefore a comparison with BLEU, which is more fluency-oriented, is not completely fair. Moreover, good systems usually make more errors in adequacy than in fluency, which makes BLEU an even worse metric these days.

Martin further explained that HTER is a metric where the humans post-edit the MT output to transform it into a correct translation, and then TER, which is actually a word-based Levenshtein distance, is computed as the score.
Matěj Korvas then pointed to an important difference between MEANT and HTER: MEANT uses reference translations, whereas HTER uses post-editations. Surprisingly, this is not noted in the paper.

Section 2 Related work was skipped.

Here we look at how the evaluation is actually done. It consists of three steps, all done by humans in HMEANT. In MEANT, the first step is done automatically.

1. SRL (semantic role labelling) of both the reference and the MT output; the labels are based on PropBank (but have nicer names)
2. aligning the frames - first, predicates are aligned, and then, for each matching pair of predicates, their arguments are aligned as well
3. ternary judging - deciding whether each matched role is translated correctly, incorrectly or only partially correctly

The group discussed whether HMEANT evaluations are really faster than HTER annotations, as some of the readers participated in HMEANT evaluation. Some readers agree that about 5 minutes per sentence is quite accurate, while others state that 5 minutes are at best the lower bound. However, it is not completely clear whether all of the three phases of annotation are claimed to be done in 5 minutes. (Probably yes, but the less do the readers agree with the necessary times indicated.)

The answer is that it contains the arguments of the predicate. It actually contains all possible roles, where the non-present ones only add a zero to the sum and therefore do not influence the score.

We further tried to compute the score for the following set of sentences:

• Reference: John loves Mary.
• MT1: Stupid John loves Mary.
• MT2: John loves Jack.
• MT3: John hates Mary.

We supposed that the semantic roles are the same in all cases, i.e. Agent for John or Stupid John, Predicate for loves or hates, and Experiencer for Mary. It was explained by Martin that Stupid John has no inner structure in HMEANT as there is no predicate in the phrase - HMEANT semantic annotation is shallow in that respect. Furthermore, we assumed (following the paper's Section 3) that the weights are uniform, i.e. w_pred = w_j = 0.1 and w_partial = 0.5.

For MT1, the HMEANT score is equal to 1, because, according to the paper, extra information is not penalized, and the translation is therefore regarded as being completely correct.

For MT2, C_precision is 1, but C_recall is only 2/3, and the HMEANT score, which is the F-score, is therefore 4/5.

For MT3, the predicates do not match, and therefore no arguments are taken into account and the score is 0. Martin and Ruda agreed that most probably not even a partial match of predicates can be annotated, as there is no support for such annotation in the formulas, which Martin suggested to be a possible flaw of the method.

Karel Bílek also noted that it is hard to annotate semantics on incorrect sentences, which is not mentioned in the paper.

Here, we reminded the difference between Kendall's τ and Spearman's τ. Kendall's τ only takes the ranks into account, disregarding the actual scores, while Spearman's τ takes the scores into account. The formula for Kendall's τ is τ = (#same_ordered_pairs - #opposite_ordered_pairs) / #all_pairs.

Martin also remarked that the authors use sentence-level BLEU to compute the correlation; however, BLEU was designed for whole documents, not for individual sentences, and therefore should preferably not be used on sentence level.

Petr notes that, although it might seem surprising that monolinguals perform better in the evaluation than bilinguals, it is probably a consequence of the fact that bilinguals try to guess what the source was, while the monolinguals cannot do that.

All other sections were basically skipped.

For the rest of the session, Martin took the lead to express some more objections to the paper. The group agreed with the objections, and even added some more.

Table 3 seems to represent the main results of the paper.It is shocking that the authors used only 40 sentences; moreover, they used it as both the training set and the test set.
The grid search they use to tune the parameters means to “try everything and find the best-correlating parameters” - in this case this is 12 parameters. They ran the grid search optimization on the 40 sentences they have, but then they evaluated HMEANT on the same data. The group agreed that such evaluation is completely flawed and it is not clear why it was performed and included in the paper.
Karel Bílek also notes that it is quite ridiculous to state the precision to 4 decimal digits when only 40 sentences are used.

In Table 4, the authors probably try to compensate for this flaw by performing cross-validation. However, please note there are only 10 sentences in one fold. Petr thinks that the table should show that the parameter weights are stable. However, Martin thinks that for only 40 sentences, it is probably easy to find 12 parameter values to achieve good performance. Moreover, Aleš Tamchyna assumes that even the formulas used might be fitted to those 40 sentences.

Martin then informed the group that Dekai Wu has still not given us the data from the annotations done on ÚFAL (which was already several months ago), which makes us even more suspicious whether the experiments were fair.

Martin also notes that the authors claim that all other existing evaluation metrics require lexical matches to consider a translation to be correct - which is not true, as the Meteor metric can also use paraphrases.

The group generally agreed that, although the ideas behind HMEANT seem reasonable, the paper itself is misleading and is not to be believed much (or probably at all). The proposed metric possibly correlates better with human judgment than automatic metrics, but it does not really seem to reach HTER.