DiaBLa: A Corpus of Bilingual Spontaneous Written Dialogues for Machine Translation

We present a new English-French test set for the evaluation of Machine Translation (MT) for informal, written bilingual dialogue. The test set contains 144 spontaneous dialogues (5,700+ sentences) between native English and French speakers, mediated by one of two neural MT systems in a range of role-play settings. The dialogues are accompanied by fine-grained sentence-level judgments of MT quality, produced by the dialogue participants themselves, as well as by manually normalised versions and reference translations produced a posteriori. The motivation for the corpus is two-fold: to provide (i) a unique resource for evaluating MT models, and (ii) a corpus for the analysis of MT-mediated communication. We provide a preliminary analysis of the corpus to confirm that the participants' judgments reveal perceptible differences in MT quality between the two MT systems used.


Introduction
The use of Machine Translation (MT) to translate everyday, written exchanges is becoming increasingly commonplace; translation tools now regularly appear on chat applications and social networking sites to enable cross-lingual communciation. MT systems must therefore be able to deal with a wide variety of topics, styles and vocabulary. Importantly, the translation of dialogue requires translating sentences coherently with respect to the conversational flow in order for all aspects of the exchange, including speaker intent, attitude and style, to be correctly communicated.
It is important to have realistic data to evaluate MT models and to guide future MT research for informal, written exchanges. In this article, we present DiaBLa (Dialogue BiLingue 'Bilingual Dialogue'), a new test set of English-French spontaneous written dialogues mediated by MT, 1 obtained by crowdsourcing and covering a range of dialogue topics and annotated with fine-grained human judgments of MT quality. To our knowledge, this is the first corpus of its kind. Our data collection protocol is designed to encourage spontaneous dialogue between speakers of two languages, using role-play scenarios to provide conversation material. Sentence-level human judgments of the quality of the MT systems are provided by the participants themselves while they are actively engaged in dialogue. The result is a rich bilingual test corpus of 144 dialogues, which are annotated with sentence-level MT quality evaluations and human reference translations.
We discuss the potential of the corpus and the collection method for MT research in Sec. 2, both for MT evaluation and for the study of language behaviour in informal dialogues. In Sec. 3 we describe the data collection protocol and interface. We describe basic characteristics and examples of the corpus in Sec. 4. This includes a description of the annotation layers (normalised versions, human reference translations and human MT quality judgments). We illustrate the usefulness of the human evaluation by providing a comparison and analysis of the MT systems used (Sec. 4.4). We compare two different types of MT system, a baseline model and a mildly context-aware model. Finally, we provide plans for future work on the corpus in Sec. 5. The corpus, interface, scripts and participation guidelines are freely available under a CC BY-SA 3.0 licence. 2 1 Participants write and receive messages in their respective native language thanks to MT systems translating between the two languages.
2 https://github.com/rbawden/ DiaBLa-dataset arXiv:1905.13354v1 [cs.CL] 30 May 2019 1.1 Related work A number of corpora of informal data do exist. However they either cover different domains or are not designed with the same aim in mind. OpenSubtitles2016 (Lison and Tiedemann, 2016) is a large-scale corpus of film subtitles, from a variety of domains, making for very heterogeneous content. However the conversations are scripted rather than being spontaneous, and are translations of monolingual texts, rather than being bilingual conversations. The MSLT corpus (Federmann and Lewis, 2016) is designed as a bilingual corpus, and is based on oral dialogues produced by bilingual speakers, who understand the other speaker's original utterances. This means that it is not possible to analyse the impact that using MT has on the interaction between participants. Other bilingual taskorientated corpora exist, for example BTEC (Basic Travel Expression Corpus) (Takezawa et al., 2002), SLDB (Spoken Language DataBase) (Morimoto et al., 1994) and Field Experiment Data of Takezawa et al. (2007), which is the most similar corpus to our own in that it contains MT-mediated dialogue. However these corpora are restricted to the travel/hotellery domains, therefore not allowing the same variety of conversation topic as our corpus. Human judgments for the overall quality are provided for the third corpus, but only at a very coarse-grained level. Feedback about the participants' perception of MT quality is therefore of a limited nature in terms of MT evaluation; sentence-level evaluations are not provided.

Motivation
The main aim of our corpus is as a test set to evaluate MT models in an informal setting in which communication is mediated by MT systems. However, the corpus can also be of interest for studying the type of language used in the dialogues as well as the way in which human interaction is affected by use of MT as a mediation tool. We develop these two motivations here, starting with the corpus' utility for MT evaluation (Sec. 2.1) and then discussing the corpus' potential for analysis of MT-assisted interaction (Sec. 2.2).

MT evaluation
The corpus is useful for MT evaluation in three ways: as (i) a test set for automatically evaluating new models, (ii) a challenge set for manual evaluation, and (iii) a validation of the effectiveness of the protocol to collect new dialogues and to compare new translation models in the future.
Test set for automatic evaluation The test set provides an example of spontaneously produced utterances in a real, unscripted setting. It could be particularly useful for evaluating contextual MT models due to the dialogic nature of the utterances, the need to take into account previous MT outputs and the presence of metadata.
Challenge set for manual evaluation It can be used as a basis for manual evaluation (as a challenge set). The sentence-level human judgments provided can be used as an indicator as to which sentences were the most challenging for MT. Manual evaluation of new translations of our test set can then be guided towards those sentences whose translations were marked as poor, to provide an informed idea of the quality of the new models of these difficult examples, and to encourage development for particularly challenging phenomena.
Validation of the protocol for the collection of human judgments of MT quality Human evaluation remains the most accurate form of MT evaluation, especially for understanding which aspects of language pose difficulties for translation. While hand-crafted examples and challenge sets provide the means to test particular phenomena (King and Falkedal, 1990;Isabelle et al., 2017), it is also important to observe and evaluate the quality of translation on real, spontaneously produced texts. Our corpus provides this opportunity, as it contains spontaneous productions by human participants and is richly annotated for MT quality by its end users (see Sec. 4.4). We provide a preliminary comparative evaluation of the two MT systems, in order to show the utility of the human judgments collected. The same collection method can be applied to new MT models for a similar evaluation.

MT-assisted interaction
As MT systems are becoming more common online, it is important for them to take into account the type of language that may be used and the way in which user behaviour may affect the system's translation quality. Non-canonical syntactic structures, spelling and typing errors, text mimicking speech, including pauses and reformulations, must be taken into account if MT systems are to be used for successful communication in more informal environments. The language used in our corpus is relatively clean in terms of spelling. However participants are encouraged to be natural with their language, and therefore a fruitful direction would be in the analysis of the type of language used.
Another interesting aspect of human-MT interaction would be to study how users themselves adapt to using such a tool during the dialogues. How do they deal with translation errors, particularly those that make the dialogue incoherent? Do they adjust their language over time, and how do they indicate when they have not understood correctly? An interesting line of research would be to use the corpus to study users' communication strategies, for example by studying breakdowns in communication as in (Higashinaka et al., 2016).

Data collection and protocol
We collected the dialogues via a dedicated web interface, allowing participants to register, log on and chat. Each dialogue involves two speakers, a native French speaker and a native English speaker. Each writes in their native language and the dialogue is mediated by two MT systems, one translating French utterances into English and the other translating English utterances into French.
Participants Participants are volunteers recruited by word of mouth and social media. They participated free of charge, motivated by the fun of taking part in fictional role-play. Users provide basic information: age bracket, gender, English and French language ability, other languages spoken and whether they work in research or Natural Language Processing (see Tab. 1 for basic statistics).
Scenarios To provide inspiration, a role-play scenario is given at the start of each dialogue and roles are randomly assigned to the speakers. We designed twelve different scenarios (cf. App. A) two of which are shown in Fig. 1. The first turn is also assigned randomly to one of the speakers to get the dialogue started. This information is indicated at the top of the dialogue screen in the participants' native languages. A minimum number of 15 sentences per speaker is recommended, and participants are informed once this threshold is reached. Participants are told to play fictional characters and not to use any personal details. We nevertheless anonymise the corpus prior to distribution to remove usernames mentioned in the text.  Table 1: Some characteristics of the participants.

Evaluation method
You are both stuck in a lift at work (1) You are an employee and are with your boss.
(2) You are the boss and are with an employee.
You are in a retirement home.
(1) You are visiting and talking to an old friend.
(2) You are a resident and you are talking with an old friend who is visiting you. gual point of view. The choice to use the participants to provide the MT evaluation is an important part of our protocol: we can collect judgments on the fly, facilitating the evaluation process, and it importantly means that the evaluation is performed from the point of view of participants actively engaged in dialogue. Athough some errors may go unnoticed (e.g. a word choice error that nevertheless makes sense in context), many errors can be detected this way through judgments about coherence and understanding of the dialogue flow. Having information about perceived mistakes could also be important for identifying those mistakes that go unperceived in translation. MT quality is evaluated twice, (i) during the dialogue and (ii) at the end of the dialogue. Evaluations are saved for later and not shown to the other participant. They evaluate each translated sentence during the dialogue by selecting an overall translation quality (perfect, medium or poor), and indicating which errors occur: grammar, meaning, style, word choice, coherence and other (see Fig. 2 for an example). Note that several problems can be indicated for the same sentence. If they wish, they can also write a free comment providing additional information or suggesting corrections. Once the dialogue is finished, they give overall feedback of the MT quality. They are asked to rank the quality of the translations in terms of grammaticality, meaning, style, word choice and coherence on a five-point scale (excellent, good, average, poor and very poor), and to indicate whether any particular aspects of the translation or of the interface were problematic. Finally, they indicate whether they would use such a system to communicate with a speaker of another language. Figure 2: The sentence-level evaluation form. The original French sentence was L'entrée c'est juste deux soupes de melon. "The starter is just two melon soups." Participants were given instructions (with examples) on how to evaluate MT quality. However there is expected to be a certain degree of variation in the way participants evaluate. This subjectivity, inevitable with any human evaluation, is interesting, as it gives an indication of the variance of the tolerance for errors, and which types of errors were considered most detrimental.

MT systems
We compare the quality of two MT model types (see Sec. 4.4). Within a dialogue, the same model type is used for both language directions, and the same number of dialogues is mediated by each model type. Both models are neural encoder-decoder models with attention (Bahdanau et al., 2015), implemented using Nematus (Sennrich et al., 2017). The first model (BASELINE) is trained to translate sentences in isolation. The second (2TO2), is trained to translate sentences in the context of the previous sentence, as in (Tiedemann and Scherrer, 2017) and (Bawden et al., 2018). This is done by concatenating each sentence with its previous sentence, separated by a special token, and translating both sentences at once. In a post-processing step, only the current sentence is kept. Note that if the previous sentence is spoken by the same speaker as the current sentence, then the original previous sentence is prepended. If the previous sentence is spoken by the other speaker (in the opposite language), then the MT output of the previous sentence is prepended to the current sentence. This means that the previous sentence is always in the same language as the current sentence, and also corresponds to the context seen by the current speaker.
Training Data and MT setup The systems are trained using the OpenSubtitles2016 parallel corpus (Lison and Tiedemann, 2016). The data is cleaned, tokenised and truecased using the Moses toolkit (Koehn et al., 2007) and tokens are split into subword units using BPE (Sennrich et al., 2016b). The data is then filtered to exclude poorly aligned or truncated sentences, resulting in a training set of 24,140,225 sentences. Hyperparameters are given in App. B. During the dialogues, the participants' text is first split into sentences and preprocessed in the same way as the training data. Translation is performed using MARIAN for fast CPU decoding (Junczys-Dowmunt et al., 2018). All processing scripts will be released should the paper be accepted.  Tab. 2 shows the basic characteristics of the 144 dialogues. 75.7% of dialogues contain ≥35 sentences and the average sentence length is 9.9 tokens, very slightly longer than the translations. An extract of dialogue, representing a fictional argument, is given in Fig. 3, providing an example of the type of language used by the participants. The language used is colloquial and contains a number of fixed expressions (e.g. get off your intellectual high-horse, Mr Fancy pants), which can prove difficult for MT, as is the case in this example. The systems are sometimes robust enough to handle spelling and grammatical errors (e.g. qui ne penses 'who think 2.sg ' instead of qui ne pense 'who thinks 3.sg " and rality instead of reality, translated into French as ralité instead of réalité, conserving the spelling error in translation). The dialogues also contain cultural references, such as references to films and actors. In many cases named entities are well conserved, although sometimes cause problems (e.g. Marcel Carné translated as Marcel Carborn). 3

Normalised versions
Although participants were encouraged to use their best spelling and grammar, such errors did occur (missing or repeated words, typographical errors, inconsistent use of punctuation). We provide manually normalised versions of sentences containing errors. The aim of this normalisation is to provide information about the presence of errors (useful for studying their impact on translation), and for providing a basis for the human reference translations, as we do not attempt to reproduce errors in the translations. Corrections are kept to a minimum (i.e non-canonical use of language was not corrected if linked to the colloquial use of language), and therefore in practice are limited to the addition of capital letters at the beginning of sentences and fullstops at the end of sentences and typographical error correction only when the correct form can easily be guessed from the context. 4

Machine translations
Each sentence is translated automatically into the other language for the other participant. A single type of MT system (BASELINE or 2TO2) is used for all sentences within a dialogue. The use of two different systems is relevant to our analysis of the human evaluations produced by dialogue participants (Sec. 4.4). The choice of MT system does of course affect the quality of the MT. However, the corpus will remain relevant and useful as a test set and for analysing human language behaviour in this setup independently of this choice.
3 Carné is segmented into two subwords, Car and né 'born', and the second element has been translated. 4 When a wrong but attested word is used, but it is made explicit that this is intentional, we do not correct the word. E.g. the use of 'deserts' instead of 'desserts' in 'The deserts are in the refrigerator. I said deserts for fun, I meant desserts!'  Figure 3: A dialogue extract with baseline MT output ("MT:", in grey) and human evaluation ("Eval:").

Human reference translations
In order for the corpus to be used as a test set for future MT models, we also produce human reference translations for each language direction. Translators were native speakers of the target language, with very good to bilingual command of . . . about ice-cream that's eaten? ou bienà une glace pour se regarder ?
Or about a mirror to look into?
Meta-discussion Tu connais un restau indonésien? Do you know an Indonesian restaurant? Ouà la limite thaï ?
Or at a push Thai? (MT: Or the Thai limit) What do you mean by the Thai limit?
Qu'est-ce que tu veux dire par la limite thaïlandaise ? the source language, and all translations were further verified by a bilingual speaker. Particular attention was paid to producing natural, spontaneous translations. The translators did not have access to the machine translated versions of the sentences they were translating to avoid any bias towards the MT models or the training data. However, they could see the machine translated sentences of the opposite language direction. This was important to ensure that utterances were manually translated in the context in which they were originally produced (as the speaker would have seen the dialogue) and to ensure cohesive translations (e.g. for discursive phenomena, such as anaphoric phenomena and lexical repetition). Spelling mistakes and other typographical irregularities (e.g. missing punctuation and capital letters) were not transferred to the translations; 5 the translations are therefore clean (as if no typographical errors had been present).
Translation difficulties The particularity of the setup makes a small number sentences difficult to translate (cf. also the examples in Tab. 3): • the informal nature means that many utterances are idiomatic and translation equivalents hard to find. We chose idiomatic equivalents based on communicative intention.
• ambiguity in one language (but not the other) can sometimes lead to seemingly nonsensical utterances. This is a more theoretical translation problem and not one that can be solved satisfactorily. We translated these utterances as best as possible, despite resulting incoherences in the target language.
• occasional coherence problems due to mistranslations (a side-effect of using MT), which led to meta-discussions. Where possible, we made translations coherent with the dialogue context available to the speaker.

Human judgments of MT quality
As described in Sec. 3, participants evaluated the translations from a monolingual (target language) perspective. We provide a preliminary analysis of these judgments to show that they are a good indicator of MT quality and that such a protocol is useful for comparing MT models in a real setting.
Overall MT quality Although an in-depth linguistic analysis is beyond the scope of this paper, we look here at global trends in evaluation. 6 Differences between models are shown in Fig. 4. They show unsurprisingly that MT quality is dependent on the language pair; translation into English is perceived as better than into French, approximately half of all EN→FR sentences being annotated as medium or poor. 7 There is little difference in perceived quality between the BASELINE and 2TO2 for FR→EN. This contrasts with EN→FR, for which the number of sentences marked as perfect is higher by +4% for 2TO2 than for BASELINE. An automatic evaluation with BLEU 8 (Papineni et al., 2002) shows that the contextual model scores mildly better than the baseline, particularly for EN→FR. We retranslate all

Types of errors encountered
The evaluation results for each problem type are shown in Fig. 5. The few number of problems classed as other indicates that our categorisation of MT errors was sufficiently well chosen. The most salient errors for all language directions and models are in word choice, especially when translating into French, with approximately 16% of sentences deemed to contain a word choice error. As with the overall evaluations, there are few differences between BASELINE and 2TO2 for FR→EN, but significant differences are seen for EN→FR: 2TO2 models perform better, with fewer errors in most problem types, except word choice. A notable difference is the lower frequency of coherence-related errors for 2TO2. Coherence errors also appear to be less serious, as there is a lower percentage of translations labelled as poor (as opposed to medium).
These results are encouraging, as they show that our data collection method is a viable way to collect human judgments, and that such judgments can reveal fine-grained differences in MT systems, even when evaluating on different sentence sets.
In spite of the errors, the translation quality is in general good, especially into English, and participant feedback is excellent concerning intelligiblity and dialogue flow. As well as sentence-level judgments, participants indicated overall MT quality once the dialogue was complete. Participants indicated that they would use such a system to communicate with a speaker of another language 89% of the time. In 81% of dialogues, grammaticality was marked as either good or excellent. Coherence, style and meaning were all indicated as being good or excellent between 76% and 79% of the time. As a confirmation of the sentence-level evaluations, word choice was the most problematic error type, indicated in only 56% of dialogues as being good or excellent (40% of dialogues had average word choice, leaving a very small percentage in which it was perceived as poor).

Focus on a discourse-level phenomenon
We study one specific discursive phenomenon: the consistent use of French pronouns tu and vous. The French translation of singular you is ambiguous between tu (informal) and vous (formal). Their inconsistent use was one of the most commented problems by French speakers, 9 and a strategy for controlling this choice has been suggested for this reason (Sennrich et al., 2016a). Neither of our models explicitly handles this choice, although 2TO2 does take into account pairs of consecutive  sentences, and therefore could be expected to have more consistent use of the pronouns across neighbouring sentences. As a proxy for its ability to account for lexical cohesion, we look at the two models' ability to ensure consistent translation of the pronouns across consecutive sentences. For each model, we take translated sentences in which tu or vous appear, and for which the previous sentence also contains either tu or vous. By comparing the number of times the current sentence contains the same pronoun as the previous sentence (see Tab. 4), we can estimate the degree of translation consistency for this particular aspect. Although the absolute figures are too low to provide statistical significance, we can see a general trend that the 2TO2 model does show greater consistency in the use of the pronouns over the baseline model, with +9% in the consistency use of tu and +6% in the consistent use of vous.

Discussion and future work
Our new corpus of MT-mediated dialogues provides a basis for many future research studies, in terms of the interaction between MT and humans: how good communication can be when using MT systems, how MT systems must adapt to real-life human behaviour and how humans handle communication errors. We have already shown through a preliminary analysis that the human judgments provide a viable form of MT evaluation and can be further analysed to give us more insight into MT of dialogue. The same protocol could be extended to other language pairs, although this would be part of a larger, international effort. We intend to further extend the English-French corpus in future work and annotate it with discourse-level information, which will pave the way for future phenomenon-specific evaluation: how they are handled by different MT systems and evaluated by the participants. In this direction, we manually annotated anaphoric phenomena in 27 dialogues (anaphoric pronouns, event coreference, possessives, etc.). Despite the small size of this sample, it already displays interesting characteristics, which could provide a strong basis for future work. Anaphoric references are common in the sample annotated: 250 anaphoric pronouns, 34 possessive pronouns, and 117 instances of event coreference. Their incorrect translation was often a cause of communication problems (see Fig. 6), the impact of which will be investigated further. The resource could also be extended to more scenarios. It is designed to cover a wide range of topics, and to be relatively unrestricted in the scenarios imposed. This is a point of difference with existing bilingual corpora, which are often taskorientated and/or limited to a particular domain. For example BTEC (Basic Travel Expression Corpus) and SLDB (Spoken Language DataBase) are both focused on the travel domain and represent a particular style of conversation structured around a task. The relatively unrestricted nature of our corpus allows participants to be freer in their language use, and will allow us to analyse conversational language over multiple domains.

Conclusion
The corpus presented in this paper therefore provides a wide range of opportunities for future research into the type of language used in this setting and the way in which participants handle breakdowns. The protocol for data and MT judgment collection presented provides a useful framework for future evaluation of MT quality, and the corpus itself can be used as a test set to guide manual evaluation of new models. Our preliminary analyses of sentence-level human judgments show that the evaluation procedure is viable, and we have observed some interesting differences between two types of MT model used in our experiments.

C Participation guidelines
The following guidelines were presented to all participants, and were available during the dialogue if needed. A French translation was presented to French-speaking participants.

DiaBLa Instructions
You will be participating in an improvised written dialogue with another user. You will each write in your own native language (English or French). Don't worry -you do not need to be able to speak or understand the other language. Machine translation systems will translate all of the other person's sentences into your language. You will also evaluate the translations from a monolingual point of view (i.e. is the sentence grammatical? Does it make sense? Was the word choice ok? Is it stylistically appropriate? Is it coherent with respect to previous sentences?) Please read all instructions carefully before continuing!

C.1 Signing up and logging in
You must first register (we require some basic information -see FAQ). Log in using the email address you registered with. Choose a username and the language you are going to speak in, which must be your mother tongue. You will be talking to real people. To increase your chances of finding someone to chat to, fill in this spreadsheet with your availabilities. Or try your luck and log in straight away!

C.2 Dialoguing
Once logged in, invite someone by clicking on their username or wait for someone to invite you. You can accept or refuse an invitation to dialogue. If the request is accepted, you will be taken to the dialogue screen. One of you is assigned the first turn, and after that, you are free to dialogue as you please. You will be presented with a setting (at the top of the chat box) in which the dialogue will take place. E.g. "You are in a forest" and your role. Now improvise a dialogue in the setting provided, as in improvised drama or role-play. I.e. play a made-up character and not yourself. The dialogues are to be like written drama transcriptions, rather than chat messages. We recommend writing at least 15 sentences each (you will receive a message when this happens). You can write more, but it is even more useful for us to have more dialogues rather than fewer very long ones.
Please do not use: • emoticons or SMS speech • your partner's username, your own username or personal details Please do use: • your best spelling, grammar and punctuation • the correct gender of you and your partner (for instance when using pronouns) • your imagination! You can refer to imaginary objects/people around you

C.3 Evaluation
You will evaluate the other person's translated utterances by selecting one of the smileys: • green smiley face: "perfect" • orange neutral face: "ok but not perfect" • sad red face: "poor" When you select a smiley, you will be prompted to indicate what is wrong with the translation: grammar, meaning, word choice, style, coherence, plus any extra comments to make your evaluation clearer. See FAQ for some examples.

C.4 Purpose
We will be using the dialogues to evaluate the machine translation systems and how easy communication was. The dialogues will be used to create a corpus of dialogues, which will be freely distributed for research purposes, and also used to analyse the machine translation models. Be natural, spontaneous and creative! However, please avoid making your sentences purposefully difficult in order to trick the machine translation system. Thank you!

C.5 FAQ
What if I don't understand what my partner says? As in real life, speak about the problem with your partner. Say that you don't understand and try to continue the dialogue as best as possible.
When evaluating, what do the error types correspond to?
• Meaning: the sentence does not appear to make sense, e.g.: I was told by my avocado that a sentence was likely.
• Word choice: a poor word choice was made, e.g.: "I did you a chocolate cake" (instead of "I made you a chocolate cake."), "He took an attempt" (instead of "He made an attempt") • Style: the level of formality is inconsistent or language usage is strange, although grammatically well-formed and understandable, e.g.: Strange/unnatural utterances, wrong level of formality: "What's up" in a job interview, etc.
• Coherence: Lack of consistency with previous utterances or the context: wrong pronoun used that refers to something previously mentioned, inconsistent use of "tu" and "vous" (for French), word choice is inconsistent with what was previously said (e.g. "I'm angry! -What do you mean by 'upset'?"), etc.
Why do you need personal information? This enables us to evaluate whether certain aspects of the conversation (e.g. gender marking in French) are correctly translated or not. It allows us to analyse how machine translation systems react to the differences in language use, which depends for instance on the age of the user. The personal information that will be distributed in the resulting corpus is the following: Why do you need to know speaker gender? Speaker gender can be important in certain languages in terms of which words agree with the gender of the speaker (e.g. French). We therefore ask you to choose between male and female for practical reasons. If you do not identify with either gender, please choose the one by which you wish to be identified linguistically (i.e. would you prefer to be referred to as "he" or "she"?). The important thing is to be coherent when you dialogue in your use of gender. Figure 7 shows an example of the interface used for dialogue collection.