Fairytale Child (Pohádkové díteˇ) is a simple chatbot trying to simulate a curious child. It asks the user to tell a fairy tale, often interrupting to ask for details and clarifications. However, it remembers what it was told and tries to show it if possible. The chatbot can communicate in Czech and in English. It analyzes the morphology of each sentence produced by the user with natural language processing tools, tries to identify potential questions to ask, and then asks one. A morphological generator is employed to generate correctly inflected sentences in Czech, so that the resulting sentences sound as natural as possible.
One of the first chatbots, and the first one that became
widely known and successful, was ELIZA [
Despite its simplicity, ELIZA proved to be able to successfully communicate and react naturally enough so that numerous users reported a belief that the chatbot was actually intelligent and did understand them. One of the reasons behind the success of ELIZA is the fact that when communicating with a psychotherapist, it is the client who “creates” the world about which he talks; it is thus natural that the psychotherapist does not know anything about that world (which would appear strange in other situations, such as discussing the current political situation). Another important decision was the choice of English as the communication language, which enabled employment of rather simple pattern matching instead of full-fledged linguistic analyses, and extremely simple language generation techniques; this is thanks to English being an analytical (nearly non-inflecting) language with a fixed word order.
In this work, we present a Czech language chatbot for a similar setting – the Fairytale Child. The chatbot tries to simulate a child which wants to hear a fairy tale. Similarly to the therapist scenario, it is the user who creates the entire world in which the story takes place. The child knows nothing about this world, and thus asks a lot of questions to learn something about the world, the people and objects in it, and their properties.
The behaviour of the chatbot is similar to the “Whyphase” of child development, in which the child repeatedly asks “Why?” as a reaction to virtually any sentence. However, as the “Why-phase” is rather annoying for the parents, and we do not want our chatbot to be annoying, we try to simulate an older child, which has a wider range of more specific questions, and which often does not ask any question but simply encourages the user to go on.
Similarly to several other chatbots, the Fairytale Child remembers what it was told, and does not ask for it again; instead, it tries to show the user that it has paid attention by showing off the knowledge it has acquired if the story touches the topic again.
The Fairytale Child chatbot is built using the Treex
Natural Language Processing (NLP) framework [
The main aim of our work is to show that when backed by a powerful and easy-to-use NLP framework, construction of a simple Czech chatbot able to produce naturally formed responses can be not only viable, but even rather straightforward.
The chatbot is being developed in two versions –
Fairytale Child for English, and Pohádkové díteˇ for Czech. As
creating a chatbot for Czech is a more interesting and
challenging endeavour, Pohádkové díteˇ is the primary focus of
our work; we provide the Fairytale Child version mainly
for our work to be intelligible also for non-Czech speakers.
ling [
Please observe the ways in which the chatbot tries to ask questions. Also please notice how it remembers what it was told about the objects, and when they are mentioned again, it can decide to show that it has been listening by repeating what it was told. Interestingly, this leads to a “mistake in understanding” towards the end of the story, which, by accident, seems very natural and appropriate at that point. 2.1
This session transcript is real and unedited (except for a few corrections of user’s typos). Therefore, it also shows several cases where the chatbot misanalyzed the sentence; these will be referred to in Section 4. (38) USER: No a kacˇátko uvideˇlo, jak k jezeru prˇilétají z teplých krajin krásní velcí bílí ptáci. (39) CHATBOT: Byli hodneˇ bílí? (40) USER: Byli bílí jako sníh. (41) CHATBOT: Pokracˇuj... (42) USER: Byly to labuteˇ. (43) CHATBOT: Pokracˇuj... (44) USER: Labuteˇ pˇristávaly na jezero, a kácˇátko k nim opatrneˇ šlo a obdivovalo je. (45) CHATBOT: Pokracˇuj... (46) USER: Cˇ ekalo, kdy si ho všimnou a budou na neˇj zase pokrˇikovat. (47) CHATBOT: Pokracˇuj... (48) USER: Ale labuteˇ na neˇj ani nepokrˇikovaly, ani ho neodháneˇly, ani se mu nesmály. (49) CHATBOT: Pokracˇuj... (50) USER: Kacˇátko se tomu divilo, a ptalo se labutí: Copak vám neprˇipadám ošklivé? (51) CHATBOT: Pokracˇuj... (52) USER: A na to se labuteˇ divily a ptaly se ho: Procˇ bys nám prˇipadalo ošklivé? Vždyt’ jsi stejneˇ krásné jako my! (53) CHATBOT: Jaké to byly labuteˇ? (54) USER: Ty, co na zimu odleˇtly do teplých krajin, stejneˇ jako divoké kachny. (55) CHATBOT: Byla hodneˇ divoká? (56) USER: Ale to se jen tak rˇíká, divoké kachny, to znamená, že žijou volneˇ v prˇírodeˇ. (57) CHATBOT: Jaká to byla prˇíroda? (58) USER: No prˇíroda kolem toho jezera. (59) CHATBOT: Aha, to, u kterého se kacˇátko narodilo. (60) USER: Ano to. No a to kacˇátko se divilo: To prˇece není pravda, vy jste prˇekrásné labuteˇ a já jsem ošklivé kacˇátko. (61) CHATBOT: Pokracˇuj... (62) USER: A labuteˇ mu rˇekly: tak pojd’ blíž a podívej se na sv˚uj odraz na hladineˇ. (63) CHATBOT: Jaká to byla hladina? (64) USER: Hladina toho jezera. (65) CHATBOT: Pokracˇuj... (66) USER: A tak se kacˇátko podívalo, ale místo ošklivého kacˇátka uvideˇlo krásnou bílou labut’. (67) CHATBOT: Aha, ty co na zimu odleˇtly do teplých krajin stejneˇ jako divoké kachny. (68) USER: Práveˇ že ne – uvideˇlo totiž samo sebe. (69) CHATBOT: Pokracˇuj... (70) USER: Nebylo to totiž ve skutecˇnosti kacˇátko, ale labut’átko. (71) CHATBOT: Pokracˇuj... (72) USER: No a to je konec. (73) CHATBOT: To už je konec? Tak deˇkuju za pohádku! :-) 2.2 .
This is the English version of the sample Ugly Duckling session from Section 2.1. The user uttrances have been translated from Czech to English. From the responses that the English version of the chatbot is able to generate, we chose the ones that correspond to the responses in the Czech session. Thus, this session is not a real one, but it is a realistic one, in the sense that the system is able (and reasonably likely) to generate exactly the responses shown here. (60) USER: Yes, this one. Well, and the duckling was startled: This is simply nonsense, you are beautiful swans and I am just an ugly duckling. (61) CHATBOT: Go on... (62) USER: And the swans told it: well then come closer and look at your reflection at the surface. (63) CHATBOT: What surface was it? (64) USER: The surface of the lake. (65) CHATBOT: Go on... (66) USER: And so the duckling looked, but instead of an ugly duckling it saw a beautiful white swan. (67) CHATBOT: I see, the swans that flew south, just as the wild ducks did. (68) USER: Well, actually not – it actually did see itself. (69) CHATBOT: Go on... (70) USER: It wasn’t in fact a duckling, but a swanling. (71) CHATBOT: Go on... (72) USER: And this is the end. (73) CHATBOT: This is the end already? Thanks for the story then! :-)
Czech sentence analysis scenario Treex::Block::W2A::CS::Tokenize Treex::Block::W2A::CS::TagMorce Treex::Block::W2A::CS::FixMorphoErrors English sentence analysis scenario Treex::Block::W2A::EN::Tokenize Treex::Block::W2A::EN::NormalizeForms Treex::Block::W2A::EN::FixTokenization Treex::Block::W2A::EN::TagMorce Treex::Block::W2A::EN::FixTags
Treex::Block::W2A::EN::Lemmatize
The chatbot is implemented as a console application in Perl, reading user input from standard input, and producing responses to standard output. It extensively uses Treex framework blocks and tools, but is not implemented as a Treex application, as Treex scenarios do not support interactive operation.
The script starts by initializing all tools and asking the user to tell a fairy tale (see utterance (1) in Section 2). Then it iteratively analyses the sentence entered by the user with NLP tools (Section 3.1), selects a suitable predefined response template and generates a response (Section 3.2).1 This loop continues until the end of the story is detected – this is done by detecting the presence of the word “end” (“konec” in Czech), see (72)-(73). The chatbot thanks for the story and terminates. 3.1
Each line entered by the user is analyzed with a standard Treex scenario – see Table 1. The sentences are split into individual tokens, and each token is assigned a lemma2 and a part-of-speech3 tag.
For Czech, the standard Czech positional tagset is
used [
For English, the standard Penn Treebank tagset [
If the chatbot asks a question, the user’s answer is stored in a simple knowledge base. Before storing, the answer is normalized in two steps. The first one is the removal of sentence-initial garbage words “Well”/“Oh” (“No” in Czech) – see (8)-(9). The second one is more intricate: it is a switch between 1st and 2nd person – see (12)-(13). This consists of switching “I”/“me” and “you” in English (“já” and “ty” in Czech), and also switching the person for Czech pronouns and verbs.
The knowledge base is simply a key-value hashmap. The key is the lemma of the word (noun or adjective) that the chatbot asked about, and the value is the normalized answer of the user to that question. 3.2
The chatbot randomly chooses from a set of the following three possible response types. It then tries to generate the response of the given type. If this is impossible, e.g. because the template requires a noun but there is no noun in the sentence – see (5) – it responds “Go on...” (“Pokra cˇuj...”) instead. This randomness makes the responses less predictable and therefore more realistic.
For generating Czech responses, a morphological
generator [
What NOUN was it? The first response type asks the user to provide details about a noun – see e.g. (23). The noun to be asked about (denoted as NOUN) is the last noun in the sentence, as this is presumably the newest and most interesting piece of information.
In Czech, NOUN is first switched to nominative case using the morphological generator (but morphological number is preserved). The response generated has the form of “ Jak_ to byl_ NOUN?”; the correct form of “Jak_” (“Jaký”, “Jaká”, “Jaké”... – “What”) and “byl_” (“byl”, “byli”... – “was”/“were”) is chosen based on the morphological gender and number of the NOUN (and generated by the morphological generator).
In English, the response has a form of “What NOUN was it/were they?”, based on the singular/plural marked in the NOUN tag (NN and NNP tags mark singular, NNS and NNPS tags mark plural).
If a question regarding NOUN had already been asked (i.e. the knowledge base already contains NOUN as a key), the chatbot does not ask the question again, but instead shows off by reapeating the then-provided answer – “I see, ANSWER.” (“Aha, ANSWER.”) – see (35)-(36) and then (58)-(59).4 What is a NOUN? The second response type asks the user to explain what a noun means – see e.g. (29). Again, the noun to be asked for (denoted as NOUN) is the last noun in the sentence.
In Czech, NOUN is first switched to nominative, preserving the number. The response generated has the form of “Co je/jsou to NOUN?”, the correct choice of verb “je” or “jsou” (“is” or “are”) being made based on the morphological number of the noun.
In English, the response has a form of “What is/are a/an NOUN?”, based on the singular/plural marked in the NOUN tag, and on the first letter of NOUN.5
Similarly to the first response type, if a question regarding NOUN had already been asked, the chatbot reapeats the answer instead.
Very ADJECTIVE? The third type asks whether a property was intense. Similarly to previous types, the adjective to be asked about (denoted ADJECTIVE) is the last adjective of the sentence.
In Czech, the ADJECTIVE is first converted to nominative. Then, the response is generated in the form “ Byl_ hodneˇ ADJECTIVE?” (the correct form of “byl_” is generated according to morphological gender and number of ADJECTIVE).
In English, the form is only “Very ADJECTIVE?”, for reasons discussed in Section 4.
If a question regarding ADJECTIVE had already been asked, the chatbot resorts to the default “Go on” answer, as for adjective questions, the answers are not reproduced. The idea is that while it is reasonable to believe that a repeated noun is a reference to the same entity, this is less likely for adjectives. However, to avoid annoying questions, the chatbot will not ask about the same adjective twice. 3.3
Although both Treex and Perl are sometimes blamed for their slowness, we have not encountered such issues with our application. This is helped by the fact that all the Treex tools are loaded only once – thus, the startup takes about 7 4Note that (10)-(11) and (18)-(19) seem to contradict this claim. However, on closer inspection of the original Czech transcript, it can be seen that the user used different spellings for the word “duckling” – “ká cˇátko” in (10) but “ka cˇátko” in (18). This difference has no influence on meaning in Czech, but as the tagger preserves this distinction in the lemmas it produces, our system treats the two as distinct, and therefore asks about the duckling again instead of showing that it already knows what the user is talking about. (The distinction has not been preserved in the English version because of the lack of such a dublet in English.) 5We resort to a simplification here by treating all vowel letters as vowels – i.e. we would incorrectly generate e.g. “an uniform” instead of the correct “a uniform”. seconds for Czech (about 3 seconds for English), but each response then takes only about 70 ms for Czech (about 15 ms for English).6 We believe this to be perfectly acceptable for real-world usage. 4
The whole conversation between the user and the chatbot has a reasonably natural feel. The chatbot usually reacts appropriately, which is demonstrated by the ability of the user to answer chatbot’s questions and continue with the story, eventually reaching the end.
Occasionally, we encounter errors that stem from incorrect morphological analysis disambiguation by the tagger. One such case can be seen in (15), where “Byla hodn eˇ smutná?” (“Was she very sad?”) was generated instead of the correct “Bylo hodn eˇ smutné?” (“Was it very sad?”), as the neuter nominative “smutné” (“sad”) in (14) was misanalyzed as feminine genitive. A similar situation happens also in (54)-(55), where feminine plural nominative is misanalyzed as feminine singular genitive, and in (24)-(25), where “starý” (“old”) is analyzed as a noun instead of adjective. However, such errors happen already in the tool that we use, and are effectively beyond our influence.
An error that we have encountered (but does not show in the sample conversation) is that of asking the “Very ADJECTIVE?” question with inappropriate addjectives, as in “He sat there the whole day.” (“Sed eˇl tam celý den.”) – “Very whole?” (“Velmi celý?”). However, it seems to occur especially with several frequent adjectives, so enumerating them manually seems to be a viable remedy.
We also experimented with the “Why?” question, but have found it hard to identify the appropriate part of the user’s utterance to use in the question – the whole sentence seems often too long and awkward, while only taking the verb often leads to a non-sentence. This is an issue to be addresed once we add deeper levels of linguistic analyses.
There are still issues with the English implementation. One of them is that the Czech question with “Jaký”, such as (23) “Jaký to byl pán?”, is conveniently ambiguous, as it can mean both “What was the man like?” and “Which man was it?”, and thus feels more natural than the English one. In English, we are not aware of such an ambiguous construction, and we thus use “What man was it?”, which tries to keep the ambiguity but is not good English.
There are also cases where the Czech rich morphology and analysis helps – especially when dealing with gender, which is not marked in English analysis. Thus, when asking the “Very ADJECTIVE?” question, we can choose the correct subject in Czech based on the morphological features of the adjective, while this is not possible in English. This can be seen e.g. in (3), where in Czech we generate the whole question “Bylo hodn eˇ ošklivé?” (“ Was it 6Measured on Intel R Core 2 Duo T5870 running 32b Ubuntu 10.10, averaged over 5 runs. The response time is averaged over the 36 sentences from Section 2. very ugly?”), because the morphological tag tells us that “ošklivé” (“ugly”) was used in a neuter gender, while in English, we currently resort only to “Very ugly?”, since we cannot easily distinguish whether we should use “he”, “she” or “it” (or another pronoun) as the subject. This is often distractive and leads to unnatural questions, as in (39). The gender is not only unmarked on the adective itself, but the English tagset does not even mark gender on nouns (so there is no simple and direct way of telling that a “king” is a “he” while a “queen” is a “she”). However, there exist tools that are able to fill that shortcoming, and we will try to use them in future. Still, it will necessitate to use not only tagging but also syntactic parsing to be able to identify the gender associated with an adjective, since it is implied by the parent noun in a dependency tree, which might not be straightforward to find without a proper syntactic analysis of the sentence.
Employing a syntactic parser and a function labeller,
such as [
An obvious and important future goal is to implement
a web interface. Currently, the application is
terminalbased, reading user utterances from the standard input and
producing chatbot responses to standard output. This does
not look very appealing; moreover, it requires the users to
have Treex installed. Even a simple web interface would
make the application available to a much wider range of
users, and would also make it more attractive. While Treex
does have a web interface [
We have implemented a Fairytale Child chatbot, which acts as a child that wants to hear a fairy tale. It analyzes the user utterances and tries to react with a question or remark if possible, simulating a curious child. The chatbot is able to communicate in Czech or in English.
We used the Treex natural language processing framework for both analysis of the user utterances and generation of the chatbot responses. This proved to be effective, especially for the morphologically rich Czech language.
Even though the implementation of the chatbot is quite simple, it responds quite naturally in most cases. This shows that even for morphologically rich languages, a simple approach can lead to a well-performing chatbot, provided that there is a powerful NLP framework at hand.
The chatbot is licenced under GNU GPL v2 and is available at https://ufal.mff.cuni.cz/dite.7
This research was supported by the grants GAUK 1572314, and SVV 260 104. This work has been using language resources developed, stored and distributed by the LINDAT/CLARIN project of the Ministry of Education, Youth and Sports of the Czech Republic (project LM2010013).