The notion of tense is a subject of lively debate in recent theoretical Chinese linguistics literature. Much of the con­troversy surrounds the question of whether tense is a use­ful theoretical construct. Huang (1984) and Li (1990) sug­gest there is a finite-nonfinite distinction in Chinese syn­tax which implies the existence of a phonologically empty tense category in the syntactic representation of a clause. Other researchers (Hu et al., 2001; Lin, 2003; Lin, 2006; Smith and Erbaugh, 2005) contend that such an abstract notion of tense is not justified for Chinese. In spite of this controversy, there is a general agreement that Chinese does not have grammatical tense. That is, Chinese does not have tense morphemes that are dedicated to mark the tempo­ral location of situations, like English. However, Chinese speakers, like speakers of other world languages, use the moment of speech as the temporal deixis to temporally lo­cate situations in their communication. One piece of ev­idence for this is that temporal adverbials like 4A ("to­day"), sji A ("tomorrow") and t A ("yesterday") all as­sume a temporal deixis that is the moment ofspeech.

Corpus annotation provides a new perspective from which we can look into how Chinese speakers interpret situations temporally. Through linguistic annotation, we can deter­mine whether Chinese speakers consistently identify the temporal locations of situations. To the extent that they do, this information would be highly useful for natural lan­guage processing systems that can learn from the human interpretations through statistical and machine learning ap­proaches. For example, this information would be highly valuable to Machine Translation. To translate a language like Chinese into a language like English in which tense is grammatically marked with inflectional morphemes, an MT system will have to infer the necessary temporal informa­tion to determine the correct tense for verbs. Statistical MT systems, the currently dominant research paradigm, typi­cally do not address this issue directly. As a result, when evaluated for tense, current MT systems often perform mis­erably. For example, when a simple sentence like "Whe BJf A/tomorrow HL ® /return _L ^/Shanghai" is given to

Google's state-of-the-art Machine Translation system, it produces the output "He returned to Shanghai tomorrow", instead of the correct "he will return to Shanghai tomor­row". The past tense on the verb "returned" contradicts the temporal expression "tomorrow". Ye et al (2006), for ex­ample, reported that the best MT systems only get the tense correct 47% of the time, worse than the 69.5% when the most frequent tense (which is the past tense) is assigned. In this paper we report a preliminary study in which we ask Chinese native speakers to temporally locate situations denoted by verbs in naturally occurring text and assign a label that indicates their temporal relations to the document creation time.

The three temporal points, Speech Time, Reference Time and Situation Time are generally accepted as important to temporal interpretation since Reichenbach (1947) first pro­posed them. Speech Time and Reference Time are impor­tant to the interpretation of tense. Speech Time is the mo­ment of speech, and Reference time is the temporal per­spective from which the speaker invites his audience to con­sider a situation, and Situation Time is the time at which the situation actually occurs. In the "tense" annotation study we report here, we annotate the relation between Speech Time and Situation Time instead of the relation between Speech Time and Reference time, which in our judgment is too subtle to be annotated consistently. In determining the different values of tense, one important question to ask is whether one should use the tense taxonomy of a partic­ular language like English, or use a tag set that abstracts away from the language-specific properties. In a previous pilot tense annotation study motivated by Chinese-English machine translation, Ye (2007) used a tagset that is a di­rect reflection of the English tense system. She used a set of 11 tags that represent different combinations of the three tenses (present, past and future) and aspect (perfect, pro­gressive). However, there are reasons for not borrowing the English tense taxonomy directly. First, the English system does not consistently encode the relation between Speech time and Situation time. For example, in "He will call me after he gets here", while his "getting here" happens at a time in the future, it is assigned the present tense because it is in a clause introduced by "after". The annotation of the Chinese tense should not be bound by such English id­iosyncrasies. Instead it should annotate temporal notions that are more universal. Similarly, in English one can ei­ther use an infinitive marker or the future tense to indicate that the temporal location of a situation is in the future. For example, in "He hopes to leave", one can infer that leav­ing is something in the future, much like the leaving in "he hopes that she will leave", where the future tense is explic­itly marked. In a language like Chinese that lacks overt tense markers, if the task is to annotate tense as it is used in English, the choice will be indeterminate in such situa­tions. It makes more sense to annotate the abstract relation between Speech Time and Situation Time which is more deterministic. Adopting the English tense taxonomy also limits the pool of potential annotators, who would have to be proficient in both Chinese and English. Such annotators are generally hard to find.

In written text, which is the primary source of data that we are dealing with, the temporal deixis is the document cre­ation time. All situations are temporally related to this doc­ument creation time except in direct quotations, where the temporal location is relative to the moment of speech of the speaker who is quoted. For example, in (1), there are sev­eral verbs involved. The temporal location of ft ("say") assumes that the temporal deixis is the document creation time, while .JL ("be"), M k ("want") are inside the direct quotations of ^ JJ, the speaker, and their temporal location is relative to the time when he performed that speech act.

Suokang once to he DE mother say : " he be I DE property , I want BA he roll into ball , put in pocket "Suokang once said to his mother: 'He is my property. If I want to roll him into a ball and put it in my pocket,

Tenses that relate the situation time to the document cre­ation time or the moment of speech are considered to be ab­solute tenses, even though the moment of speech is variable and constantly changes. The term situation covers events and states in the broad sense of those terms and they are generally signaled by the presence of a verb. In other cases, the temporal location of a situation cannot be defined in re­lation to the moment of speech, at least not directly. For example in (2), The temporal location of ^ ("intend") cannot be determined independently of the temporal loca­tion of if Jt("reveal"). The temporal location of ^ k is simultaneous with if 3J. If the temporal location of if JJ is in the past, then the temporal location of ^ k is also in the past. If the temporal location of if JJ is in the future, then the temporal location of ^kis also in the future. In this specific case, the situation denoted by the matrix verb if J is in the past. Therefore the situation denoted by ^ k is also located in the past.

he also reveal Russia intend in next ten years within.

to Iran provide weapons .

"He also revealed that Russia intended to provide weapons to Iran within the next ten years."

In our Chinese "tense" annotation task, we annotate both absolute and relative tenses. By "tense", we mean the tem­poral relation between the temporal deixis and the situation (in the case of absolute tense) and between situations (in the case of relative tense). The situation time can be anterior to (in the past), simultaneous with (in the present), or poste­rior to (in the future) the moment of speech. Since there is no grammatical tense that can be directly observed in Chi­nese text, the tense annotation as defined here is a temporal inference exercise. The "situations" that we are interested in are expressed as clauses centered around a verb, and for the sake of convenience we mark the "tense" on the verb itself instead of the entire clause. However, when infer­ring the temporal location of the situation, we have to take into consideration the entire clause, because the arguments and modifiers of a verb are just as important as the verb itself when determining the temporal location of the situa­tion. For example, in the absence of additional information such as temporal adverbials, the default temporal location of (3 a) is the present while the default temporal location of (3b) is the past, even though the verb is the same in these two clauses. (3a) describes a habitual situation where the person described sleeps a lot repeatedly, even though he is not in a perpetual state of sleep. The habitual interpre­tation is induced by the adverbial modifier 4_- A ("all the time"). In contrast, the duration of three hours quantifies the sleep situation in (3b) and establishes temporal bound for B§ ("sleep"). By default bounded situations are tempo­rally located in the past (Smith and Erbaugh, 2005), a point we will elaborate further.

he all the time sleep sleep .

"He sleeps all the time".

b. -fell 7 =- 'hat. he sleep ASP three hour .

"He slept for three hours."

We performed two experiments with two annotators that are native speakers of Chinese who are also fluent in English. In the first experiment, we replicated the annotation condi­tion reported in (Ye, 2007) under which the annotators were asked to choose a tense they would use to translate a Chi­nese verb into English without providing them with a set of guidelines. The experiment was done on 1076 verb in­stances from 6 articles randomly selected from the Chinese Treebank. The annotators were presented with each verb in its context in the article. The results are presented in Table 2. The agreement is higher than the 43% is reported in (Ye, 2007), but on a reduced tag set that is only relevant to tense. In the second experiment, we provided the annotators with a simple set of guidelines that are summarized in Section 3.. The guidelines summarize factors that the annotators should consider when making their judgment on the tense selection. The guidelines are presented as a set ofrules that can be overwritten when "hard evidence" such as temporal adverbials are present. The second experiment was done on a different set of verbs taken from six different articles from the first experiment. The results (Table 3) show that even with this set ofvery simple guidelines, there has been sub­stantial improvement (17%) in inter-annotator agreement, from the 58% in Experiment 1 to the 75% in Experiment 2. One interesting observation is that the past tense is anno­tated with very high inter-annotator agreement (86%) while the future and present tenses are much harder to annotate. This is something that we will try to explain in future work.

CarlotaS. Smith and Mary Erbaugh. 2005. Temporal inter­pretation in Mandarin Chinese. Linguistics, 43(4):713-756.

Z. Xiao and A. McEnery. 2004. Aspect in Mandarin Chi­nese: A corpus-based study. Amsterdam, John Ben­jamins.

Yang Ye, Victoria Li Fossum, and Steven Abney. 2006. Latent features in automatic tense translation between Chinese and English. In The Proceedings of the 5th SIGHAN Workshop on Chinese Language Processing, Sydney, Australia. Yang Ye. 2007. Automatica Tense and Aspect Translation between Chinese and English. Ph.D. thesis, University of Michigan.

In the context of natural language processing, annotation is about recovering implicit information that is useful for natural language applications. In this sense annotation of Chinese tense would be a highly rewarding endeavor if suc­cessful. Our preliminary experiments show that providing a set of intuitive of guidelines is crucial to Chinese tense an­notation. Although the inter-annotator agreement we report here has not yet reached satisfactory levels, we are confi­dent that with refined guidelines the inter-annotator agree­ment can be further improved.